sensors

Article
UAV-YOLOv8: A Small-Object-Detection Model Based on
Improved YOLOv8 for UAV Aerial Photography Scenarios
Gang Wang, Yanfei Chen *, Pei An, Hanyu Hong, Jinghu Hu and Tiange Huang

Hubei Key Laboratory of Optical Information and Pattern Recognition, School of Electrical and Information
Engineering, Wuhan Institute of Technology, Wuhan 430205, China; [email protected] (G.W.);
[email protected] (P.A.); [email protected] (H.H.); [email protected] (J.H.); [email protected] (T.H.)
* Correspondence: [email protected]

Abstract: Unmanned aerial vehicle (UAV) object detection plays a crucial role in civil, commercial,
and military domains. However, the high proportion of small objects in UAV images and the limited
platform resources lead to the low accuracy of most of the existing detection models embedded
in UAVs, and it is difficult to strike a good balance between detection performance and resource
consumption. To alleviate the above problems, we optimize YOLOv8 and propose an object detec-
tion model based on UAV aerial photography scenarios, called UAV-YOLOv8. Firstly, Wise-IoU
(WIoU) v3 is used as a bounding box regression loss, and a wise gradient allocation strategy makes
the model focus more on common-quality samples, thus improving the localization ability of the
model. Secondly, an attention mechanism called BiFormer is introduced to optimize the backbone
network, which improves the model’s attention to critical information. Finally, we design a feature
processing module named Focal FasterNet block (FFNB) and propose two new detection scales based
on this module, which makes the shallow features and deep features fully integrated. The proposed
multiscale feature fusion network substantially increased the detection performance of the model
and reduces the missed detection rate of small objects. The experimental results show that our model
has fewer parameters compared to the baseline model and has a mean detection accuracy higher
than the baseline model by 7.7%. Compared with other mainstream models, the overall performance
of our model is much better. The proposed method effectively improves the ability to detect small
Citation: Wang, G.; Chen, Y.; An, P.; objects. There is room to optimize the detection effectiveness of our model for small and feature-less
Hong, H.; Hu, J.; Huang, T. objects (such as bicycle-type vehicles), as we will address in subsequent research.
UAV-YOLOv8: A
Small-Object-Detection Model Based Keywords: UAVs; small-object detection; YOLOv8; WIoU; BiFormer; FasterNet
on Improved YOLOv8 for UAV
Aerial Photography Scenarios.
Sensors 2023, 23, 7190. https://
doi.org/10.3390/s23167190
1. Introduction
Academic Editor: Pasquale Daponte With the continuous reduction in the production cost of UAVs and the gradual maturity
Received: 23 July 2023
of flight control techniques, the application of UAVs is becoming increasingly widespread
Revised: 8 August 2023 in areas such as power line inspections [1], traffic monitoring [2], and crop analysis [3].
Accepted: 12 August 2023 Object detection plays an increasingly important role as a key link in the missions carried
Published: 15 August 2023 out by UAVs and has very great research significance. The UAV has a wide field of view
due to its high flying altitude, leading to problems such as a high proportion of small
objects and complex backgrounds in the captured images, which increases the difficulty of
the object detection task. Moreover, UAV platforms have limited resources, making it hard
Copyright: © 2023 by the authors. to embed high computational and storage-demanding object detection models. Therefore,
Licensee MDPI, Basel, Switzerland. enhancing the performance of object detection while considering the limited resources of
This article is an open access article the hardware platform is one of the core issues in object detection in UAV aerial scenes.
distributed under the terms and
The essential difference among object detection algorithms is that the features of the
conditions of the Creative Commons
image are extracted differently. Most of the traditional object detection algorithms are
Attribution (CC BY) license (https://
reusing classifiers, such as the deformable parts model (DPM) [4].The DPM first uses
creativecommons.org/licenses/by/
classifiers to slide over the image, and then the output of the classifiers is aggregated as
4.0/).

Sensors 2023, 23, 7190. https://doi.org/10.3390/s23167190 https://www.mdpi.com/journal/sensors

Sensors 2023, 23, 7190 2 of 27

the object detection results. This detection method is very time-consuming and tedious,
and the detection effect is poor. The current mainstream object detection algorithms mainly
use deep learning methods, and we classify them into two categories: two-stage and
one-stage. The R-CNN family [5–7] is a very classical two-stage algorithm, which first
extracts candidate frames, then uses a classifier to filter them, and finally removes duplicate
boxes and finetunes the predicted boxes using non-maximal value suppression. The two-
stage detector has some advantages in terms of detection accuracy but has disadvantages
such as difficulty in training, slow detection speed, and difficulty in optimization. One-
stage detectors include methods such as the you only look once (YOLO) series [8–13]
and single-shot multibox detector (SSD) [14], which use a separate neural network for
one forward inference to generate the coordinates and category results of the prediction
boxes. The one-stage object detection algorithm improves the speed of detection but loses
some of the detection accuracy. In addition, thanks to the excellent performance of the
transformer [15] model in the field of natural language processing, Carion et al. introduced
the model into the field of computer vision and proposed the detection with transformers
(DETR) model [16], which achieved desirable results and provided a new research idea
for object detection. However, the objects in natural scenes are multiscale in form, and the
objects in the UAV viewpoint are mainly small. So, the mainstream detection algorithms
mentioned above are not feasible to be used directly for the object detection task in UAV
aerial photography scenes.
A large amount of research work has emerged in the field of object detection in
UAV aerial photography scenarios in recent years. Luo et al. [17] optimized the detection
performance by improving the network module in YOLOv5. The effectiveness of the
proposed strategy is verified by numerous datasets, but the detection of small objects is
inferior. Zhou et al. [18] solved the problem of monotonous backgrounds in UAV images by
using background replacement from the perspective of data enhancement. However, this
work was not effective in improving the detection accuracy of small objects. Du et al. [19]
designed sparse convolution to optimize the detection head from the perspective of a
lightweight model. The sparse convolution reduces the computational effort but decreases
the detection accuracy of the model. Deng et al. [20] proposed a lightweight network in
order to improve the efficiency of insulator fault detection in transmission lines by UAVs.
The method uses YOLOv4 as the baseline model. Firstly, the original backbone network is
replaced with MobileNetv3 [21], which effectively reduces the parameters of the model.
Secondly, the method also improves the generalization ability of the model by improving
the loss function. Finally, the binary particle swarm optimization idea is introduced to
reduce the delay of fault detection. Zheng et al. [22] proposed a multispecies oil palm tree
detection method called MOPAD, which can detect oil palm trees well and can accurately
monitor the growth of oil palm trees. The method combines Faster RCNN, Refined Pyramid
Feature (RPF), and Hybrid Balanced Loss Module. MOPAD achieves desirable observations
on three large oil palm tree datasets and the average F1 scores outperform other state-of-
the-art detection models. Liu et al. [23] proposed a small-target-detection method in the
UAV view to reduce the leakage and false detection rate of small targets. The method
uses YOLOv3 as the base model, and it optimizes the backbone network by introducing
ResNet [24] units and adding convolutional operations to enhance the receptive field
of the model. For the current situation of more and more illegal flights of multi-rotor
UAVs, Liu et al. [25] proposed a novel detection method to improve the detection accuracy
of multi-rotor UAVs. The method uses Efficientlite to replace the backbone network of
YOLOv5, which reduces the complexity of the model and improves the detection efficiency.
In addition, adaptive spatial feature fusion is used at the head of the baseline model to
optimize the detection accuracy of the model. To improve the detection performance of
UAV aerial images, Wang et al. [26] proposed a lightweight detection model called MFP-
YOLO by optimizing YOLOv5. Firstly, the method designs a multiplexed inverted residual
block (MIRB) and introduces the convolutional block attention module (CBAM) [27], which
effectively improves the model’s detection effect under an environment of scale variation
Sensors 2023, 23, 7190 3 of 27

and background complexity. Secondly, the formula introduces a parallel convolutional

spatial pyramid pooling framework, which takes into account targets at different scales.
Finally, a lightweight decoupled detection header is applied to the baseline model, which
reduces the parameters of the model while maintaining its detection accuracy. Liu et al. [28]
proposed a multi-branch parallel feature pyramid network (MPFPN) to reduce the leakage
rate of small-target detection in UAV images. Meanwhile, supervised spatial attention
module (SSAM) was added to this network to suppress the interference of background
noise. Finally, the effectiveness of the method was demonstrated at the public data level.
Most of the current research methods generally have low accuracy for object detection in
UAV aerial photography scenarios, and it is difficult to balance the relationship between
the accuracy of the model and resource consumption.
To alleviate the above problems, we propose an object detection model based on UAV
aerial photography scenarios, called UAV-YOLOv8, using YOLOv8 as the backbone network.
This model not only improves the performance of target detection but also does so without
too much resource consumption. The main contributions of this paper are as follows:
• We propose an efficient and fast feature processing module called the FFNB based
on the FasterNet block [29]. Utilizing this module, we design two new detection
scales that enable the comprehensive fusion of shallow and deep features, significantly
reducing the missed detection rate of small objects.
• We introduce a low-computational-cost dynamic sparse attention mechanism BiFormer [30]
in the backbone network, which improves the model’s attention to the critical information
in the feature map and optimizes the detection performance of the model.
• We incorporate WIoU v3 [31] in our bounding box regression loss, which employs a
dynamic non-monotonic mechanism to design a more reasonable gradient gain alloca-
tion strategy. WIoU v3 effectively reduces the gradient gain of high-quality samples
and low-quality samples, which enhances the model’s localization performance and
generalization ability.
• Compared with some mainstream YOLO series models as well as six other classical
detection models, the experimental results demonstrate the superiority of our method.
In addition, we perform visual analysis from three perspectives to explain that our
proposed method effectively improves the detection performance for small objects.

2. YOLOv8 Detection Algorithm

The YOLO model has been a great success in the field of computer vision; based
on this, researchers have improved and added new modules to the method, proposing
many classical models. YOLOv8 is an algorithm released by the Ultralytics company on
10 January 2023. Compared to previous excellent models in the YOLO series (such as
YOLOv5 and YOLOv7), YOLOv8 is an advanced and cutting-edge model that offers higher
detection accuracy and speed.
The YOLOv8 network structure mainly consists of a backbone, neck, and head, as
shown in Figure 1.

2.1. Backbone
YOLOv8 uses modified CSPDarknet53 [10] as the backbone network, and the input
features are down-sampled five times to obtain five different scale features, in turn, which
we denote as B1–B5. The structure of the backbone network is shown in Figure 1a. The
Cross Stage Partial (CSP) module in the original backbone network is replaced by the
C2f module, and the structure of the C2f module is shown in Figure 1f (n denotes the
number of bottlenecks). The C2f module adopts a gradient shunt connection to enrich
the information flow of the feature extraction network while maintaining a light weight.
The CBS module performs a convolution operation on the input information, followed
by batch normalization, and finally activates the information stream using SiLU to obtain
the output result, as shown in Figure 1g. The backbone network finally uses the spatial
pyramid pooling fast (SPPF) module to pool the input feature maps to a fixed-size map for
Sensors 2023, 23, 7190 4 of 27

adaptive size output. Compared with the structure of spatial pyramid pooling (SPP) [32],
Sensors 2023, 23, x FOR PEER REVIEW
reduces the computational effort and has lower latency by sequentially 4connecting
SPPF of 28

the three maximum pooling layers, as shown in Figure 1d.

Figure
Figure 1. The
1. The networkstructure
network structure of
of YOLOv8.
YOLOv8.The Theww(width)
(width)andand
r (ratio) in Figure
r (ratio) 1 are parameters
in Figure 1 are parameters
used to represent the size of the feature map. The size of the model can be controlled by setting the
used to represent the size of the feature map. The size of the model
values of w and r to meet the needs of different application scenarios.
can be controlled by setting the
values of w and r to meet the needs of different application scenarios.
2.1. Backbone
2.2. Neck
YOLOv8 uses modified CSPDarknet53 [10] as the backbone network, and the input
Inspired
features areby PANet [33], five
down-sampled YOLOv8 is obtain
times to designed
five with a PAN-FPN
different structure
scale features, at which
in turn, the neck, as
shown in Figure 1b. Compared with the neck structure of YOLOv5
we denote as B1–B5. The structure of the backbone network is shown in Figure 1a. Theand YOLOv7 models,
YOLOv8
Cross Stage Partial (CSP) module in the original backbone network is replaced by the C2fwhich
removes the convolution operation after up-sampling in the PAN structure,
maintains
module, the
andoriginal performance
the structure while achieving
of the C2f module is shown inaFigure
lightweight model.
1f (n denotes theWe use P4-P5
number
andofN4-N5
bottlenecks). The C2f
to denote themodule adopts ascales
two different gradient shunt connection
of features in the PANto enrich the infor-
structure and FPN
mationof
structure flow
theofYOLOv8
the feature extraction
model, network Conventional
respectively. while maintaining FPNa uses
light aweight.
top-downThe CBS
approach
module deep
to convey performs a convolution
semantic operation
information. The on
FPN theenhances
input information,
the semanticfollowed by batch of the
information
normalization, and finally activates the information stream using
features by fusing B4-P4 and B3-P3, but some object localization information willSiLU to obtain the be
out-
lost. To
put result, as shown in Figure 1g. The backbone network finally uses the spatial pyramid
alleviate this problem, PAN-FPN adds PAN to FPN. PAN enhances the learning of location
pooling fast (SPPF) module to pool the input feature maps to a fixed-size map for adaptive
information by fusing P4-N4 and P5-N5 to realize path enhancement in a top-down form.
size output. Compared with the structure of spatial pyramid pooling (SPP) [32], SPPF re-
PAN-FPN constructs a top-down and bottom-up network structure, which realizes the
duces the computational effort and has lower latency by sequentially connecting the three
complementarity
maximum pooling of shallow
layers, aspositional information
shown in Figure 1d. and deep semantic information through
feature fusion, resulting in feature diversity and completeness.
2.2. Neck
2.3. Head
Inspired by PANet [33], YOLOv8 is designed with a PAN-FPN structure at the neck,
asThe
shown detection part
in Figure 1b.of YOLOv8with
Compared usesthe
a decoupled head
neck structure structure,
of YOLOv5 andas YOLOv7
shown inmod-
Figure 1e.
Theels,
decoupled head structure
YOLOv8 removes uses two
the convolution separate
operation afterbranches
up-samplingfor in
object classification
the PAN structure, and
which maintains
predicted bounding thebox
original performance
regression, while achieving
and different a lightweight
loss functions are model.
used forWethese
use two
P4-P5 and N4-N5 to denote the two different scales of features in the PAN structure
types of tasks. For the classification task, binary cross-entropy loss (BCE Loss) is used. For and
theFPN structure
predicted boxofbounding
the YOLOv8 model, task,
regression respectively. Conventional
distribution focal lossFPN
(DFL)uses[34]
a top-down
and CIoU [35]
are employed. This detection structure can improve detection accuracy andinfor-
approach to convey deep semantic information. The FPN enhances the semantic accelerate
mation of the features by fusing B4-P4 and B3-P3, but some object localization information
model convergence. YOLOv8 is an anchor-free detection model that concisely specifies
will be lost. To alleviate this problem, PAN-FPN adds PAN to FPN. PAN enhances the
positive and negative samples. It also uses the Task-Aligned Assigner [36] to dynamically
learning of location information by fusing P4-N4 and P5-N5 to realize path enhancement
assign samples, which improves the detection accuracy and robustness of the model.
Sensors 2023, 23, 7190 5 of 27

3. Method
YOLOv8 is a state-of-the-art object detection model and takes into account the multi-
scale nature of objects, using three scale-detection layers to accommodate objects of different
scales. However, the images acquired by UAVs have the problems of complex backgrounds
and a high proportion of small objects. This results in the detection structure of YOLOv8
not meeting the detection requirements in UAV aerial photography scenarios. To mitigate
the above problems, this paper uses YOLOv8 as the base model and optimizes the model
from the perspectives of loss function, attention mechanism, and multiscale feature fusion.
The main ideas of the improvement strategy are as follows:
First, WIoU v3 is utilized as the bounding box regression loss. WIoU v3 incorporates
a dynamic non-monotonic mechanism and designs a sensible gradient gain allocation
strategy, which reduces the occurrence of large or harmful gradients from extreme samples.
WIoU v3 focuses more on samples of ordinary quality, thereby improving the model’s
generalization ability and overall performance.
Then, the dynamic sparse attention mechanism BiFormer is introduced into the back-
bone network. BiFormer reduces computation and memory consumption by filtering out
most of the low-relevance regions in the feature graph and then applying attention to the
high-relevance features. BiFormer improves the model’s attention to the key information
in the input features and optimizes the detection performance of the model.
Finally, the efficient feature processing module FFNB is proposed based on FasterNet.
FFNB has fewer computation and memory accesses during feature processing. Based
on FFNB, we design two new detection layers. Our proposed multiscale feature fusion
network makes the shallow feature and deep feature fully complement each other, which
effectively improves the detection effect of the model on small objects.
We refer to the final improved network model as UAV-YOLOv8, and the overall
framework of the model is shown in Figure 2. To show the structure of the improved model
more concisely and intuitively, we omit the drawing of the C2f module and SPPF module
in Figure 2. In Figure 2, large, medium, small, X-small, and XX-small are used to represent
the size of the object. The improved model changes from the original 3-scale detection
tox 5-scale
Sensors 2023, 23, FOR PEERdetection,
REVIEW which effectively improves the overall detection performance of 6the of 28
model, especially for small objects.

Figure 2. The overall structure of the proposed improved model.

Figure 2. The overall structure of the proposed improved model.
For ease of reading, we give a notation of the terms involved with the formulas in
this paper, as shown in Table 1.

Table 1. Explanation of notations.

Notation Explanation
Sensors 2023, 23, 7190 6 of 27

For ease of reading, we give a notation of the terms involved with the formulas in this
paper, as shown in Table 1.

Table 1. Explanation of notations.

Notation Explanation
Loss function
(cbx , cby ), w, h Denote the coordinates of the center point of the prediction box, width, and height, respectively
gt gt
(cbx , cby ), w gt , h gt Denote the coordinates of the center point of the real box, width, and height, respectively
ρ(w, w gt ), ρ(h, h gt ) Denote the distance of width and height between the real box and the predicted box, respectively
cw , ch Denote the width and height of the smallest closed box enclosed by the real and predicted boxes
L∗ IoU , r Monotonic focus coefficient and non-monotonic focus factor
β Degree of outliers
γ, δ, α Hyperparameters (we can adjust these hyperparameters to fit different models)
BiFormer
Q, K, V Queries, keys, and values
Linear transformation matrices corresponding to
WQ, WK, WV
Q, K, and V, respectively
dK Dimension of the matrix K
X ∈ R H × E×C Dimension of the input feature map
2 2
Ar ∈ R S × S Adjacency matrix
2
I r ∈ N S ×k Routing index matrix
I(ri,1) , I(ri,2) , . . . , I(ri,k) Index values corresponding to the k routing regions with the highest correlation in region i
gather (·) Gathering key–value pairs
Attention(·) Regular attention computation

3.1. Improved Loss Function

The object detection task in the UAV aerial photography scene has a high proportion of
small objects, so a reasonably designed loss function can significantly improve the detection
performance of the model. YOLOv8 uses DFL and CIoU to calculate the regression loss of
the bounding box, but CIoU has the following drawbacks: first, CIoU does not consider
the balance of difficult and easy samples. Second, CIoU uses the aspect ratio as one of the
penalty factors of the loss function, and if the aspect ratio of the real box and the predicted
box are the same, but the values of width and height are different, the penalty term cannot
reflect the real difference between these two boxes. Third, the calculation of the CIoU
formula involves an inverse trigonometric function, which will increase the consumption
of model arithmetic power. The formula of CIoU is shown in Equation (1):

ρ2 (b, b gt ) 4 −1 w
gt w
LCIoU = 1 − IoU + + ( tan − tan−1 ) (1)
( c w )2 + ( c h )2 π2 h gt h

In Equation (1), Intersection over Union (IoU) denotes the intersection ratio of the
prediction box and the real box. Some of the parameters involved in Equation (1) are shown
in Figure 3. ρ(b, b gt ) denotes the Euclidean distance between the centroids of the real box
and the prediction box; h and w denote the height and the prediction box; h gt and w gt
denote the height and the width of the real box; ch and cw denote the height and the width
of the minimum enclosing box formed by the prediction box and the real box.
Sensors
Sensors2023,
2023,23,
23, x7190
FOR PEER REVIEW 8 7ofof28
27

Schematicdiagram
Figure3.3.Schematic
Figure diagramof
ofthe
theparameters
parametersof
ofthe
theloss
loss function.
function.

EIoU[38]
SIoU [37]introduces
improves based on between
the angle CIoU by thetreating the length
predicted and
box and width
the separately
real box as
as a pen-
penalty terms, reflecting the difference in width and height between the real box and
alty factor for the first time. Firstly, based on the magnitude of the angle (as in Figure 3, the
θpredicted
and α )box, whichthe
between is predicted
more reasonable
box andcompared with
the real box, thethe penalty box
predicted term of CIoU.
rapidly The
moves
formula of EIoU is shown in Equation (2):
towards the nearest axis and then regresses towards the real box. SIoU reduces the degrees
of freedom of the regression and speeds up the convergence of the model.
ρ2 (b, b gt ) ρ2 (w, w gt ) ρ2 (h, h gt )
While several L EIoU = 1 − IoU loss
mainstream + functions introduced
+ above
+ take2 a static focusing (2)
mechanism, WIoU not only considers )2 +aspect,
(cwthe (ch )2 centroid(cw )2distance,(cand h ) overlap area but
also introduces a dynamic non-monotonic focusing mechanism. WIoU applies a reasona-
Some of the parameters involved in Equation (2) are shown in Figure 3; ρ(w, w gt )
ble gradient gain allocation strategy to evaluate the quality of the anchor box. Tong et al.
and ρ(h, h gt ) denote the Euclidean distance of width and Euclidean distance of height
[31] proposed three versions of WIoU. WIoU v1 was designed with attention-based gt gt pre-
between the real box and the prediction box, respectively; (cbx , cby ) and (cbx , cby ) denote
dicted box loss, and WIoU v2 and WIoU v3 added focusing coefficients.
the coordinates of the center
WIoU v1 introduces pointsasofathe
distance real box
metric and the prediction
of attention. When the box, respectively.
object box and the
SIoU [38] introduces the angle between the predicted box
predicted box overlap within a certain range, the penalty of reducing the geometric and the real box as a penalty
metric
factor for the first time. Firstly, based on the magnitude of the
makes the model obtain better generalization ability. The formula for calculating angle (as in Figure 3, θ WIoU
and α)
between
v1 is showntheinpredicted
Equations box and the real box, the predicted box rapidly moves towards the
(3)–(5):
nearest axis and then regresses towards the real box. SIoU reduces the degrees of freedom
of the regression and speeds up theLconvergenceWIoUv1 = RWIoUof×the LIoUmodel. (3)
While several mainstream loss functions introduced above take a static focusing
mechanism, WIoU not only considers the (bcgtxaspect,
− bcx ) 2centroid
+ (bcgty −distance,
bc y ) 2 and overlap area but
WIoU = exp(
also introduces a dynamicRnon-monotonic ) applies a reasonable
focusing2 mechanism. 2 WIoU (4)
gradient gain allocation strategy to evaluate the quality
( c w + c )
h of the anchor box. Tong et al. [31]

proposed three versions of WIoU. WIoU v1 was designed with attention-based predicted
LIoU = 1 − IoU (5)
box loss, and WIoU v2 and WIoU v3 added focusing coefficients.
WIoUv2
WIoU v1isintroduces
applied to distance
WIoU v1as byaconstructing
metric of attention. When focus
the monotonic the object box andL∗ the
coefficient IoU

,predicted box overlap

which effectively within
reduces theaweight
certainofrange,
simple the penalty of
examples in reducing the geometric
the loss value. However,metric
con-
makes the
sidering thatmodelL∗ IoUobtain better generalization
decreases as LIoU decreases ability.during
The formula
modelfor calculating
training, WIoU in
resulting v1
is shown in Equations (3)–(5):
slower convergence, the average of LIoU is introduced to normalize L∗ IoU . The formula
of WIoU v2 is shown in EquationL(6): W IoUv1 = RW IoU × L IoU (3)
L∗ γ 2
LWIoUv 2 = ((bcgtxIoU
− )bcx×) L+ (bc1y, γ− >
2
bcy0)
gt
WIoUv (6)
RW IoU = exp( IoU L 2 2
) (4)
(cw + ch )
WIoU v3 defines the outlier β to measure the quality of the anchor box, constructs
a non-monotonic focus factor r based on =β1,−
L IoU and applies r to WIoU v1. A small value
IoU (5)
of β indicates a high anchor box quality, and a smaller r is assigned to it, reducing the
weight of high-quality anchor frames in the larger loss function. A large value of β
Sensors 2023, 23, 7190 8 of 27

WIoU v2 is applied to WIoU v1 by constructing the monotonic focus coefficient L∗ IoU ,

which effectively reduces the weight of simple examples in the loss value. However,
considering that L∗ IoU decreases as L IoU decreases during model training, resulting in
slower convergence, the average of L IoU is introduced to normalize L∗ IoU . The formula of
WIoU v2 is shown in Equation (6):

L∗ IoU γ
LW IoUv2 = ( ) × LW IoUv1 , γ > 0 (6)
L IoU

WIoU v3 defines the outlier β to measure the quality of the anchor box, constructs a
non-monotonic focus factor r based on β, and applies r to WIoU v1. A small value of β
indicates a high anchor box quality, and a smaller r is assigned to it, reducing the weight
of high-quality anchor frames in the larger loss function. A large value of β indicates a
low-quality anchor box, and a small gradient gain is assigned to it, which reduces the
harmful gradients generated by low-quality anchor boxes. WIoU v3 uses a reasonable
gradient gain allocation strategy to dynamically optimize the weight of high- and low-
quality anchor boxes in the loss, which makes the model focus on the average quality
samples and improves the overall performance of the model. The WIoU v3 formulas
are shown in Equations (7)–(9). δ and α in Equation (8) are hyperparameters that can be
adjusted to fit different models.

LW IoUv3 = r × LW IoUv1 (7)

β
r= (8)
δα β−δ
L∗ IoU
β= ∈ [0, +∞) (9)
L IoU
By comparing the several mainstream loss functions above, we finally introduce
WIoU v3 in the object bounding box regression loss. On the one hand, WIoU v3 takes into
account some advantages of EIoU and SIoU, which is in line with the design concept of
the excellent loss function. On the other hand, WIoU v3 uses a dynamic non-monotonic
mechanism to evaluate the quality of anchor boxes, which makes the model focus more on
anchor boxes of ordinary quality and improves the model’s ability to localize objects. For
the object detection task in the UAV aerial photography scene, the high proportion of small
objects increases the detection difficulty, and WIoU v3 can dynamically optimize the loss
weights of small objects to improve the detection performance of the model.

3.2. Efficient Attention Mechanism

Due to the complex backgrounds and high proportion of small objects in images
captured by UAVs, many detection models have a poor ability to suppress background
information. To make the detection model focus more on the key information in the input
features and less on the background information, we introduce a dynamic sparse attention
mechanism called BiFormer in the backbone network of the model. BiFormer utilizes
query adaptation to first filter out the most irrelevant key–value pairs in coarse-grained
regions of the input feature graph, efficiently find key–value pairs with high relevance, and
then perform attention computation. This significantly reduces computation and storage
consumption and enhances the model’s perception of the input content.
YOLOv8 is a convolutional neural network (CNN) model, and a CNN is essentially
local processing, so the relationships between global features cannot be obtained. Compared
with traditional CNN models, the transformer uses an attention mechanism to obtain the
degree of correlation between data and other data, with the property of a global sensing
field. An effective attention mechanism can build robust and powerful data-driven models,
making the models more flexible when dealing with complex and large data. The attentional
mechanism works as follows: first, [ x1 , x2 , x3 , · · · , x T ] is acquired by encoding the input
Sensors 2023, 23, 7190 9 of 27

data sequence [ a1 , a2 , a3 , · · · , a T ]. Then, three matrices of queries Q, keys K, and values V

are obtained by linear transformation matrices W Q , W K , and W V , respectively. The dot
product between the query and the corresponding key is computed, √ then normalized, and
finally multiplied with matrix V to obtain the weighted sum. dK is introduced to prevent
the gradient of the result from vanishing, and dK denotes the dimensionality of the matrix
K. The
Sensors 2023, 23, x FOR PEERformula
REVIEW for attention is shown in Equation (10): 10 of

QK T
Attention( Q, K, V ) = so f tmax ( √ )V (10)
dK
However, the conventional attention mechanism has the drawbacks of high comp
However,tational complexityattention
the conventional and large memory usage.
mechanism Detection
has the drawbacks models deployed
of high computa- on UAV pl
forms are resource-constrained, and if a conventional attention
tional complexity and large memory usage. Detection models deployed on UAV platforms module is introduced
rectly into the
are resource-constrained, andmodel, it will occupy
if a conventional most ofmodule
attention the platform resources
is introduced and reduce
directly into the inf
the model, it will occupy most of the platform resources and reduce the inference speed of research
ence speed of the model. To ease the computational and memory problems,
the model. To have proposed
ease the to reduce
computational resource
and memory consumption by replacing
problems, researchers global
have queriestowith spa
proposed
queries that focus on only some key–value pairs. Since then,
reduce resource consumption by replacing global queries with sparse queries that focus onmany related works bas
on this research idea have appeared, such as local attention,
only some key–value pairs. Since then, many related works based on this research idea deformable attention, a
expansive attention, but all of them are manually produced static
have appeared, such as local attention, deformable attention, and expansive attention, but patterns and conte
all of them areindependent sparsity. To
manually produced solve
static these problems,
patterns Lei Zhu et al. [30]sparsity.
and content-independent proposed Toa novel d
namic sparse attention: the Bi-Level Routing Attention, whose
solve these problems, Lei Zhu et al. [30] proposed a novel dynamic sparse attention: the workflow is shown in F
ure 4a.
Bi-Level Routing Attention, whose workflow is shown in Figure 4a.

(a) (b)
Figure 4.
Figure 4. (a) Structure of(a)
theStructure
Bi-Level of the Bi-Level
Routing Routing
Attention; (b) Attention; (b)the
Structure of Structure of block.
Biformer the Biformer block.

From Figure From

4a, it can be 4a,
Figure seen thatbe
it can theseen
input
thatfeature map
the input feature R H ×WX×C∈ is
X ∈ map R Hfirstly
×W × C
is firstly
HW
divided into S × S subregions, and each region contains S2 featureHW vectors. We change
vided into S × S subregions,
S2 × HW ×C and each region contains 2
feature vectors. We chan
the shape of X to obtain X r ∈ R S2 . Then, the feature vectorsS are linearly trans-
HW
formed to derive the three matrices, Q, K, and S 2 ×The
V. ×C calculation formulae are shown in
the shape of X to obtain X r ∈ R S
2
. Then, the feature vectors are linearly tran
Equations (11)–(13):
formed to derive the three matrices,r QQ , K , and V . The calculation formulae are show
Q=XW (11)
in Equations (11)–(13):

K = Xr W k Q = X rW Q (12) (1

r V
K = X rW k (1
V=XW (13)
Then, the attention relation from region to region V = X rW Vby constructing a directed
is obtained
(1
graph to locate its related regions
Then, the for a relation
attention given region. The specific
from region implementation
to region process
is obtained by is
constructing a
as follows: Q and
rectedV for each
graph toregion
locate are processed
its related by region
regions averaging
for a given to obtain
region. the region
The specific implementati
r ∈ RS2is
level Qr and Kprocess ×Cas
. Then, theQdotand
follows: V forofeach
product Qr and K r isare
region processedtoby
calculated region
obtain theaveraging
2
×C
obtain the region level Q r and K r ∈ R S . Then, the dot product of Q r and K r is c
2 2
culated to obtain the adjacency matrix Ar ∈ R S ×S , which is used to measure the int
region correlation, and the formula is shown in Equation (14):
r r r T
Sensors 2023, 23, 7190 10 of 27

2 2
adjacency matrix Ar ∈ RS ×S , which is used to measure the inter-region correlation, and
the formula is shown in Equation (14):

Ar = Qr ( K r ) T (14)

Next, Ar is pruned. The least relevant token in Ar is filtered out at the coarse-grained
level, and the top k most relevant regions in Ar are retained to obtain the routing index
2
matrix, I r ∈ N S ×k . The calculation formula is shown in Equation (15):

I r = topkIndex ( Ar ) (15)

Subsequently, token-to-token attention is used at the fine-grained level. For queries in

the region i, this attention is focused only on the k routing regions where I(ri,1) , I(ri,2) , . . . , I(ri,k)
are indexed and collects all the K and V tensors in these k regions to acquire K g and V g .
The calculated formula are shown in Equations (16)–(17):

K g = gather (K, I r ) (16)

V g = gather (V, I r ) (17)

Finally, the collected K g and V g are processed with attention, and a local context
enhancement term LCE(V ) is added to obtain the output tensor O. The formula is shown
in Equation (18):
O = Attention( Q, K g , V g ) + LCE(V ) (18)
The BiFormer block is designed based on Bi-Level Routing Attention, as shown in
Figure 4b. The DWConv in this block denotes deep separable convolution, which can reduce
the number of parameters and computation of the model. LN denotes layer normalization
processing, which can accelerate the training and improve the generalization ability of the
model. MLP denotes a multilayer perceptron, which further processes and adjusts the
attention weights to enhance the model’s attention to different features. The add symbol in
Figure 4b indicates connecting two feature vectors.
In this paper, the BiFormer block is introduced into the backbone network. On the one
hand, BiFormer can take into account the limited computing power and storage resources
of the UAV hardware platform. On the other hand, the dynamic attention mechanism
in this block can improve the model’s attention to the vital information of the object and
optimize the detection performance of the model. To make full use of the efficient attention
mechanism in this block, we use the BiFormer block between B3 and B4 of the model’s
backbone network, replacing the original C2f block.

3.3. Multiscale Feature Fusion Network

Poor detection of small objects is one of the challenges in object detection tasks in the
context of UAV aerial photography. In many existing works [39–42], detection scales are
added to the model to reduce the missed detection rate of small objects, which is an effective
improvement method. However, this approach can complicate the structure of the model
and increase the consumption of computational and storage resources. To mitigate this
problem, a feature processing block called FFNB is proposed in this paper, and a multiscale
feature fusion network is designed based on this block. The detection accuracy of small
objects is greatly improved while reducing the excessive consumption of resources.
Object detection tasks for UAV platforms are limited by computational resources, and
models with simple structure, low latency, and high data throughput are sought. Some
classical lightweight networks, such as MobileNet [43], ShuffleNet [44], and GhostNet [45],
are using deep convolution or group convolution to extract the spatial features of images.
Deep convolution reduces the input of feature dimensions by convolving the input images
grouped in feature dimensions, reducing the number of parameters while keeping the
Sensors 2023, 23, 7190 11 of 27

feature information largely unchanged. Group convolution can be seen as a sparse form of
traditional convolution, where the input channels are convolved one by one, which can be
used to reduce the model parameters and achieve the purpose of lightweight models. Most
Sensors 2023, 23, x FOR PEER REVIEW
of these lightweight models focus on reducing the number of floating-point operations 12 of 28
(FLOPs), and very little related work considers the low floating-point operations per second
(FLOPS) of the model. However, the reduction in model parameters does not translate
exactly intoper
operations an second
increase in the computational
(FLOPS) speed of the
of the model. However, the reduction
model. Therefore,
in model some work
parameters
using deep
does not convolution
translate orinto
exactly group convolution
an increase in computational
in the an attempt to design
speed lightweight
of the model.and fast
There-
neural network
fore, some workblocks,
using in some
deep cases, does
convolution ornot
groupspeed up the model
convolution in anoperation
attempt and even
to design
exacerbates
lightweight theandlatency.
fast neural network blocks, in some cases, does not speed up the model
For an input
operation and even feature h × latency.
of size the
exacerbates w × c, the required FLOPs using regular convolution
of size k × k
For an input feature of size h × w × cThe
are shown in Equation (19). , thec required
in Equation (19)using
FLOPs represents theconvolution
regular number of
channels of the input data.
of size k × k are shown in Equation (19). The c in Equation (19) represents the number
of channels of the input data.
FLOPsConv = h × w × k2 × c2 (19)
FLOPsConv = h × w × k 2 × c 2 (19)
The deep convolution kernel performs a sliding operation on the input channel space
The deep
to derive convolution
the output kernel
channel performs
features, a sliding
and the FLOPsoperation on the input are
for deep convolution channel space
calculated
to derive the output channel
as shown in Equation (20): features, and the FLOPs for deep convolution are calculated
as shown in Equation (20):
FLOPs DWConv = h × w × k2 ×
FLOPsDWConv = h × w × k 2 ×cc (20)
(20)
The
The depth
depth convolution
convolution computation
computation process
process is
is shown
shown in
in Figure
Figure5a.
5a.

(a) (b)
Figure 5.
Figure 5. Comparison
Comparison of
of DWConv
DWConv and
and PConv. (a) Structure
PConv. (a) diagram of
Structure diagram of deep
deep convolution;
convolution; (b)
(b) struc-
struc-
ture diagram of partial convolution.
ture diagram of partial convolution.

The popular
The populardeepdeepconvolution
convolutioneffectively
effectivelyreduces
reducesthethe parameters
parameters of of
thethe model.
model. ButBut
in
in practical application, depth convolution needs to be followed by additional
practical application, depth convolution needs to be followed by additional point-by-point point-by-
point convolution
convolution or otherorcomputational
other computational
costs tocosts to compensate
compensate for the reduction
for the reduction in accu-
in accuracy after
racy after the convolution operation. This introduces additional memory
the convolution operation. This introduces additional memory access costs and increases access costs and
increasesTo
latency. latency.
relieveTo therelieve
abovethe above problem,
problem, Chen et al. Chen
[29]etproposed
al. [29] proposed
a simple aand simple and
efficient
efficient convolution:
convolution: partial convolution
partial convolution (PConv). (PConv).
PConv PConv uses regular
uses regular convolution
convolution to per-
to perform a
form a convolution
convolution operation operation
on someon ofsome of the continuous
the continuous features features in the
in the input input channel,
channel, and the
and the remaining
remaining features arefeatures are processed
processed by identityby identitykeeping
mapping, mapping, thekeeping
channel the channel un-
unchanged. We
changed. We perform the convolution calculation for the first consecutive
perform the convolution calculation for the first consecutive feature with channel number feature with
cchannel
p in the number c p in the
input features, input features,
as shown in Figureas 5b,shown in Figure
and derive 5b, and derive
the formula the formula
for calculating the
FLOPs of PConv as shown in Equation (21):
for calculating the FLOPs of PConv as shown in Equation (21):
2 2
FLOPs
FLOPs = h=×
PConv
PConv h× ××k2k××c pc2p
ww (21)
(21)
If c pc pis 1/4
If is 1/4
ofof the
the number
number of of input
input feature
feature channels
channels c ,FLOPs
c, the the FLOPs of PConv
of PConv are 1/16
are only only
1/16
of theofconventional
the conventional convolution.
convolution. This convolution
This convolution reducesreduces the number
the number of memoryof memory
accesses
accesses
while while reducing
reducing the parameters,
the parameters, and efficiently
and efficiently extractsextracts the spatial
the spatial featuresfeatures of the
of the input
input information. Chen et al. proposed the FasterNet block based on PConv, which is a
module consisting of a PConv layer and two 1 1 convolutional layers connected sequen-
tially, as shown in Figure 6a. The add symbol in Figure 6a indicates connecting two feature
vectors. The overuse of normalization and activation layers may lead to a reduction in
 Sensors 2023, 23, 7190 12 of 27

information. Chen et al. proposed the FasterNet block based on PConv, which is a module
Sensors 2023, 23, x FOR PEER REVIEW
consisting of a PConv layer and two 1 × 1 convolutional layers connected sequentially,
13 of 28
as shown in Figure 6a. The add symbol in Figure 6a indicates connecting two feature
vectors. The overuse of normalization and activation layers may lead to a reduction in
feature diversity, which may affect the performance of the model. Therefore, the FasterNet
FasterNet block
block uses has a simpleand
normalization structure andlayers
activation a low only
number ofthe
after parameters for faster opera-
second convolutional layer.
tion.
FasterNet block has a simple structure and a low number of parameters for faster operation.

(a) (b)
Figure 6. Comparison
Figure 6. Comparison of FasterNet blockblock
of FasterNet and Focal FasterNet
and Focal block.block.
FasterNet (a) Structure diagramdiagram
(a) Structure of Fast- of
erNet block; block;
FasterNet (b) structure diagram
(b) structure of our of
diagram proposed module.
our proposed module.

In Inthisthispaper,
paper, we we revisit
revisitthethestructure
structureofofthe theFasterNet block.AA1 1× ×
FasterNetblock. 1 convolutional
1 convolutional
layer
layer is used
is used in this
in this block,block,
which which can reduce
can reduce the number
the number of parameters,
of parameters, speed up speed up the
the train-
ingtraining
and increaseand increase the nonlinear
the nonlinear fittingofability
fitting ability of theHowever,
the model. model. However,
the receptivethe receptive
field of
1 ×field of 1 × 1 convolution
1 convolution is relativelyissmallrelatively small
and lacks in and lacks in
acquiring acquiring
global global
features. It is features.
also con- It
is also considered that the FasterNet block uses only one
sidered that the FasterNet block uses only one shortcut connection, and the input featuresshortcut connection, and the
input features are convolved through three layers, in turn,
are convolved through three layers, in turn, which may lead to network degradation and which may lead to network
degradation
feature and feature
disappearance as thedisappearance
depth of the as the depth
model of thetomodel
continues deepen. continues
To solvetothe deepen.
aboveTo
solve thethe
problems, above
FFNB problems,
is proposedthe FFNB
in thisispaper
proposedbased in on
thisthe
paper based block,
FasterNet on the and
FasterNet block,
the struc-
and
ture is the
shown structure is shown
in Figure 6b. in Figure 6b.
First,
First, PConv
PConv is is
usedused toto replace
replace the two1 1× ×
thetwo 1 convolutional
1 convolutional layers
layers in in
thethe FasterNet
FasterNet
block, which improves the receptive field while making
block, which improves the receptive field while making the original module faster and the original module faster and
more efficient. Second, the residual concatenation is added
more efficient. Second, the residual concatenation is added to the last two convolutional to the last two convolutional
layers
layers in inthethe block
block to enrich
to enrich the the features
features of output
of the the output information,
information, reducereduce
the lossthe of
loss
ef- of
effective
fective features,
features, andand optimize
optimize the detection
the detection performance
performance of model.
of the the model.
Most of the current mainstream object detection models useconvolutional
Most of the current mainstream object detection models use convolutionalneural neuralnet-
works to extract object features. As the number of convolutions
networks to extract object features. As the number of convolutions increases, the semantic increases, the semantic
information
information of of
thethe input
input features
features gradually
gradually becomes
becomes richer,
richer, butbutthethe detailed
detailed features
features will
will
become fewer and fewer, which is one of the main reasons for the low detection accuracy of
become fewer and fewer, which is one of the main reasons for the low detection accuracy
of many
manyobject objectdetection
detection models
models for for
small objects.
small Although
objects. AlthoughYOLOv8YOLOv8uses auses
multiscale detection
a multiscale
method, it still cannot meet the detection needs of UAV aerial photography scenarios, which
detection method, it still cannot meet the detection needs of UAV aerial photography sce-
leads to the unsatisfactory detection accuracy of the model for small objects.
narios, which leads to the unsatisfactory detection accuracy of the model for small objects.
To improve the detection accuracy of small objects and consider the limited resources
To improve the detection accuracy of small objects and consider the limited resources
of the platform, this paper uses the efficient FFNB to design the feature fusion network. We
of the platform, this paper uses the efficient FFNB to design the feature fusion network.
add two new detection scales to the original three detection scales of YOLOv8 and fuse the
We add two new detection scales to the original three detection scales of YOLOv8 and
shallow information of B1 and B2, which are richer in location information. The FFNB is
fuse the shallow information of B1 and B2, which are richer in location information. The
added as a feature processing block between B1 and B2 of the backbone network, and then
FFNB is added as a feature processing block between B1 and B2 of the backbone network,
the original C2f block between B2 and B3 is replaced using this block. The introduced FFNB
and then the original C2f block between B2 and B3 is replaced using this block. The intro-
can ease the resource consumption caused by multiscale feature fusion. The improved
duced FFNB can ease the resource consumption caused by multiscale feature fusion. The
model achieves a five-scale detection, as shown in Figure 7, which effectively improves the
improved model achieves a five-scale detection, as shown in Figure 7, which effectively
detection performance of the model.
improves the detection performance of the model.
 Sensors 2023, 23, x FOR PEER REVIEW 14 of 28
Sensors 2023,
Sensors 23,23,
2023, x FOR
7190PEER REVIEW 14 of
1328
of 27

(a) (b)
(a) (b)
Figure 7. (a) Detection head of YOLOv8; (b) detection head of our proposed method.
Figure 7. (a)
Figure Detection
7. (a) head
Detection of YOLOv8;
head (b)(b)
of YOLOv8; detection head
detection of our
head proposed
of our method.
proposed method.
4. Experiments
4. Experiments
4. Experiments
4.1. Experiment Introduction
4.1. Experiment Introduction
4.1. Experiment Introduction
This section first introduces the dataset used in this paper, then introduces the exper-
This section first introduces the dataset used in this paper, then introduces the experi-
imental environment
This section and training
first introduces strategy,
the dataset andinfinally
used introduces
this paper, the evaluation
then introduces metrics
the exper-
mental environment and training strategy, and finally introduces the evaluation metrics
related
imental to the experimental
environment and training results.
strategy, and finally introduces the evaluation metrics
related to the experimental results.
related to the experimental results.
4.1.1. Dataset
4.1.1. Dataset
4.1.1. Dataset
The VisDrone2019
The VisDrone2019 dataset dataset[46][46]isisone
oneofof
thethe
mainstream
mainstream UAVUAV aerial photography
aerial photography da-
tasets,
The which was
VisDrone2019 collected
dataset and developed
[46] is one of by
the Tianjin University
mainstream UAV and
aerial
datasets, which was collected and developed by Tianjin University and the data mining the data mining
photography team
da-
AISKYEYE.
tasets,
team which
AISKYEYE.Thecollected
was dataset is
and
The dataset framed in more
developed
is framed than athan
bymore
in Tianjin dozen different
University
a dozen and cities in China
the data
different mining
cities and
team
in China uses
and
a variety
AISKYEYE. of
The UAVs for
dataset multi-angle,
is framed in multiscene,
more than a and
dozen multi-task
different photography,
cities
uses a variety of UAVs for multi-angle, multiscene, and multi-task photography, making thein China making
and usesthe
dataset
a variety ofvery
UAVs informative,
for including
multi-angle, the
multiscene,category
and of detection
multi-task objects
photography,
dataset very informative, including the category of detection objects (monotonous and rich), (monotonous
making and
the
rich),
dataset
the the number
very
number of detection
informative,
of detection objectsobjects
including the (fewer
(fewer category
and moreandofnumerous),
more numerous),
detection objects the distribution
(monotonous
the distribution of de-
and
of detection
tection
rich), the objects
number (sparse
of and
detection dense),
objects and
(fewer the light
and moreintensity (day
numerous), and
the
objects (sparse and dense), and the light intensity (day and night). Some representative night). Some
distribution of repre-
de-
sentative
tection
images images
objects in
in the(sparse the
datasetand dataset are
dense),inand
are shown shown in
the light
Figure 8. Figure 8.
intensity (day and night). Some repre-
sentative images in the dataset are shown in Figure 8.

(a) (b) (c)

(a) (b) (c)

(d) (e) (f)

(d) (e) (f)
Figure 8. Cont.
Sensors 2023, 23, x FOR PEER REVIEW 15 of 28

Sensors 2023,
Sensors 2023, 23,
23, 7190
x FOR PEER REVIEW 15
14of 28
of 27

(g) (h) (i)

Figure 8. Some representative images from the VisDrone2019 dataset. (a) Sparse object distribution;
(g) (b) Dense object distribution; (c) Low(h) number of objects; (d) High number of objects;
(i) (e) Many types
ofFigure
objects; (f) Objects
8. Some are very small; (g) Morning; (h) Evening; (i) Night.
Figure 8. Some representative
representative images
images from
from the
the VisDrone2019 dataset. (a)
VisDrone2019 dataset. (a) Sparse
Sparse object
object distribution;
distribution;
(b) Dense
(b) Dense object
object distribution;
distribution; (c)
(c) Low
Low number
number ofof objects;
objects; (d)
(d) High
High number
number ofof objects;
objects; (e)
(e) Many
Many types
types
VisDrone2019
of objects; (f) Objectscontains 10 different
are very small; types(h)
(g) Morning; of Evening;
objects, such as pedestrians, cars, bicy-
(i) Night.
of objects; (f) Objects are very small; (g) Morning; (h) Evening; (i) Night.
cles, etc. Figure 9 shows the information related to the manual labeling of the objects in
VisDrone2019
this dataset. We present
VisDrone2019 contains
contains 10
10 different
the subfigures
different in types
Figureof
types of9objects,
in ordersuch
objects, fromas
such asleftpedestrians,
to right and
pedestrians, cars,
topbicy-
cars, to
bicy-
bottom.
cles, etc.
cles, The
etc. first9subfigure
Figure
Figure 9shows
showsthe shows
the the number
information
information oftoobjects
related
related to
thethe of each
manual
manual type
labeling
labeling in
of the of dataset,
the inand
the objects
objects in
this
indicates
dataset. thatpresent
this dataset.
We the present
We objects are
the dominated
the subfiguressubfigures byFigure
in
in Figure cars and
9 inpedestrians.
9 in order orderleft
from from The
left to
to right second
and topsubfigure
right and top to
to bottom.
shows
bottom.
The theThe
first sizefirst
of the
subfigure object
subfigure
shows thebounding
shows
number boxes
the of
number inofthe
objects dataset,
ofobjects
each and
of each
type in the the
type coordinates
in theand
dataset, of the
dataset, and
indicates
centers
indicatesof all
thatobject
the boxes
objects are
are fixed at
dominated one point.
by cars Theandsize of the
pedestrians.
that the objects are dominated by cars and pedestrians. The second subfigure shows the object The bounding
second box size
subfigure
shows
shows
size ofthat
thethesizedataset
object ofbounding
thecontains
objectboxesbounding
a large
in the boxes
number
dataset,in the dataset,
of and
small-area and the
objects.
the coordinates The
ofcoordinates
third
the centers of
subfigureofthe
all
isobject
the distribution
centers of allare
boxes of the
object
fixed atcoordinates
boxes oneare fixedThe
point. of
atthe
one
sizecenter
point.
of thepoints
The
object ofbounding
size the object
of the objectbounding
box bounding
size boxes,
shows box
thatand
size
the
itdataset
can bethat
shows seenthe
containsthata the
dataset
large center
contains
number points of the
a large
of objects
number
small-area ofare
objects. mainlythirdconcentrated
small-area
The objects.
subfigure The in the
third
is the middle
subfigure
distribution
is the
and
of the distribution
right below theofarea
coordinates the coordinates
center imageofof
of the points the
thecenter
data. The
object points
fourth
boundingof theboxes,
subfigureobject abounding
isand scatter beboxes,
it can plot seen and
of the
that
corresponding
it can
the be seen
center pointswidth
thatoftheand
the height
center
objects of the
points
are ofobject
mainly bounding
theconcentrated
objects box,
are mainly
in thewith the darkest
concentrated
middle and right incolor
the at the
middle
below the
bottom
area ofleft
and right of
thebelow the the
image plot.
data. ItThe
area further
of fourth shows
the image that the
data.
subfigure is acurrent
The dataset
fourth plot
scatter subfigure
of theis is
dominated
a scatter plot
corresponding by small
of the
width
corresponding
objects.
and height of the width
object and height ofbox,
bounding thewith
objectthebounding box, with
darkest color at thethe bottomdarkest leftcolor
of theatplot.
the
It furtherleft
bottom shows
of thethat theItcurrent
plot. furtherdataset
showsisthat dominated
the current by small objects.
dataset is dominated by small
objects.

Figure 9. Information about the manually labeling of objects in VisDrone2019 dataset.

Sensors 2023, 23, 7190 15 of 27

From the above introduction and analysis of the VisDrone2019 dataset, it can be
surmised that the dataset contains a large number of small objects, and they exist mainly
in a dense and uneven distribution. Compared with the dataset of traditional computer
vision tasks, this dataset is a large UAV dataset with multiple scales, scenes, and angles,
which is more challenging than the general computer vision tasks.

4.1.2. Experimental Environment and Training Strategies

The hardware platform and environmental parameters used in the experimental
training phase are shown in Table 2.

Table 2. Training environment and hardware platform parameters table.

Parameters Configuration
CPU i5-12490F
GPU NVIDIA GeForce RTX 3060
GPU memory size 12G
Operating systems Win 10
Deep learning architecture Pytorch1.9.2 + Cuda11.4 + cudnn11.4

To facilitate flexible deployment on hardware devices in various application scenarios,

the YOLOv8 model has been adapted to generate five different scaled models by adjusting
two parameters: width and depth. These models are referred to as YOLOv8n, YOLOv8s,
YOLOv8m, YOLOv8l, and YOLOv8x. The parameters and resource consumption of the
five models increase sequentially, and the detection performance becomes better and better.
The width, depth, and maximum number of channels corresponding to these five models
are shown in Table 3.

Table 3. Parameters corresponding to different sizes of YOLOv8.

Model Depth Width Max Channels

YOLOv8n 0.33 0.25 1024
YOLOv8s 0.33 0.50 1024
YOLOv8m 0.67 0.75 768
YOLOv8l 1.00 1.00 512
YOLOv8x 1.00 1.25 512

To better study and improve the models, we choose YOLOv8s as the baseline model.
Some of the key parameter settings during model training are shown in Table 4.

Table 4. Some key parameters set during model training.

Parameters Setup
Epochs 200
Momentum 0.932
Initial learning rate 0.01
Final learning rate 0.0001
Weight decay 0.0005
Batch size 4
δ (WIoU v3) 1.9
α (WIoU v3) 3
Input image size 640 × 640
Optimizer SGD
Data enhancement strategy Mosaic

Under the dataset division rules of the VisDrone 2019 Challenge, this paper divides
the dataset into training sets (6471 images), testing sets (1610 images), and validation sets
Sensors 2023, 23, 7190 16 of 27

(548 images). To accelerate the model convergence, Mosaic data enhancement is turned off
in the last 10 epochs of the training process.

4.1.3. Evaluation Indicators

To test the detection performance of our proposed improved model, we use precision,
recall, mAP0.5, mAP0.5:0.95, number of model parameters, model size, and detection
speed as evaluation metrics. The following parameters are used in the formulae for some
of the above evaluation metrics: TP (predicted as a positive sample and actually as a
positive sample as well), FP (predicted as a positive sample, though it is actually a negative
sample), and FN (predicted as a negative sample, though it is actually a positive sample).
Intersection over Union (IoU) represents the ratio of intersection and concatenation between
the bounding box and the true box.
Precision is the ratio of the number of positive samples predicted by the model to the
number of all detected samples and is calculated as shown in Equation (22):

TP
Precision = (22)
TP + FP
Recall is the ratio of the number of positive samples correctly predicted by the model
to the number of positive samples that actually appeared. Recall is calculated as shown
in Equation (23):
TP
Recall = (23)
TP + FN
The average precision (AP) is equal to the area under the precision–recall curve and is
calculated as shown in Equation (24):

Z1
AP = Precision(Recall)d( Recall) (24)
0

Mean average precision (mAP) is the result obtained by the weighted average of AP
values of all sample categories, which is used to measure the detection performance of the
model in all categories, and the formula is shown in Equation (25):

1 N
N ∑ i =1
mAP = APi (25)

The APi in Equation (25) denotes the AP value with category index value i, and N
denotes the number of categories of the samples in the training dataset (in this paper,
N is 10). mAP0.5 denotes the average accuracy when the IoU of the detection model is set
to 0.5, and mAP0.5:0.95 denotes the average accuracy when the IoU of the detection model
is set from 0.5 to 0.95 (with values taken at intervals of 0.5).

4.2. Experiment Results

4.2.1. Comparative Experiment of Loss Function
To verify the superiority of introducing WIoU v3, we conducted comparison exper-
iments on YOLOv8s using WIoU v3 and some mainstream loss functions, keeping the
other training conditions consistent. The experimental results are shown in Table 5. The
experimental results show that the model achieves the best detection performance when
using WIoU v3 as the bounding box regression loss. In addition, the model’s mAP50 when
using WIoU v3 is 0.7% higher than when using CIoU, demonstrating the effectiveness of
introducing WIoU v3.
Sensors 2023, 23, x FOR PEER REVIEW 18 of 28

using WIoU v3 is 0.7% higher than when using CIoU, demonstrating the effectiveness of
Sensors 2023, 23, 7190 17 of 27
introducing WIoU v3.

Table 5. Comparison of detection results for different loss functions introduced by YOLOv8s. (The
5. Comparison
Tabledata
bold of detection
in the table indicate results
the best for different loss functions introduced by YOLOv8s. (The
results).
bold data in the table indicate the best results).
Metrics Precision/% Recall/% mAP0.5/% mAP0.5:0.95/%
Metrics
CIoU Precision/%
50.9 Recall/%
38.2 mAP0.5/%
39.3 mAP0.5:0.95/%
23.5
DIoU [35]
CIoU 51.0
50.9 38.3
38.2 39.5
39.3 23.6
23.5
DIoU [47]
GIoU [35] 51.0
50.3 38.3
38.4 39.5
39.6 23.6
23.6
GIoU
EIoU[47] 50.3
49.1 38.4
38.0 39.6
38.7 23.6
23.4
EIoU 49.1 38.0 38.7 23.4
SIoU
SIoU 51.5
51.5 38.5
38.5 39.4
39.4 23.4
23.4
WIoU
WIoU v1v1 50.1
50.1 38.5
38.5 39.3
39.3 23.3
23.3
WIoU
WIoU v2v2 50.6
50.6 38.4
38.4 39.3
39.3 23.2
23.2
WIoU
WIoU v3 v3 51.3
51.3 38.6
38.6 40.0
40.0 23.6
23.6

4.2.2. Comparison
4.2.2. Comparisonwith withYOLOv8
YOLOv8
To demonstrate
To demonstratethe thedetection performance
detection performance improvement
improvement effecteffect
of theofimproved model,
the improved
we conducted
model, comparison
we conducted experiments
comparison betweenbetween
experiments the improved model and
the improved modelthe and
baseline
the
model YOLOv8s. Table 6 shows the AP values for each category and
baseline model YOLOv8s. Table 6 shows the AP values for each category and the mAP0.5the mAP0.5 values
for all for
values categories for thefor
all categories improved model
the improved and YOLOv8s.
model and YOLOv8s.As shown
As shown in the comparison
in the compar-
results in Table 6, the mAP value of the improved model is improved
ison results in Table 6, the mAP value of the improved model is improved by 7.7%. by 7.7%. TheTheAP
values of all the categories are improved to different degrees, and the AP
AP values of all the categories are improved to different degrees, and the AP values of values of three
categories
three (pedestrian,
categories people,
(pedestrian, and motor)
people, are improved
and motor) by moreby
are improved than 10%.
more This10%.
than indicates
This
indicates that the improved model can effectively increase the detection accuracy ofobjects
that the improved model can effectively increase the detection accuracy of small small
and improve
objects the detection
and improve performance.
the detection performance.
Table 6.
Table Comparisonof
6. Comparison ofthe
theproposed
proposedimproved
improvedmodel
modelandandYOLOv8s
YOLOv8sdetection
detectionaccuracy.
accuracy. (The
(Thebold
bold
data in the table indicate the best results. All data units in the table are in percent.)
data in the table indicate the best results. All data units in the table are in percent.)

Models
Models Pedestrian People People
Pedestrian Bicycle Bicycle
Car Van
Car Truck
Van Tricycle
Truck Awning-Tricycle
Tricycle Awning-TricycleBus BusMotor
Motor mAP
mAP
44.0 36.5 28.1 15.9 57.0
YOLOv8s
YOLOv8s 42.7 32.0
42.7 12.4
32.0 79.1 12.4 79.1 44.0 36.5 28.1 15.9 57.0 44.9 44.9 39.3
39.3
Ours
Ours 56.8 56.8 44.9 44.9
18.8 18.8
85.8 50.8
85.8 39.0
50.8 33.3
39.0 33.3 19.7 19.7 64.3 64.3 56.2 56.2 47.0
47.0

Figure 10 shows the change curves of some important evaluation metrics of our pro-
Figure 10 shows the change curves of some important evaluation metrics of our
posed model
proposed modelandand
YOLOv8s
YOLOv8s during
duringthethe
training process.
training From
process. FromFigure
Figure10,
10,we
wecan
cansee
seethat
that
our
our proposed
proposed model
model outperforms
outperforms YOLOv8s
YOLOv8sin in three
three detection
detection metrics:
metrics: precision,
precision,recall,
recall,
and
andmAP0.5
mAP0.5after
afterabout
about15 15epochs
epochsofoftraining
trainingfrom
fromthethebeginning,
beginning,and
and our
ourmodel
modelstarts
startsto
to
stabilize
stabilize after about 50 epochs of training. Compared with YOLOv8s, our method isfaster
after about 50 epochs of training. Compared with YOLOv8s, our method is faster
to
totrain
train and
and better
betterto
to detect.
detect.

(a) (b) (c)

Figure
Figure 10.
10. (a)
(a)Training
Trainingcurve
curveof
ofUAV-YOLOv8s
UAV-YOLOv8sand andYOLOv8s
YOLOv8sin inmAP;
mAP;(b)
(b)training
trainingcurve
curveofofUAV-
UAV-
YOLOv8s and YOLOv8s in precision; (c) training curve of UAV-YOLOv8s and YOLOv8s
YOLOv8s and YOLOv8s in precision; (c) training curve of UAV-YOLOv8s and YOLOv8s in in recall.
recall.

To further demonstrate the effectiveness of the proposed method, we compared the

proposed model with some different sizes of YOLOv8 (YOLOv8n, YOLOv8s, YOLOv8m,
and YOLOv8l) on the VisDrone2019 dataset. The experimental results are shown in Table 7.
Sensors 2023, 23, 7190 18 of 27

According to the data in Table 7, compared with other models, the improved model has
the highest values of the three evaluation indexes Recall, mAP0.5, and mAP0.5:0.95, and
the detection performance is better than the model with a larger size than itself. From the
experimental results, it can be seen that our proposed five-scale detection structure can
improve the detection accuracy of small objects. In addition, the low-computing-power
attention mechanism BiFormer, introduced by us, improves the detection performance of
the model without consuming too many resources.

Table 7. Comparative experimental results of the proposed model with four different sizes of YOLOv8.
(The bold data in the table indicate the best results.)

Models Precision/% Recall/% mAP0.5/% mAP0.5:0.95/% Model Size/MB Detection Time/ms Parameter/106
YOLOv8n 43.8 33.0 33.3 19.3 6.6 4.2 3.0
YOLOv8s 50.9 38.2 39.3 23.5 22.5 7.7 11.1
YOLOv8m 56.0 42.5 44.6 27.1 49.6 16.6 25.9
YOLOv8l 57.5 44.3 46.5 28.7 83.5 25.6 43.7
Ours 54.4 45.6 47.0 29.2 21.5 19.5 10.3

4.2.3. Adding BiFormer Block

To obtain the best performance and facilitate the subsequent ablation experiments
after adding the BiFormer block to the model, the following comparison experiments were
conducted in this study. We used YOLOv8s after the introduction of WIoU v3 as the
baseline model and replaced the C2f module with the BiFormer block in different layers
of the backbone network, and attained the experimental results shown in Table 8. The +
in Table 8 indicates that the BiFormer block is added to the baseline model. B3-BiFormer
indicates that the C2f module between layer B3 and layer B4 is replaced using the BiFormer
block, and B4-BiFormer indicates that the C2f module between layer B4 and layer B5 is
replaced using the BiFormer block. According to the experimental results in Table 8, the
best detection performance is achieved when the C2f module between layers B4 and B5
of the baseline model is replaced by the BiFormer module. Compared with the baseline
model, mAP0.5 is improved by 0.5%. When adding the BiFormer block, the model can
better focus on the essential information in the input features.

Table 8. Experimental results after introducing the BiFormer block at different layers of the backbone
network of the baseline model. (The bold data in the table indicate the best results.)

Model Precision/% Recall/% mAP0.5/% mAP0.5:0.95/%

Baseline 50.7 38.7 40.0 23.6
+B3-BiFormer 49.7 38.6 39.7 23.5
+B4-BiFormer 50.6 39.2 40.5 24.1
+B3-BiFormer+
50.3 39.1 40.2 24.0
B4-BiFormer

4.2.4. Ablation Experiments

To verify the effectiveness of each improvement strategy proposed in this paper, we
performed ablation experiments on the baseline model using the VisDrone2019 dataset, and
the experimental results are shown in Table 9. B2-FFNB and B1-FFNB in Table 9 indicate
the use of FFNB module-based detection layers that fuse the shallow features of B2 and B1
√
layers, respectively. indicates that this improved strategy was used.
Sensors 2023, 23, 7190 19 of 27

Table 9. Detection results after the introduction of different improvement strategies. (The bold data
in the table indicate the best results.)

Model Detection
Baseline WIoU v3 BiFormer B1-FFNB B2-FFNB Precision/% Recall/% mAP0.5/% mAP0.5:0.95/% Parameter/106
Size/MB Time/ms

√ 50.9 38.2 39.3 23.5 22.5 7.7 11.1

√ √ 50.7 38.7 40.0 23.6 22.5 7.2 11.1
YOLOv8s √ √ √ 50.6 39.2 40.5 24.1 22.1 7.8 10.9
√ √ √ √ 55.8 42.8 44.8 27.2 21.6 11.2 10.6
54.4 45.6 47.0 29.2 21.5 19.5 10.3

The experimental results in Table 9 show that each improvement strategy improved the
detection performance to different degrees when applied to the baseline model. WIoU v3 is
introduced in the prediction box regression loss. WIoU v3 improves the localization ability
of the model using a smarter sample allocation strategy, resulting in improved mAP50 by
0.7%. BiFormer was introduced into the backbone network, replacing the C2f module in the
original model. The efficient attention mechanism in BiFormer improves the attention to
the key information in the feature map, which increases the mAP50 by 0.5%. The relatively
simple structure of the BiFormer module reduces the size of the model by 0.4 MB and the
parameters by about 0.2 M (million). The B2-FFNB detection layer was added to the baseline
model, increasing the mAP by 4.3%. The currently used dataset contains a large number of
small targets, so adding a detection scale to fuse the shallow (B2) feature information can
effectively reduce the missed detection rate of small objects. Our proposed efficient and fast
FFNB module alleviates the computational and memory burden of feature fusion and reduces
the parameters of the model. The B1-FFNB detection layer was added to the baseline model,
resulting in the growth of mAP by 2.2%. The model fuses the B1 layer features with richer
detail information, which makes the shallow and deep information fully fused and further
reduces the missed detection rate of small objects. Our proposed feature fusion network
effectively improves the detection performance of the model.
The average detection accuracy of the improved model is increased by 7.7%, most
of the detection metrics are effectively improved, and the size and number of parameters
of the model are lower than those of the baseline model. However, the addition of two
detection layers causes the model structure to become complex and the inference time to
become longer.

4.2.5. Comparison Experiments

To show the superiority and effectiveness of the improved algorithm proposed in this
paper, we conducted two sets of comparison experiments. In the first set of comparison
experiments, we compared the proposed model with some YOLO series algorithms. In
the second set of comparison experiments, we compared the proposed model with other
excellent models.
The YOLO family of algorithms used in this paper are as follows: YOLOv3 [10], which
uses multiscale detection for the first time, and its lightweight version, YOLOv3-tiny;
YOLOv4 [11] uses the idea of CSPNet [48] to construct a new backbone network struc-
ture called CSPDarknet53, which is lightweight while maintaining the detection accuracy;
YOLOv5 uses Mosaic data augmentation to increase the model training speed and detection
accuracy. In addition, YOLOv5 also uses the Focus structure, which reduces the compu-
tation and parameter count of the model while maintaining the detection performance.
YOLOv7 uses the efficient network architecture ELAN [13]. The results of the comparison
experiments are shown in Table 10.
Sensors 2023, 23, 7190 20 of 27

Table 10. Detection results of some YOLO series models and the proposed model. (The bold data in
the table indicate the best results).

Model Detection
Models Precision/% Recall/% mAP0.5/% mAP0.5:0.95/% Parameter/106
Size/MB Time/ms
YOLOv3 54 43.6 41.9 23.3 213 18.3 103.7
YOLOv3-Tiny 38.2 24.8 23.8 13.3 24.4 2.9 12.1
YOLOv4 36 48.6 42.1 25.7 245 25.3 64.4
YOLOv5s 46.4 34.6 34.4 19 14 12.0 7.2
YOLOv7 51.4 42.1 39.9 21.6 72 1.7 64
YOLOv8s 50.9 38.2 39.3 23.5 22.5 7.7 11.1
UAV-YOLOv8s 54.4 45.6 47.0 29.2 21.5 19.5 10.3

According to the experimental results in Table 10, the earlier YOLO series algorithms,
such as YOLOv3 and YOLOv4, which have complex structures and large parameter counts,
are not conducive to deployment on unmanned platforms and have relatively low detection
accuracy. YOLOv3-Tiny achieves a lightweight model but loses a large portion of the
detection accuracy. YOLOv5s and YOLOv8s have smaller model sizes, fewer parameters,
and better detection performance than previous YOLO versions. However, both models
use three-scale detection structures, which cannot fulfill the detection needs of a high
proportion of small objects, and therefore result in lower model detection accuracy than
the proposed method in this paper at the same size. Compared with UAV-YOLOv8s,
YOLOv7 has an advantage in detection speed, but it is deficient in model size and detection
performance. In the comparative experimental results, UAV-YOLOv8s has the highest
average detection accuracy and the best overall detection performance. The inference time
of the improved model increases, but real-time detection can still be achieved.
In this study, a comparative experiment was conducted to evaluate the performance
of UAV-YOLOv8s and other mainstream models. The following is a brief introduction
to the models used in the comparative experiments: Faster R-CNN [7] optimizes the
problems of high computational and structural complexity of R-CNN; Cascade R-CNN [49]
proposes a multilevel detection architecture based on R-CNN; RetinaNet [50] proposes
Focal loss; CenterNet [51] proposes anchorless frame detection; FSAF [52] algorithm solves
the drawbacks of heuristic feature selection; ATSS [53] proposes an adaptive training
sample selection mechanism. The experimental results are shown in Table 11.

Table 11. Detection results of the classical model and the proposed model. (The bold data in the table
indicate the best results.)

Models AP0.5/% AP0.75/% AP0.5:0.95/%

Faster R-CNN [7] 37.2 22.8 21.9
RetinaNet [50] 19.1 10.5 10.6
Cascade R-CNN [49] 39.1 26.2 24.3
CenterNet [51] 33.7 18.0 18.8
FSAF [52] 36.5 20.6 20.9
ATSS [53] 36.4 23.1 22.3
UAV-YOLOv8s 47.0 30.6 29.2

According to the comparative experimental results in Table 11, the model proposed in
this paper has the best detection performance compared to other excellent models. Faster
R-CNN, as a two-stage detection algorithm, has a slower detection speed compared to
one-stage algorithms Moreover, the resolution of the feature map extracted by the backbone
network is small, so it is relatively poor in accuracy for small-object detection. RetinaNet
and ATSS focus on how to determine the positive and negative samples. CenterNet removes
the computational effort due to the use of anchors, but there are situations such as dense
scenes and occlusion of small objects that can lead to some items being missed if multiple
Sensors 2023, 23, x FOR PEER REVIEW 22 of 28
Sensors 2023, 23, 7190 21 of 27

proposed with Cascade R-CNN improves the overall detection performance of the model
but increases
prediction computational
centroids complexity
are overlapping. Theand trainingdetection
multilevel difficulty.architecture
Both RetinaNet and FSAF
proposed with
Cascade
use R-CNN
multiscale improves
feature fusionthetooverall
considerdetection
objects performance of theHowever,
at different scales. model butinincreases
the con-
computational
text complexity
of visual tasks and training
where small difficulty.
objects are Both RetinaNet
the primary and FSAF
detection objects, usetwo
these multiscale
model
feature fusion
structures to consider
cannot address the objects at different
detection needs.scales. However,
This results in bothin the contextand
RetinaNet of visual
FSAF
tasks where small objects are the primary detection objects,
having poorer detection results than the methods proposed in this paper. these two model structures
cannot address the detection
Summarizing the resultsneeds.
of theThis results
above in both
two sets RetinaNet and
of comparison FSAF having
experiments, thepoorer
UAV-
detection results
YOLOv8s thaninthe
proposed methods
this paper proposed
has betterindetection
this paper.performance compared to other
Summarizing
models. Our proposedthe results of thefeature
multiscale above two setsnetwork
fusion of comparison
achieves experiments, the UAV-
five-scale detection.
YOLOv8s proposed
This structure in this paperdetection
for small-object has betterhas
detection performanceover
some advantages compared to other
the models models.
involved
Our proposed multiscale feature fusion network achieves five-scale detection.
in the above comparison experiments, so our detection results are better than the other This structure
for small-object
models. detection
In addition, has some advantages
we introduced WIoU v3 and overBiFormer
the models involved
in the in the
baseline above
model com-
to opti-
parison experiments, so our detection results are better than the other models.
mize the model’s localization ability and noise suppression. The improvement strategy In addition, we
introduced WIoU v3 and BiFormer in the baseline model to optimize the model’s
we introduced takes into account the consumption of resources and achieves better detec- localization
ability
tion and noise suppression. The improvement strategy we introduced takes into account
results.
the consumption of resources and achieves better detection results.
4.3. Visualization Analysis
4.3. Visualization Analysis
Deep learning models are characterized by poor interpretability, which somehow
Deep learning models are characterized by poor interpretability, which somehow
hinders the development and application of deep learning. To chart the detection effect of
hinders the development and application of deep learning. To chart the detection effect of
the model proposed in this paper intuitively and conveniently, we conducted comparative
the model proposed in this paper intuitively and conveniently, we conducted comparative
experiments to analyze the detection performance of the model from three perspectives:
experiments to analyze the detection performance of the model from three perspectives:
confusion matrix, model inference results and heat map. Finally, to verify the generaliza-
confusion matrix, model inference results and heat map. Finally, to verify the generalizabil-
bility of our method, we conducted inference experiments on self-made data.
ity of our method, we conducted inference experiments on self-made data.
To
To visualize
visualize the
the ability
abilityofofour
ourmethod
methodtoto predict thethe
predict object categories,
object we we
categories, plotted the
plotted
confusion matrices of UAV-YOLOv8s and YOLOv8s, as shown in Figure
the confusion matrices of UAV-YOLOv8s and YOLOv8s, as shown in Figure 11. The 11. The rows and
columns
rows andof the confusion
columns matrices represent
of the confusion the true and
matrices represent predicted
the true categories,
and predicted respec-
categories,
tively, and the values in the diagonal region indicate the proportion of correctly
respectively, and the values in the diagonal region indicate the proportion of correctly predicted
categories, while the values
predicted categories, whileinthethevalues
other regions indicate
in the other the proportion
regions of incorrectly
indicate the proportionpre-
of
dicted categories.
incorrectly predicted categories.

(a) (b)
Figure 11. (a)
Figure 11. (a) Confusion
Confusion matrix
matrix plot
plot of
of YOLOv8s; (b) confusion
YOLOv8s; (b) confusion matrix
matrix plot
plot of
of our
our model.
model.

As can be seen from Figure 11, the diagonal

diagonal region
region of
of the
the confusion
confusion matrix
matrixinin UAV-
UAV-
YOLOv8s is darker in color than YOLOv8s, indicating that the ability of our model to
correctly predict
predict the
theobject
objectcategory
categoryhas
hasbeen
beenenhanced.
enhanced.However,
However, forfor small
small objects
objects such
such as
as
the the bicycle,
bicycle, tricycle,
tricycle, and awning-tricycle,
and awning-tricycle, the proportion
the proportion of objects
of objects judged judged as
as background is
higher, implying that a large portion of these categories are missed during the detection
Sensors 2023, 23, x FOR PEER REVIEW 23 of 28

Sensors 2023, 23, 7190 22 of 27

background is higher, implying that a large portion of these categories are missed during
the detection
process. process. The
The improved improved
model reducesmodel reduces
the missed the missed
detection detection
rate for rate for these
these categories, but
categories, but the percentage of them being correctly predicted is still
the percentage of them being correctly predicted is still low. The bicycle category low. The bicycle of
category of transportation
transportation is small andisusually
small and usually
exists existsand
in a dense in aoccluded
dense and occluded
form, makingform, mak-
it difficult
ing it detected
to be difficult toin be detected in environments
environments with complex backgrounds.
with complex backgrounds.
To visually demonstrate
To visually demonstrate the detection effect
the detection of our
effect of our method,
method, inference
inference experiments
experiments
using
using UAV-YOLOv8s,
UAV-YOLOv8s, YOLOv8s
YOLOv8s and and YOLOv5s
YOLOv5s were were performed
performed in in this
this study.
study. We
We chose
chose
four representative scenarios, urban roads, public facility places, market intersections,
four representative scenarios, urban roads, public facility places, market intersections, and
traffic intersections, as the experimental data. These scenarios contain numerous
and traffic intersections, as the experimental data. These scenarios contain numerous and and di-
verse
diverse small objects, which are suitable for inference experiments. The detection results of
small objects, which are suitable for inference experiments. The detection results of
the three models are shown in Figure 12.

(a) YOLOv5s (b) YOLOv8s (c) UAV-YOLOv8s

Figure
Figure 12.
12. Inference
Inference results
results of
of three
three different
different models
models on
on VisDrone2019
VisDrone2019 dataset.
dataset. (a)
(a) Inference
Inference results
results
of YOLOv5s; (b) inference results of YOLOv8s; (c) inference results of our model.
of YOLOv5s; (b) inference results of YOLOv8s; (c) inference results of our model.

As shown in the results in Figure 12, compared with YOLOv8s and YOLOv5s, the
method proposed
proposed ininthis
thispaper
paperhas
hasthe
the optimal
optimal detection
detection accuracy
accuracy for for objects
objects at far
at the theend
far
end of the field of view. In addition, our method improves the model’s leakage detection
of the field of view. In addition, our method improves the model’s leakage detection rate
for occluded and dense objects, which effectively improves the detection performance.
Sensors 2023, 23, x FOR PEER REVIEW 24 of 28

Sensors 2023, 23, 7190 23 of 27

rate for occluded and dense objects, which effectively improves the detection perfor-
mance.
Gradient-weighted class activation mapping (Grade CAM)
Gradient-weighted CAM) waswas utilized
utilized to
to generate
generate
heat maps
heat maps for
for YOLOv8s
YOLOv8s and and UAV-YOLOv8s
UAV-YOLOv8s[54]. [54]. The
The heat
heat maps
maps visually
visually and
and easily
easily reflect
reflect
which areas
which areasofof the
the feature
featuremap
map the
the model
model is is focusing
focusing on.on. The
The gradient
gradient values
values are
are obtained
obtained
by backpropagation
by backpropagation of of the
the confidence
confidence of of the
the model
model output
output categories
categories byby Grade
Grade CAM,
CAM, andand
the pixels
the pixels with
with higher
higher gradients
gradients in
in the
the feature
feature maps
maps areare represented
represented by by deeper
deeper shades
shades ofof
red ininthe
red theheatmaps.
heatmaps. Conversely,
Conversely, the the pixels
pixels withwith
lowerlower gradients
gradients are represented
are represented by deeper by
deeper of
shades shades
blue. of blue.
The The experimental
experimental results
results are shownare in
shown
Figure in13.
Figure
From13. From13,
Figure Figure
we can13,
we can surmise
surmise that YOLOv8s
that YOLOv8s pays poorpays poor attention
attention to smallto small and
objects objects and is insensitive
is insensitive to distantto
objects. The model
distant objects. The proposed in this paper
model proposed in this has
paperbetter
has suppression of background
better suppression noise
of background
and
noisepays
andmore
pays attention to small
more attention objects.
to small The model’s
objects. The model’sattention is more
attention focused
is more on the
focused on
center point
the center of the
point ofobject, which
the object, makes
which the predicted
makes bounding
the predicted box more
bounding accurate
box more and thus
accurate and
improves
thus improvesthe overall detection
the overall performance
detection performance of theofmodel.
the model.

(a) original images (b) YOLOv8s (c) UAV-YOLOv8s

Figure 13.
Figure 13. (a)
(a) Original
Original images;
images; (b)
(b) heat
heat maps
maps of
of YOLOv8s;
YOLOv8s; (c)
(c) heat
heat maps
maps of
of our
our model.
model.

To validate
To validate the
the generalizability
generalizability of of our
our proposed
proposed method,
method, representative
representative image
image data
data
werecollected
were collectedfor forinference
inference experiments
experiments in this
in this study.
study. TheThe image
image datadata
werewere
mainly mainly col-
collected
from
lectedscenes
from on the campus
scenes of WuhanofInstitute
on the campus Wuhan of Technology
Institute and various
of Technology andscenes within
various the
scenes
Wuhan
within thecityWuhan
area in city
HubeiareaProvince,
in HubeiChina. These
Province, scenes
China. contain
These a large
scenes number
contain of small
a large num-
objects, which
ber of small are challenging
objects, which arefor the inference
challenging for experiments.
the inference The results ofThe
experiments. the inference
results of
experiments are shown in Figure 14. Among them, Figure 14a–c
the inference experiments are shown in Figure 14. Among them, Figure 14a–c show the show the detection
results in the
detection campus
results scenario,
in the campusand Figureand
scenario, 14d–f show14d–f
Figure the detection
show theresults in the
detection Wuhan
results in
downtown
the Wuhanscenario.
downtown Thescenario.
results inThe
Figure 14 show
results that our
in Figure 14 method exhibits
show that good detection
our method exhibits
results in a variety
good detection of representative
results in a variety of scenes such as scenes
representative streets,such
campuses, andcampuses,
as streets, intersections.
and
Some of the detection results in Figure 14 show that our model has almost
intersections. Some of the detection results in Figure 14 show that our model has almost no missed
detection. These experimental results effectively prove the generalizability of our method.
Sensors 2023, 23, x FOR PEER REVIEW 25 of 28

Sensors 2023, 23, 7190 24 of of

no missed detection. These experimental results effectively prove the generalizability 27

our method.

(a) (b)

(c) (d)

(e) (f)
Figure
Figure 14.
14. Examples
Examples of
of detection
detection results
results on
on self-made data.
self-made data.

5. Conclusions
5. Conclusions
For the
For the object
object detection
detection task
task in
in UAV
UAV aerial
aerial photography
photography scenarios,
scenarios, there
there are
are problems
problems
such as
such as aa high
high proportion
proportionofofsmallsmallobjects,
objects,complex
complexbackgrounds,
backgrounds, and limited
and limitedhardware
hardware re-
sources. Most existing models suffer from poor detection accuracy
resources. Most existing models suffer from poor detection accuracy and struggle toand struggle to achieve
a balance
achieve between
a balance detection
between performance
detection and resource
performance consumption.
and resource consumption. To optimize the
To optimize
detection performance of the model while considering platform resource
the detection performance of the model while considering platform resource consump- consumption, this
paperthis
tion, proposes a UAV aerial
paper proposes a UAV scene object
aerial detection
scene model called
object detection UAV-YOLOv8,
model based on
called UAV-YOLOv8,
YOLOv8. Firstly, the WIoU v3 loss function is introduced, which incorporates
based on YOLOv8. Firstly, the WIoU v3 loss function is introduced, which incorporates a dynamic a
sample allocation strategy to effectively reduce the model’s attention to
dynamic sample allocation strategy to effectively reduce the model’s attention to extreme extreme samples
and improve
samples overall performance.
and improve Secondly,Secondly,
overall performance. the efficient
thedynamic
efficient sparse
dynamic attention
sparse mech-
atten-
anism BiFormer is integrated into the backbone network, enhancing
tion mechanism BiFormer is integrated into the backbone network, enhancing the model’s the model’s focus
on critical information in the feature maps and further optimizing
focus on critical information in the feature maps and further optimizing detectiondetection performance.
Finally, the fast and hardware-friendly FFNB is proposed, and two new detection scales are
designed based on this block to realize five-scale detection, which fully integrates shallow
and deep features and drastically reduces the missed detection rate of small objects. The
Sensors 2023, 23, 7190 25 of 27

improved model achieves an average detection accuracy improvement of 7.7% over the
baseline model without increasing its size or parameters, leading to significant enhance-
ments in object detection performance. Moreover, the improved model outperforms some
classical algorithms of similar types in terms of detection accuracy.
As the improved model adds two detection layers, the structure of the model becomes
complex and the feature map size of the shallow layer is relatively large, resulting in
different degrees of improvement in the computation and inference time of the model.
The FLOPs value of YOLOv8s is 28.7 billion times and that of UAV-YOLOv8s is 53 billion
times, which is nearly double the improvement, and there is still room for optimization
of computational resource consumption. The detection accuracy of the improved model
is still not high for very small objects such as bicycles, and the next major research focus
is to continue to optimize the detection accuracy of the model while keeping resource
consumption in mind.

Author Contributions: Conceptualization, G.W. and H.H.; methodology, G.W.; software, G.W.;
validation, G.W., H.H. and P.A.; formal analysis, G.W. and Y.C.; investigation, G.W.; resources, G.W.
and J.H.; data curation, G.W. and T.H.; writing—original draft preparation, G.W.; writing—review
and editing, G.W. and P.A.; visualization, G.W.; supervision, G.W. and Y.C.; project administration
G.W. and Y.C. All authors have read and agreed to the published version of the manuscript.
Funding: This study was funded by the Graduate Innovative Fund of Wuhan Institute of Technology
(No: CX2022148).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Li, Z.; Zhang, Y.; Wu, H.; Suzuki, S.; Namiki, A.; Wang, W. Design and Application of a UAV Autonomous Inspection System for
High-Voltage Power Transmission Lines. Remote Sens. 2023, 15, 865. [CrossRef]
2. Byun, S.; Shin, I.-K.; Moon, J.; Kang, J.; Choi, S.-I. Road Traffic Monitoring from UAV Images Using Deep Learning Networks.
Remote Sens. 2021, 13, 4027. [CrossRef]
3. Bouguettaya, A.; Zarzour, H.; Kechida, A.; Taberkit, A.M. A survey on deep learning-based identification of plant and crop
diseases from UAV-based aerial images. Cluster. Comput. 2022, 26, 1297–1317. [CrossRef]
4. Felzenszwalb, P.F.; Girshick, R.B.; McAllester, D.; Ramanan, D. Object Detection with Discriminatively Trained Part-Based Models.
IEEE Trans. Pattern Anal. Mach. Intell. 2010, 32, 1627–1645. [CrossRef] [PubMed]
5. Girshick, R.; Donahue, J.; Darrell, T.; Malik, J. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation.
In Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus, OH, USA, 23–28 June 2014;
pp. 580–587.
6. Girshick, R. Fast R-CNN. In Proceedings of the 2015 IEEE International Conference on Computer Vision (ICCV), Santiago, Chile,
7–13 December 2015; pp. 1440–1448.
7. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. IEEE
Trans. Pattern Anal. Mach. Intell. 2017, 39, 1137–1149. [CrossRef]
8. Redmon, J.; Divvala, S.; Girshick, R.; Farhadi, A. You Only Look Once: Unified, Real-Time Object Detection. In Proceedings of the
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 27–30 June 2016; pp. 779–788.
9. Redmon, J.; Farhadi, A. YOLO9000: Better, Faster, Stronger. In Proceedings of the 2017 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), Honolulu, HI, USA, 21–26 July 2017; pp. 6517–6525.
10. Redmon, J.; Farhadi, A. YOLOv3: An Incremental Improvement. arXiv 2018, arXiv:1804.02767.
11. Bochkovskiy, A.; Wang, C.Y.; Liao, H.Y.M. YOLOv4: Optimal Speed and Accuracy of Object Detection. arXiv 2020, arXiv:2004.10934.
12. Li, C.; Li, L.; Jiang, H.; Weng, K.; Geng, Y.; Li, L.; Ke, Z.; Li, Q.; Cheng, M.; Nie, W.; et al. YOLOv6: A Single-Stage Object Detection
Framework for Industrial Applications. arXiv 2022, arXiv:2209.02976.
13. Wang, C.Y.; Bochkovskiy, A.; Liao, H.Y.M. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object
detectors. arXiv 2022, arXiv:2207.02696.
14. Liu, W.; Anguelov, D.; Erhan, D.; Szegedy, C.; Reed, S.; Fu, C.Y.; Berg, A.C. SSD: Single Shot MultiBox Detector. arXiv 2016,
arXiv:1512.02325.
Sensors 2023, 23, 7190 26 of 27

15. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, Ł.; Polosukhin, I. Attention is all you need. In
Advances in Neural Information Processing Systems; MIT Press: Cambridge, MA, USA, 2017; pp. 5998–6008.
16. Carion, N.; Massa, F.; Synnaeve, G.; Usunier, N.; Kirillov, A.; Zagoruyko, S. End-to-end object detection with transformers. arXiv
2020, arXiv:2005.12872.
17. Luo, X.; Wu, Y.; Wang, F. Target Detection Method of UAV Aerial Imagery Based on Improved YOLOv5. Remote Sens. 2022, 14, 5063.
[CrossRef]
18. Zhou, H.; Ma, A.; Niu, Y.; Ma, Z. Small-Object Detection for UAV-Based Images Using a Distance Metric Method. Drones 2022, 6, 308.
[CrossRef]
19. Du, B.; Huang, Y.; Chen, J.; Huang, D. Adaptive Sparse Convolutional Networks with Global Context Enhancement for Faster
Object Detection on Drone Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,
Vancouver, BC, Canada, 18–22 June 2023; pp. 13435–13444.
20. Deng, F.; Xie, Z.; Mao, W. Research on edge intelligent recognition method oriented to transmission line insulator fault detection.
Int. J. Electr. Power Energy Syst. 2022, 139, 108054. [CrossRef]
21. Howard, A.; Pang, R.; Adam, H.; Le, Q.; Sandler, M.; Chen, B.; Wang, W.; Chen, L.C.; Tan, M.; Chu, G.; et al. Searching for
MobileNetV3. In Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, Republic of Korea, 27
October–2 November 2019; pp. 1314–1324.
22. Zheng, J.; Fu, H.; Li, W.; Wu, W.; Yu, L.; Yuan, S.; Tao, W.Y.W.; Pang, T.K.; Kanniah, K.D. Growing status observation for oil palm
trees using Unmanned Aerial Vehicle (UAV) images. ISPRS J. Photogramm. Remote Sens. 2021, 173, 95–121. [CrossRef]
23. Liu, M.; Wang, X.; Zhou, A.; Fu, X.; Ma, Y.; Piao, C. UAV-YOLO: Small Object Detection on Unmanned Aerial Vehicle Perspective.
Sensors 2020, 20, 2238. [CrossRef]
24. He, K.M.; Zhang, X.Y.; Ren, S.Q.; Sun, J. Deep Residual Learning for Image Recognition. In Proceedings of the 2016 IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 27–30 June 2016; pp. 770–778.
25. Liu, B.; Luo, H. An Improved Yolov5 for Multi-Rotor UAV Detection. Electronics 2022, 11, 2330. [CrossRef]
26. Wang, J.; Zhang, F.; Zhang, Y.; Liu, Y.; Cheng, T. Lightweight Object Detection Algorithm for UAV Aerial Imagery. Sensors 2023,
23, 5786. [CrossRef]
27. Woo, S.; Park, J.; Lee, J.-Y.; Kweon, I.S. CBAM: Convolutional Block Attention Module. In Proceedings of the Computer
Vision—ECCV 2018, Cham, Switzerland, 8–14 September 2018; pp. 3–19.
28. Liu, Y.; Yang, F.; Hu, P. Small-object detection in UAV-captured images via multi-branch parallel feature pyramid networks. IEEE
Access 2020, 8, 145740–145750. [CrossRef]
29. Chen, J.; Kao, S.-H.; He, H.; Zhuo, W.; Wen, S.; Lee, C.-H.; Chan, S.-H.G. Run, Don’t walk: Chasing higher FLOPS for faster neural
networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Vancouver, BC, Canada,
18–22 June 2023.
30. Zhu, L.; Wang, X.; Ke, Z.; Zhang, W.; Lau, R. BiFormer: Vision Transformer with Bi-Level Routing Attention. arXiv 2023,
arXiv:2303.08810.
31. Tong, Z.; Chen, Y.; Xu, Z.; Yu, R. Wise-IoU: Bounding Box Regression Loss with Dynamic Focusing Mechanism. arXiv 2023,
arXiv:2301.10051.
32. He, K.; Zhang, X.; Ren, S.; Sun, J. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition. IEEE Trans.
Pattern Anal. Mach. Intell 2015, 37, 1904–1916. [CrossRef] [PubMed]
33. Liu, S.; Qi, L.; Qin, H.; Shi, J.; Jia, J. Path Aggregation Network for Instance Segmentation. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 18–23 June 2018; pp. 8759–8768.
34. Li, X.; Wang, W.; Wu, L.; Chen, S.; Hu, X.; Li, J.; Tang, J.; Yang, J. Generalized Focal Loss: Learning Qualified and Distributed
Bounding Boxes for Dense Object Detection. arXiv 2020, arXiv:2006.04388.
35. Zheng, Z.; Wang, P.; Liu, W.; Li, J.; Ye, R.; Ren, D. Distance-IoU loss: Faster and better learning for bounding box regression. In
Proceedings of the AAAI Conference on Artificial Intelligence, New York, NY, USA, 7–12 February 2020; pp. 12993–13000.
36. Feng, C.; Zhong, Y.; Gao, Y.; Scott, M.R.; Huang, W. TOOD: Task-Aligned One-Stage Object Detection. In Proceedings of the 2021
IEEE International Conference on Computer Vision (ICCV), Montreal, QC, Canada, 10–17 October 2021; pp. 3490–3499.
37. Zhang, Y.F.; Ren, W.; Zhang, Z.; Jia, Z.; Wang, L.; Tan, T. Focal and efficient IOU loss for accurate bounding box regression.
Neurocomputing 2022, 506, 146–157. [CrossRef]
38. Gevorgyan, Z. SIoU Loss: More Powerful Learning for Bounding Box Regression. arXiv 2022, arXiv:2205.12740.
39. Cao, X.; Zhang, Y.; Lang, S.; Gong, Y. Swin-Transformer-Based YOLOv5 for Small-Object Detection in Remote Sensing Images.
Sensors 2023, 23, 3634. [CrossRef]
40. Lu, S.; Lu, H.; Dong, J.; Wu, S. Object Detection for UAV Aerial Scenarios Based on Vectorized IOU. Sensors 2023, 23, 3061.
[CrossRef]
41. Zhang, T.; Zhang, Y.; Xin, M.; Liao, J.; Xie, Q. A Light-Weight Network for Small Insulator and Defect Detection Using UAV
Imaging Based on Improved YOLOv5. Sensors 2023, 23, 5249. [CrossRef]
42. Jiang, X.; Cui, Q.; Wang, C.; Wang, F.; Zhao, Y.; Hou, Y.; Zhuang, R.; Mei, Y.; Shi, G. A Model for Infrastructure Detection along
Highways Based on Remote Sensing Images from UAVs. Sensors 2023, 23, 3847. [CrossRef]
43. Howard, A.G.; Zhu, M.; Chen, B.; Kalenichenko, D.; Wang, W.; Weyand, T.; Andreetto, M.; Adam, H. MobileNets: Efficient
Convolutional Neural Networks for Mobile Vision Applications. arXiv 2017, arXiv:1704.04861.
Sensors 2023, 23, 7190 27 of 27

44. Zhang, X.; Zhou, X.; Lin, M.; Sun, J. ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices. arXiv
2017, arXiv:1707.01083.
45. Han, K.; Wang, Y.H.; Tian, Q.; Guo, J.Y.; Xu, C.J.; Xu, C. GhostNet: More Features from Cheap Operations. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 14–19 June 2020; pp. 1580–1589.
46. Zhu, P.; Wen, L.; Du, D.; Bian, X.; Fan, H.; Hu, Q.; Ling, H. Detection and Tracking Meet Drones Challenge. IEEE Trans. Pattern
Anal. Mach. Intell. 2021, 44, 7380–7399. [CrossRef] [PubMed]
47. Rezatofighi, H.; Tsoi, N.; Gwak, J.; Sadeghian, A.; Reid, I.; Savarese, S. Generalized intersection over union: A metric and a loss
for bounding box regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long
Beach, CA, USA, 15–20 June 2019; pp. 658–666.
48. Wang, C.Y.; Liao, H.Y.M.; Wu, Y.H.; Chen, P.Y.; Hsieh, J.W.; Yeh, I.H. CSPNet: A new backbone that can enhance learning
capability of CNN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle,
WA, USA, 13–19 June 2020; pp. 390–391.
49. Cai, Z.; Vasconcelos, N. Cascade R-CNN: Delving into High Quality Object Detection. In Proceedings of the 2018 IEEE/CVF
Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 18–22 June 2018; pp. 6154–6162.
50. Lin, T.; Goyal, P.; Girshick, R.; He, K.; Dollár, P. Focal Loss for Dense Object Detection. In Proceedings of the 2017 IEEE
International Conference on Computer Vision (ICCV), Venice, Italy, 22–29 October 2017; pp. 2999–3007.
51. Duan, K.; Bai, S.; Xie, L.; Qi, H.; Huang, Q.; Tian, Q. Centernet: Keypoint triplets for object detection. In Proceedings of the
IEEE/CVF International Conference on Computer Vision, Seoul, Republic of Korea, 27 October–2 November 2019; pp. 6569–6578.
52. Zhu, C.; He, Y.; Savvides, M. Feature Selective Anchor-Free Module for Single-Shot Object Detection. In Proceedings of the 2019
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, USA, 16–20 June 2019; pp. 840–849.
53. Zhang, S.; Chi, C.; Yao, Y.; Lei, Z.; Li, S.Z. Bridging the Gap Between Anchor-Based and Anchor-Free Detection via Adaptive
Training Sample Selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle,
WA, USA, 14–19 June 2020; pp. 9759–9768.
54. Selvaraju, R.R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; Batra, D. Grad-CAM: Visual Explanations from Deep Networks
via Gradient-Based Localization. In Proceedings of the 2017 IEEE International Conference on Computer Vision (ICCV), Venice,
Italy, 22–29 October 2017; pp. 618–626.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.