Journal of Physics: Conference Series

PAPER • OPEN ACCESS You may also like

- A smoking behavior detection method
Research on Flower Image Classification Method based on the YOLOv5 network
Xiangkui Jiang, Haochang Hu, Xun Liu et
Based on YOLOv5 al.

- YOLO-A2G: An air-to-ground high-

precision object detection algorithm based
To cite this article: Ming Tian and Zhihao Liao 2021 J. Phys.: Conf. Ser. 2024 012022 on YOLOv5
Lei Li, Ziyu Yao, Zhiyong Miao et al.

- Faster Detection Method of Driver

Smoking Based on Decomposed YOLOv5
Fangfei Shi, Hui Zhou, Chunyang Ye et al.
View the article online for updates and enhancements.

This content was downloaded from IP address 199.244.60.117 on 29/10/2022 at 05:18

2nd International Conference on Computer Vision and Data Mining (ICCVDM 2021) IOP Publishing
Journal of Physics: Conference Series 2024 (2021) 012022 doi:10.1088/1742-6596/2024/1/012022

Research on Flower Image Classification Method Based on

YOLOv5

Ming Tian1*, Zhihao Liao2

1
Department of Automation, Faculty of Artificial Intelligence and Automation, Hust,
Wuhan, China
2
Department of Artificial Intelligence, Faculty of Information and Engineering,
Nanchang University, Nanchang, China
*[email protected]

Abstract—The rapid development of deep learning has accelerated the progress of related
technologies in the computer vision field and it has broad application prospects. Due to flower
inter-class similarity and intra-class differences, flower image classification has essential
research value. To achieve flower image classification, this paper proposes a deep learning
method using the current powerful object detection algorithm YOLOv5 to achieve fine-grained
image classification of flowers. Overlap and occluded objects often appear in the images of the
flowers, so the DIoU_NMS algorithm is used to select the target box to enhance the detection of
the blocked objects. The experimental dataset comes from the Kaggle platform, and experimental
results show that the proposed model in this paper can effectively identify five types of flowers
contained in the dataset, Precision reaching 0.942, Recall reaching 0.933, and mAP reaching
0.959. Compared with YOLOv3 and Faster-RCNN, this model has high recognition accuracy,
real-time performance, and good robustness. The mAP of this model is 0.051 higher than the
mAP of YOLOv3 and 0.102 higher than the mAP of Raster-RCNN.

1. INTRODUCTION
The study of flowers has many aspects, such as the study of ornamental value, nutritional, and medicinal
value of flowers[1]. Accurate classification of flowers is the primary work of related research.
Researchers can realise flower classification accurately under knowledge accumulation and data access,
but it takes much time. With the development of society and the improvement of living standards, people
have an increasing demand for flower appreciation, but ordinary people are often challenging to
accurately identify the types of flowers. Therefore, the automatic classification of flower images can
assist researchers in flower research and provide popular science and convenience for people.
Flower image classification belongs to fine-grained image classification[2], which classifies different
flower subcategories. There are three difficulties in the image classification of flowers. (1) There are
similarities between different categories of flowers, as shown in Figure 1; (2) Flowers of the same
category have different characteristics, as shown in Figure 2; (3) Plants usually have more than one flower,
so there are overlapping flowers in the image, as shown in Figure 3.

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
2nd International Conference on Computer Vision and Data Mining (ICCVDM 2021) IOP Publishing
Journal of Physics: Conference Series 2024 (2021) 012022 doi:10.1088/1742-6596/2024/1/012022

(a) tulip (b)rose

Figure 1. Inter-class similarity between different
flower classes ( (a) tulip; (b) rose)

(a)daisy (b)sunflower
Figure 2 .Intra-class difference between flowers of
the same class( (a)daisy; (b)sunflower)

(a) tulip (b)daisy

Figure 3 . Flower images are blocked and overlapping
( (a)tulip; (b)daisy)
Traditional methods extract image features mainly by feature extraction operators and combine the
extracted features with traditional machine learning algorithms to realise image classification. Many
researchers have used complex algorithms to divide flower images and perform feature extraction in
recent years, achieving good results [3-6]. Nilsback et al. [3] proposed to fuse the SIFT and HOG features
and classify them using SVM. Wang L et al. [4] proposed transforming original colour flower images
from RGB space to Lab space and performing segmentation using the OTSU algorithm. Angelova et al.
[5] proposed a target segmentation method and extracted 4-scale HOG features using locality-constrained
linear cording (LLC) for feature coding. Xie X et al. [6] proposed the significance-based flower image
segmentation method Saliency Based flower Image Segmentation. However, the algorithm for extracting
the image characteristics in the traditional methods is closely related to the current image's characteristics,
and the algorithm's versatility is not good.
Deep learning can automatically extract features from images, avoiding the subjectivity of manually
extracting features, and is widely used in object detection and classification problems. Xia et al. [7] use
Inception v3 network and migration learning techniques to improve classification accuracy, but the model
is complex and flawed in real-time. Yin H et al.[8] proposed an unsupervised flower image classification
method based on selective convolution descriptor aggregation. For fine-grained image classification it
refers to the literature for bird classification [9]. Literature [9] Use bounding box and part annotation to
train the R-CNN model to obtain the optimal candidate regions and classify them with SVM.
The YOLO algorithm is a one-stage target detection algorithm in deep learning. The YOLO algorithm
was proposed in 2016 by Redmon et al. [8], which implements the classification and localisation of object
detection with one network, and YOLO has been widely used in object detection. Currently, YOLO has

2
2nd International Conference on Computer Vision and Data Mining (ICCVDM 2021) IOP Publishing
Journal of Physics: Conference Series 2024 (2021) 012022 doi:10.1088/1742-6596/2024/1/012022

experienced development from v1 to v5 [10-13]. YOLOv5, launched in 2020, has the advantages of small
volume, fast speed, high precision, and implementation in ecologically mature PyTorch. R-CNN belongs
to the Two-stage object detection algorithm, which can effectively improve the problem of the target to
be tested in the image, but the Two-stage model is more complex and computational than the One-stage
model. From all of the above, the latest YOLOv5 target detection algorithm is selected to classify the
flower dataset of the Kaggle platform.

2. YOLOV5 MODEL
YOLOv5 is the latest object detection algorithm in the YOLO algorithm, which has high detection
accuracy and is fast and good in real-time. YOLOv5 includes four models of YOLOv5s, YOLOv5m,
YOLOv5l, YOLOv5x, where the YOLOv5s has the smallest volume. This paper selects the YOLOv5s
model, consisting of four parts, Input, Backbone, Neck, Prediction, respectively. The network structure
is shown in Figure 4.

Figure 4. YOLOv5 network

3
2nd International Conference on Computer Vision and Data Mining (ICCVDM 2021) IOP Publishing
Journal of Physics: Conference Series 2024 (2021) 012022 doi:10.1088/1742-6596/2024/1/012022

2.1 Input
On the input side, YOLOv5 draws on the CutMix method and uses Mosaic data enhancement to improve
the recognition of small targets effectively. Add adaptive scaling processing, the image unified scaling to
a unified size and then sent to the network learning, to enhance the ability of network data processing.

2.2 Backbone
Backbone includes CSP networks and Focus structures, etc. The Focus structure contains four-slice
operations and one convolution of 32 convolutional cores, transforming the original 608 * 608 * 3 picture
into 304 * 304 * 32 feature maps. CSPNet performs local cross-layer fusion, utilising the feature
information of different layers to obtain richer feature maps.

2.3 Neck
The Neck section contains both PANet and SPP. PANet (PathAggregation), to fully integrate the image
features of different layers, aggregate the top feature information and the output features of different CSP
networks in top-down order, and then aggregate the shallow features from the bottom-up. SPP (space
pyramid pooling) uses four different sized nuclei for maximum pooling and then performs tensor splicing.

2.4 Prediction

2.4.1 loss function: The loss function of YOLOv5s uses GIOU_Loss, which alleviates the situation that
IOU_Loss cannot handle the two boxes. As shown in Figure 5, we assume that the minimum external
rectangle of the prediction box and the real box is C, The union set of the forecast box and the real box is
N, The intersection of the prediction box and the real box is M, IOU is the ratio of intersection to the
union, as shown in (1). D is the difference set of C and N, as shown in (2). GIOU is IOU minus the ratio
of D TO C, as shown in (3), then the formula for GIOU_Loss is shown in (4).

Figure 5. Geometric relationship

M
IOU = (1)
N

D = |C − N| (2)
D
GIOU = IOU − (3)
C
M D
GIOU_LOSS = 1 − GIOU = 1 − ( − ) (4)
N C

4
2nd International Conference on Computer Vision and Data Mining (ICCVDM 2021) IOP Publishing
Journal of Physics: Conference Series 2024 (2021) 012022 doi:10.1088/1742-6596/2024/1/012022

2.4.2 Non-Maximum Suppression: locally removes redundant detection boxes, retaining the best one.
YOLOv5 uses NMS to select the detection box, and this article uses DIOU_NMS. DIOU_NMS can
improve the detection accuracy of the overlapping and occluded targets.

3. RESULTS AND ANALYSIS OF THE EXPERIMENTS

3.1 Flower detection framework

The framework of this model is shown in Figure 6, divided into two parts: the training phase and flower
classification phase. During the training phase, the training set is entered into the model for training.
During the flower classification phase, the model completes the flower detection and classification of the
input image. The model processes the prediction scores in the five categories of daisies, sunflowers,
dandelion, roses, and tulips, and then plotted the predicted flower species and scores on the output image.

Figure 6. Flower detection framework

3.2 Dataset
The dataset of this paper adopts the flower dataset disclosed on the Kaggle platform, which includes
flower images of daisy, dandelion, rose, sunflower and tulip. See Fig. 7 for the schematic diagram of
various flower images. The flowers were labelled with LabelMe, and the data set was divided into the
training set, verification set and test set according to the ratio of 8:1:1. The dataset consists of 400 images,
each of different sizes.

Figure 7. Five types of flower images

3.3 Results of the training

The Pytorch framework is adopted in this experiment, and the model is trained on GeForce GTX 1080ti
(video memory 12G) GPU, CUDA 10.0 experimental environment. The training parameters are set as
follows: the input image pixel is 640*640, the Batch size is 16, and the initial learning rate is 0.01. The

5
2nd International Conference on Computer Vision and Data Mining (ICCVDM 2021) IOP Publishing
Journal of Physics: Conference Series 2024 (2021) 012022 doi:10.1088/1742-6596/2024/1/012022

momentum factor is set to 0.95, the weight decays to 0.001, and the epoch is 300. Precision, Recall and
mAP after model training are shown in Figure 8.

Figure 8. Training results

It can be seen from Figure 8 that the accuracy, recall, and mAP of the model after training is high and
stable, which shows that the training effect of the model is good.

3.4 The evaluation index

This experiment uses Precision (P), Recall (R), and mean Average Precision (mAP) as evaluation
indicators. The results of the detection contained four cases, namely True Positive (TP), False Positive
(FP), True Negative (TN), and False Negative (FN). The definition is shown in Table 1.
Table 1 Definition of different samples
Abbreviation Define
True Positive TP A positive sample with a positive forecast
False Positive FP A negative sample with a positive forecast
True Negative TN A negative sample with a negative forecast
False Negative FN A positive sample with a negative forecast

For category C, the accuracy is the ratio of the number of correct samples detected to the number of
samples detected, and the calculation formula is shown in (5). The Recall of category C is the ratio of the
correct number of samples detected to the total number of samples of this class, and the calculation
formula is shown in (6). AP is the area of the curve enclosed by P and R, and mAP is the average of the
AP for all categories, and the calculation formula is shown in (7).
TP
Precision =
TP+FP
(5)
TP
Recall = (6)
TP+FN
ΣAveragePrecision
Mean Average Precision = (7)
C

6
2nd International Conference on Computer Vision and Data Mining (ICCVDM 2021) IOP Publishing
Journal of Physics: Conference Series 2024 (2021) 012022 doi:10.1088/1742-6596/2024/1/012022

3.5 Test results and analysis

The experimental results are shown in Figure 9. The proposed model can correctly detect and classify
flower images with insect interference, overlapping ones, and fuzzy flower images from the detection
results.

(a) General flower image (b)Insect interference image

(c) Overlapping flower image (d)Fuzzy flower image

Figure 9. Test results
Furthermore, we detect and classify the test sets on YOLOv3 and Faster-RCNN models, and the
indicators of the different models are shown in Figure 10.

Figure 10. Precision, Recall and mAP in different models

From Figure 10, the algorithm in this paper, which is YOLOv5s, has higher Precision，Recall and
mean Average Precision than other algorithms and performs better than other algorithms, which is
suitable and excellent as an algorithm for implementing the automatic classification of flower images

4. CONCLUSION
Based on the powerful YOLOv5 algorithm in target detection, this paper realised flower image detection
and fine-grained classification. The experimental results show that the model can detect and recognise
the flower images with insect interference, overlapping flower images and fuzzy flower images.
Compared with the YOLOv3 and Faster-RCNN algorithms, the flower classification model proposed in
this paper has good detection and classification results and obvious performance advantages. In addition,
this experiment mainly focuses on the detection and classification of large flower targets in the image.
The next step is to optimise the model further to classify and detect small flower targets in the image.

REFERENCES
[1] Ou J,YANG C H. Quantitative evaluation of the ornamental value of wild herbal flowers [J].

7
2nd International Conference on Computer Vision and Data Mining (ICCVDM 2021) IOP Publishing
Journal of Physics: Conference Series 2024 (2021) 012022 doi:10.1088/1742-6596/2024/1/012022

Guizhou Agricultural Science,2009,37(06):166-170.

[2] Weng Y C，Tian Y，Lu D M，et al． Fine-grained bird classifica-tion based on deep region
networks［J］． Journal of Image and Graphics，2017，22 ( 11 ) : 1521-1531．
[3] Nilsback M E，Zisserman A． Automated flower classification over a large number of classes
［C］/ / Proceedings of the 6th Indian Conference on Computer Vision，Graphics ＆ Image
Processing ． Bhubaneswar ， India: IEEE ， 2008: 722-729 ． ［ DOI: 10 ． 1109 /
ICVGIP． 2008． 47］
[4] Wang L, Fang L, Chen E,et al.Flower image segmentation algorithm based on Lab color space[J].
Journal of Zhejiang Wanli University, 20181 12(3);67-73
[5] Angelova A，Zhu S H． Efficient object detection and segmentation for fine-grained recognition
［ C ］ / / Proceedings of 2013 IEEE Conference on Computer Vision and Pattern Ｒ
ecognition． Portland，OＲ，USA: IEEE，2013: 811-818． ［DOI: 10． 1109 / CVP
Ｒ． 2013． 110］
[6] Xie X D，Lyu Y P，Cao D L． Saliency detection based flower image foreground segmentation
［EB / OL］． 2014-09-18［2018-06-20］
[7] Xia X L，Xu C，Nan B． Inception-v3 for flower classification［C］/ /Proceedings of the 2nd
International Conference on Image，Vision and Computing． Chengdu: IEEE， 2017: 783-
787［DOI: 10． 1109 /ICIVC． 2017． 7984661］
[8] Yin H，Fu X，Zeng J X，Duan B，Chen Y． Flower image classification with selective
convolutional descriptor aggregation［J］． Journal of Image and Graphics，2019，24( 05) :
0762-0772．
[9] Zhang N，Donahue J，Girshick Ｒ，et al． Part-Based Ｒ-CNNs for fine-grained category
detection ［ C ］ / / Proceedings of the 13th European Conference on Computer
Vision． Zurich，Switzerland: Springer，2014: 834-849． ［DOI: 10． 1007 /978-3-319-
10590-154］
[10] REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: Unified, Real-time Object
Detection [C]. Las Vegas: IEEE Conference on Computer Vision 56 and Pattern
Recognition, 2016.
[11] REDMON J, FARHADI A. YOLO9000: Better, Faster, Stronger[C]. Honolulu, USA,: IEEE
Conference on Computer Vision and Pattern Recognition, 2017.
[12] REDMON J, FARHADI A. YOLOv3: An incremental improvement[EB/OL]. (2018-04-
08)[2021-01-15].
[13] BOCHKOVSKIY A, WANG C Y, LIAO H M. YOLOv4: Optimal speed and accuracy of object
detection[EB/OL]. (2020-04-23)[2021-01-15].
.