Hindawi

Computational and Mathematical Methods in Medicine

Volume 2022, Article ID 6554371, 10 pages
https://doi.org/10.1155/2022/6554371

Research Article
Application of Graphic Design with Computer Graphics and
Image Processing: Taking Packaging Design of Agricultural
Products as an Example

Tianhe Xie, Rongyi Sun, Jiahao Zhang, Ruiqi Wang, and Jiashu Wang
College of Landscape Architecture and Art, Northwest A&F University, Xianyang 712100, China

Correspondence should be addressed to Jiashu Wang; [email protected]

Received 10 April 2022; Revised 3 May 2022; Accepted 7 May 2022; Published 2 June 2022

Academic Editor: Naeem Jan

Copyright © 2022 Tianhe Xie et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

With development of economy, all industries have undergone earthshaking changes. Various new technologies are starting to be
employed in all aspects of life, and graphic design is no exception. The use of computer graphics and image processing
technologies in graphic design can substantially improve design efficiency and make graphic design job more convenient to
develop. The requirements for the quality of graphic design are higher. Quality inspection has become a necessary step in the
production process, in which the detection of graphic design defects is an indispensable and important link. The traditional
graphic design defect detection adopts the method of manual visual inspection, which has the disadvantages of poor stability,
long consumption time, and high labor cost. As an efficient computer graphics and image processing technology, convolutional
neural network has received extensive attention in graphic design defect detection because of its advantages of high speed,
efficiency, and high degree of automation. Taking agricultural product packaging as an example, this paper studies application
technology for graphic design defect detection with convolutional neural network (CNN). The main contents are as follows:
construct the original YOLOv3 network model, input the graphic design images of agricultural product packaging into the
network model in batches according to the computing power of the hardware equipment, train the YOLOv3 network, and
deeply study and analyze the experimental results. The related improvement techniques are then given, based on the
characteristics of agricultural product packaging design faults. The backbone network, multiscale feature map, a priori frame,
and activation function of YOLOv3 are improved, and then performance of the improved model is verified by experiments.

1. Introduction the content of graphic design. In the process of using com-

puter graphics and image processing technology, it is neces-
At this point, the development and applicability of computer sary to fully exploit the role and value of computers.
graphics and image processing have made graphic design However, we must also realize that there are still many
work much easier. Computer graphics and image processing unreasonable problems, which cannot give full play to the
technology have rapidly become an essential competence for role of computer graphics and image processing technology,
designers in the work process in the graphic design business. which affects the sustainable and healthy development of
Various picture information may be modified and optimized graphic design. Among them, graphic design defects have
in terms of style, color, and theme using this technology, become a long-standing problem, and most graphic design
which can not only enhance the design effect but also products have certain defects. How to detect the defects of
improve job efficiency, effectively encouraging the design graphic design products has become an important and
industry’s development. The application of computer practical topic [1–5].
graphics and image processing technology in graphic design Manual quality inspection is still widely used, especially
can solve the problems of diversity and aesthetics in graphic in small and medium-sized businesses that use human
design to a certain extent and can also add more richness to detection methods, and it accounts for a significant portion
 7396, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2022/6554371 by CAPES, Wiley Online Library on [27/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2 Computational and Mathematical Methods in Medicine

of the market. Manual quality inspection, on the other hand, model can improve the detection accuracy of agricultural
must remain in a fixed position, which can lead to a variety product graphic design defects.
of inspection issues when determining the presence of flaws The paper’s organization paragraph is as follows: the
by eye observation. Manual quality inspection cannot even related work is presented in Section 2. Section 3 analyzes
ensure the stability and consistency of the detection of the the methods of the proposed work. Section 4 discusses the
same batch of graphic design products due to the influence experiments and results. Finally, in Section 5, the research
of factors such as individual variances. In addition, missed work is concluded.
detection and false detection are prone to occur during the
detection process, the accuracy of the detection results is 2. Related Work
reduced, and the product quality is uneven. In addition,
the upper limit of manual inspection efficiency is low, the Reference [16] proposed a real-time algorithm for weaving
cost is high, and a sampling inspection strategy is usually defect detection based on machine vision. The technique
adopted to ensure production efficiency. The quality assess- detects five types of fabric surface defects using image pro-
ment of the entire product group by randomly selecting sev- cessing methods such as wavelet transform, double-
eral products from the same group of products is far less threshold binarization, and morphological operations, with
rigorous than large-scale testing. Therefore, in the manual a 93.4 percent defect detection and 96.3 percent defect type
quality inspection process, quality control and production identification accuracy. Reference [17] proposed a camshaft
efficiency cannot always be achieved at the same time. In surface defect detection method based on neighborhood
addition, defect detection requires not only qualitative weighted segmentation. The method realizes the detection
inspection of the appearance of graphic design products of typical defects such as trauma, trachoma, and poor grind-
but also statistics of defect size and other data. However, ing on the surface of the camshaft through defect segmenta-
manual quality inspection can only rely on paper and pen tion and defect area marking. At a speed of 0.44 s per shaft,
to record, and the obtained data is not comprehensive and the method was able to detect defects larger than 1 mm in
cannot be used as valuable information to guide production diameter on the camshaft surface. Reference [18] proposed
and production optimization and improvement [6–10]. a method for detection of bottle mouth defects based on
Machine vision detection technology is a modern detec- threshold segmentation, which combines three-circle posi-
tion technology that integrates photoelectric sensing, com- tioning method, residual analysis dynamic threshold, and
puter science, image processing, pattern recognition, and global threshold segmentation to detect five kinds of bottle
other disciplines. It converts the captured target into an mouth defects. Reference [19] adopts NCC for defect detec-
image signal, transmits it to a dedicated image processing tion and adopts the integral graph strategy to speed up the
system, and extracts the parameters to be detected, thereby traditional NCC calculation, so that the calculation time
realizing the defect detection of graphic design products. does not change with the change of the template window
Machine vision inspection technology offers many advan- size. Reference [20] proposed an engine block defect detec-
tages over hand visual inspection: noncontact measurement tion based on small area template matching. The method
of the product to be inspected causes no damage to the prod- combines rough search and fine search strategy improve-
uct. The wider spectral response range and higher resolution ment algorithm to determine the exact position and similar-
extend the visual range and resolution of the naked eye. Sta- ity of the images to be matched, realizes defect detection, and
ble, reliable, and fast work for a long time makes up for the has a faster operation speed and meets real-time require-
shortcomings of poor visual stability and low work effi- ments. Reference [21] proposed a method for detecting
ciency. The mechanized nature has greatly increased the appearance defects of smart meters based on machine vision.
degree of industrial automation. In recent years, due to the The method uses image processing technologies such as
rapid development of machine vision and digital image pro- median filtering, binarization, edge detection, template
cessing, defect detection technology based on machine matching, and OCR to detect the defects of smart meter
vision has received more and more attention in industrial LCD screen, signage characters, barcodes, and LED indica-
quality inspection and has been widely used in various tors, and the detection takes 3 s. Reference [22] proposed a
industries. Based on the above background, it is of great sig- keyboard defect detection method based on template match-
nificance to study the defect detection technology of graphic ing. The method utilizes local threshold segmentation and
design based on machine vision [11–15]. coordinate projection to locate keys and uses template
This work offers a computer graphics and image pro- matching to match key characters to realize keyboard defect
cessing technology based on convolutional neural networks detection. Reference [23] proposes a steel surface defect
to detect faults in agricultural product packaging design, detection algorithm, which uses Hough transform, PCA,
using the packaging design of agricultural products as an and self-organization map to detect three kinds of steel sur-
example. The characteristics of surface defects in agricultural face defects, and achieves an 87% accuracy on the Arcelor
product packaging design are studied in depth in this Mittal image set. Reference [24] proposed a TFT-LCD defect
research, and then enhancements to the network model, fea- detection algorithm based on machine vision. The algorithm
ture maps, prior boxes, and activation functions are pro- uses Gabor filtering, adaptive binarization, connected
posed. Finally, the revised network model is experimentally domain extraction, and blob algorithm to identify TFT-
tested and compared to the original network model’s exper- LCD point defects and line defects and achieves a high rec-
imental results. The results show that the improved network ognition rate. Reference [25] proposed a fabric defect
 7396, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2022/6554371 by CAPES, Wiley Online Library on [27/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Computational and Mathematical Methods in Medicine 3

detection method based on SVM. The method extracts the pooling layers, and fully connected layers, and finally, the
geometric features of the fabric surface pattern and uses output layer will input the result.
the SVM classifier to achieve defect detection, which can The convolutional layer is an important part of the con-
detect 5 kinds of fabric defects with an accuracy of 94.84%. volutional neural network, and the layers are directly con-
Reference [26] proposed a defect detection and identification nected locally to map the underlying information into
algorithm based on the PCA algorithm, which was applied high-level features. Each convolutional layer has multiple
to the surface of unshaded cover glass. Algorithms combine convolution kernels, and the size and number of convolution
surface defect detection and identification processes into kernels have an important impact on the recognition ability
one. The missed detection rate and false alarm rate in the of the convolutional neural network. The essence of a convo-
defect detection process reached 12% and 6%, and the recog- lution kernel is a set of trainable weight matrices. When con-
nition rate in the defect identification process reached more volution operation is performed on the input image, the
than 90%. Reference [27] proposed a defect detection algo- convolution kernel covers part of the input image according
rithm based on CNN variant network Overfeat. to a certain rule, and the value of the convolution kernel is
The algorithm uses Overfeat CNN combined with ASR multiplied by the value of the corresponding position pixel
for surface defect detection and achieves 98.7% and 60.3% in the image. Adding the products obtained in the previous
accuracy on NEU and MO databases, respectively. Literature step is the result of this convolution operation. Convolu-
[28] developed a welding seam detection method based on tional layers are often used to extract image features, and
multilayer perceptron and achieved good detection results the result after a series of convolution operations is often
on its data set. Literature [29] studied a variety of surface referred to as a feature map. The mathematical expression
defect detection techniques based on texture feature extrac- of the convolution operation is as follows:
tion and summarized the progress of anomaly detection
 
methods in the field of texture detection. Reference [30] used ylj = σ 〠xl−1 ∗ w l
bl
i ij + j : ð1Þ
fractal dimension as feature quantity in hot-rolled strip sur-
face defect identification and estimated an optimal scale to
detect defects through fractal dimension curve graph to Weight sharing is an important feature of the convolu-
obtain better detection effect. Reference [31] proposes a tion layer. Weight sharing means that when performing con-
method for detecting foreign matter in liquid medicine volution operations, all positions of the image use
based on the idea of difference between frames, which convolution kernels with the same parameters. Weight shar-
achieves high detection accuracy and speed. Convolutional ing drastically decreases the network’s parameters, allowing
neural networks based on deep learning have demonstrated for much faster model training and inference. Another ele-
their powerful capabilities in feature extraction and pattern ment of the convolution layer is multiconvolution operation,
recognition; so, more and more defect detection methods which refers to the employment of multiple separate convo-
incorporating deep learning have emerged. Reference [32] lution kernels to scan the input image. The scan result of
developed a visual inspection device for inspecting cigar each convolution kernel is a feature map, and finally, the
pack labels at a speed of 500 packs per minute. Reference multifeature map is used for prediction, and the accuracy
[33] used texture analysis to detect the seal quality of food can be greatly improved.
packaging bags. The pooling layer is generally located after the convolu-
tional layer, and its main function is to compress the size of
3. Method the feature map to reduce the amount of network parame-
ters. The output of the pooling layer is the same as the con-
This work uses convolutional neural network as a computer volutional layer, which is a feature map. These feature maps
graphics image processing technology to detect the defects in uniquely correspond to a feature map of the previous layer,
the graphic design of agricultural product packaging. The and the size will be reduced to a certain extent. In theory,
convolutional network used is YOLOv3, and a series of it is possible to directly input the feature map output by
improvement measures have been taken to effectively the convolutional layer into the classifier for result predic-
improve its performance. tion. However, after the convolution operation of the image,
the size of the image is not significantly reduced. Instead,
3.1. Convolutional Neural Network. CNN is an artificial neu- due to the convolution operation of multiple convolution
ral network designed to imitate the working mode of neu- kernels, the feature map of the network increases. This
rons in the human brain. The basic unit is artificial increases the number of network parameters and the
neurons. Artificial neurons also transmit signals hierarchi- amount of computation, which seriously affects the training
cally and perform extremely well when dealing with gridded speed of the network. Therefore, it is necessary to perform a
data, especially for large image processing. Convolutional dimensionality reduction operation on the convolutional
neural networks usually include the following structures: feature map, which is called a pooling operation. Pooling is
input layer, convolutional layer, pooling layer, fully con- the use of filters to filter out invalid information in the fea-
nected layer, and output layer. The input layer is responsible ture map after convolution operation, saving important
for the input of the network, which is generally a matrix of information while reducing the amount of information and
numbers representing an image. The image input to the net- complexity. The pooling operation has strong advantages
work will then go through a series of convolutional layers, in reducing network parameters, reducing the amount of
 7396, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2022/6554371 by CAPES, Wiley Online Library on [27/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
4 Computational and Mathematical Methods in Medicine

computation, and enhancing the robustness of the network. network’s output as an input, does multiscale prediction
The pooling operation is as follows: using regression, and outputs the full network model’s detec-
tion findings. Figure 1 shows a schematic diagram of
 
ylj = down yl−1 : ð2Þ YOLOv3’s overall network topology. The backbone network
j
DarkNet-53 is shown on the left, while the detection net-
work on the right forecasts the outcomes based on
The pooling operation first divides the feature map into DarkNet-53’s output.
small matrices of equal size and does not overlap and then The process of performing target detection in the
selects a value to replace the entire small matrix according YOLOv3 network is as follows. The YOLOv3 network model
to different criteria. There are two commonly used criteria, first scales the image input to the network to a size of 416
one is max pooling, which is to keep the one with the largest × 416 × 3 and then inputs it into the backbone network to
element in each small matrix. The second is average pooling, extract feature maps of three scales. The dimensions of the
which is to average and retain all elements in each small three feature maps are 13 × 13 × 1024, 26 × 26 × 512, and
matrix. After the feature map is pooled, the number will 52 × 52 × 256, respectively. Then, the feature maps of these
not change, but its scale will be significantly reduced, and three scales are input into the detection network module,
only the most effective information in the feature map will the lower layer feature map performs upsampling operation,
be retained, thereby improving the robustness of the convo- and the upper layer feature map performs feature fusion.
lutional neural network model. Then, after a series of convolutional layers, three final feature
The fully connected layer is the most commonly used maps are obtained. The dimensions of the three final feature
hidden layer of the network before the convolutional layer maps are 13 × 13 × 255, 26 × 26 × 255, and 52 × 52 × 255,
is proposed. Each artificial neuron in the layer establishes a respectively, and these three feature maps are finally used
connection relationship with all the neurons in the previous to predict the results.
layer. The main feature is that there are many parameters DarkNet-53 is mainly responsible for extracting feature
and a large amount of calculation. In fact, it plays the role maps of different scales. The network performs feature
of integrating information through a large number of extraction through a series of 1 × 1 convolutions and 3 × 3
parameters and calculations. In a convolutional neural net- convolutions. A batch normalization layer is also included
work, one or more fully connected layers are connected after to standardize the data, speed up the network’s convergence,
multiple convolutional and pooling layers. Its main function and improve the network model’s training efficiency.
is to fuse the extracted high-level features and then trans- DarkNet-53 contains a total of 52 convolutional layers,
form these fused features into a probability distribution. which are composed of these 52 convolutional layers. The
The classification result of the network is output according first is through a convolutional layer containing 32 convolu-
to the probability distribution through the output layer. tional kernels of size 3 × 3 and then through 5 sets of
Activation functions are an important part of convolu- repeated residual block structures. Each residual block struc-
tional neural networks and are usually used in convolutional ture consists of a single convolutional layer with a set of
layers. The activation function is essentially a mathematical repeatedly executed convolutional modules. Before perform-
function, and its main task is to add nonlinear elements to ing the convolution operation performed separately, the zero
the network model. The activation function is created by padding operation must be completed first; that is, the zero
imitating the way the neurons in the human brain transmit expansion of the upper border and the left border of the
information; that is, when the stimulation of the neuron input image is first performed. Then, a convolution opera-
reaches a certain threshold, the neuron will be activated to tion with a convolution kernel size of 3 × 3 and a stride of
transmit the information of the neurons in the previous 2 is performed. The number of specific convolution kernels
layer. Activation functions work in a similar way. Common is related to the depth of the network. After that, the convo-
activation functions are as follows: lution module operation is repeated, and it is repeated 1
time, 4 times, 8 times, 8 times, and 4 times, respectively. In
1 each repeated convolution module, the convolution opera-
SigmoidðxÞ = ,
1 + e−x tion with the convolution kernel size of 1 × 1 and the num-
ex − e−x ð3Þ ber of convolutions is halved first. Then, perform a normal
TanhðxÞ = x −x , number of convolution operations with a kernel size of 3 ×
e +e
3. This cycle is repeated for many times; that is, the feature
ReLUðxÞ = max ð0, xÞ: extraction work is completed. Subsequently, a fully con-
nected layer completes the screening of feature maps. How-
3.2. Defect Detection of Agricultural Product Packaging ever, the fully connected layer is actually implemented by
Design with YOLOv3. The overall network structure of 1 × 1 convolution; so, it can also be recorded as a convolu-
YOLOv3 is divided according to functions, which can be tion layer. So far, the entire network contains 53 convolu-
divided into backbone network and detection network. The tional layers, from which DarkNer-53 is named.
backbone network is essentially a convolutional neural net- YOLOv3 outperforms YOLO and YOLOv2 in terms of
work, which is mainly responsible for the feature extraction backbone network performance. The backbone network,
of input images. The DarkNet-53 network is selected in the for starters, uses DarkNet-53 with deeper layers. DarkNet-
YOLOv3 model. The detection network takes the backbone 53 employs a significant number of residual structures in
 7396, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2022/6554371 by CAPES, Wiley Online Library on [27/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Computational and Mathematical Methods in Medicine 5

Input

Conv2D 3×3

Residual block 1×64 Conv2D block 5L 128 Conv2D 3×3+conv2D 1×1

Residual block 2×128 Concat Conv2D+upsampling2D

Residual block 8×256 Concat Conv2D block 5L 256

Residual block 8×512 Conv2D+upsampling2D Conv2D 3×3+conv2D 1×1

Residual block 4×1024 Conv2D block 5L 1024 Conv2D 3×3+conv2D 1×1

Backbone network Detection network

Figure 1: The structure of YOLOv3.

Input
addition to layer deepening. A schematic diagram of the
residual structure’s operation flow is given in Figure 2. The Zero Padding
second point is that before DarkNet-53 performs the convo-
lution operation, it will first normalize the input data of the Conv 3×3
convolution layer and then send the processed data to the
convolution layer. The advantage of using this execution Conv 1×1
process is that it can greatly speed up the convergence speed
of the network training. Conv 3×3
For the detection network, we mainly rely on the feature
maps of different sizes output by DarkNet-53 for result pre- +
diction. YOLOv3 uses a total of 3 feature maps for object
detection. These three feature maps are located after the Output
residual convolution module repeated 8 times by the
DarkNet-53 network, after the residual convolution module Figure 2: The structure of residual module.
repeated 8 times, and after the residual convolution module
repeated 4 times. Then, use a series of convolution opera- tion of small targets. It is precisely because the feature
tions to further extract features. First, a 13 × 13 × 255 output maps of different scales are used in the YOLOv3 network
is obtained according to the feature map with the smallest model that the network can achieve better detection accu-
size, denoted as y1. A series of convolution operations and racy when detecting targets of different sizes.
an upsampling operation are performed on the underlying
feature map of DarkNet-53 to expand the size of the feature 3.3. Backbone Network Improvement. The original network
map. Then, perform a splicing operation with a feature map model of YOLOv3 was designed to conduct target detection,
with a size of 26 × 26 × 512 and then go through a series of with the aim of detection being everyday items, which is
convolution operations to obtain a feature map with a size considerably different from the target job in this topic. This
of 26 × 26 × 255, denoted as y2. Repeat this operation to subject’s goal is to find flaws in the graphic design of agricul-
get a new feature map, and the output size is 52 × 52 × 255, tural product packaging, and the features of numerous flaws
denoted as y3. differ significantly from those of the original target object.
Finally, three different scales of y1, y2, and y3 are simul- Continuing to use the original network model, although bet-
taneously input into the subsequent network for the final ter detection results can be achieved, it is still not optimal. In
result prediction. Among them, the feature map of y1 is view of this, this topic will carefully study and analyze the
small and contains high-level feature information, which is defect characteristics and make certain improvements to
mainly responsible for the prediction of large targets. The the original YOLOv3 detection algorithm accordingly.
feature map of y2 is of moderate size and is mainly respon- To improve the detection algorithm of YOLOv3, we first
sible for the prediction of medium-sized objects. The feature need to improve the backbone network DarkNet-53. The final
map of y3 is the largest, which contains more detailed fea- output feature map size of the DarkNet-53 feature extraction
ture information, and is mainly responsible for the predic- network is 13 × 13, and the feature map of this scale is
 7396, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2022/6554371 by CAPES, Wiley Online Library on [27/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
6 Computational and Mathematical Methods in Medicine

responsible for the detection of large objects. However, in this Input

project, some defects have a size of more than 1000 pixels, and
the detection effect of this feature map may be difficult to meet
Conv2D 32×3×3
the project requirements. Therefore, this topic decided to add
a residual convolution module on the basis of DarkNet-53, the
internal convolution module is repeated 4 times, and the entire Residual block 1×64
added module contains a total of 9 convolution operations.
Aside from adding more convolutional layers, the number of
Residual block 2×128
repetitions of the residual convolution module of the feature
extraction network must be adjusted, as well as the number
of convolution kernels inside the residual convolution module. Residual block 4×256
The new backbone network contains a total of 62 convolu-
tional layers, which can be called DarkNet-62. The number Residual block 8×512
of network layers is deepened, and the expressive ability of
the network is enhanced. Figure 3 is a schematic diagram of
the DarkNet-62 network model. Residual block 8×1024

3.4. Multiscale Improvement. The original network model Residual block 4×512
selected 3 feature maps for the final detection result predic-
tion. And the small-scale feature map is responsible for the Figure 3: The structure of DarkNet-62.
detection task of large objects, and the large-scale feature
map is responsible for the detection task of small objects. Table 1: Improved prior box.
Due to the unsatisfactory results of previous experiments,
this project plans to improve the selection of feature maps, Priori box Feature map size
increasing from 3 feature maps to 5 feature maps. In the pre- 352 × 1127, 465 × 272, 725 × 787 6×6
vious subsection, a residual convolution module has been 189 × 797, 228 × 195, 296 × 62 13 × 13
added, taking the output of this part as one of the selected
104 × 42, 122 × 103, 147 × 132 26 × 26
feature maps. Then, the output of the second residual convo-
lution module of the network model is selected as one of the 57 × 172, 76 × 77, 87 × 427 52 × 52
feature maps, where the size of the feature map is 104 × 104. 27 × 26, 42 × 40, 52 × 62 104 × 104
And keep the original 3 feature maps and complete the
screening of the improved YOLOv3 model feature maps. A
Table 2: Experimental environment information.
total of 5 feature maps participate in subsequent detection.
In the subsequent feature fusion process, since the 6 × 6 fea- Name Parameter
ture map has a size of 12 × 12 after upsampling, it does not
CPU Intel i9-9900K
match the 13 × 13 feature map; so, the upsampled feature
map needs to be filled. Here, the method of zero padding is used GPU GeForce RTX 2080Ti(11GB)
to fill the upper and left sides of the 12 × 12 feature map to Memory 32GB
ensure that the subsequent feature maps can be aligned. Framework PyTorch 1.6

3.5. Priori Box Improvement. The YOLOv3 target detection

network follows the a priori frame design of the YOLOv2 3.6. Activation Function Adjustment. The Leaky ReLU acti-
network. In the original YOLOv3 algorithm model, there vation function used in the original YOLOv3 network is an
are 9 a priori boxes calculated on the dataset using the K improved version of the ReLU activation function. But in
-means algorithm. Most of the COCO datasets are common the negative semiaxis region, Leaky ReLU presets a nonzero
objects and animals in life, such as TV shows, chairs, cars, slope to ensure that the output of the function on the nega-
cats, and dogs. YOLOv3’s initial a priori frame is capable tive semiaxis is not equal to 0. This ensures that all neurons
of detecting these things. The flaws in this topic are some can participate in network training, and parameters can be
new YOLOv3 targets, and the original a priori frame is not updated in time. However, since the slope of the negative
up to the task of detecting them. semiaxis region is preset and fixed, this makes it impossible
Therefore, this project plans to reset the prior frame suit- to find the perfect parameter values that are exactly suitable
able for the agricultural product packaging graphic design for this topic, and there is no guarantee that the final result
defect dataset. And on the basis of the original 3, it has been of the network is optimal.
increased to 5. The reason is that the different types of In view of this, this project plans to use PReLU as the
defects in this subject are quite different, and the sizes of activation function of this defect detection network model.
other types of defects are also different. As a result, addi- The advantage of PReLU is that the slope of the negative
tional a priori frames are required to improve flaw detection semiaxis region is a parameter that can be learned, not pre-
accuracy in this subject. Table 1 shows the distribution of the set, but determined according to the data during the training
enhanced prior frame on the feature map. of the network model. Moreover, the parameter values are
 7396, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2022/6554371 by CAPES, Wiley Online Library on [27/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Computational and Mathematical Methods in Medicine 7

100

98

96

94

Precision and recall

92

90

88

86

84

82

80
Precision Recall

No BNI
Have BNI

Figure 4: Result of backbone network improvement.

100

98

96

94
Precision and recall

92

90

88

86

84

82

80
Precision Recall
No MSI
Have MSI

Figure 5: Result of multiscale improvement.

fixed when the model is trained and are also constant in the improvement, result of priori box improvement, and result
subsequent testing process. The expression of PReLU is as of activation function improvement in detail.
follows:
4.1. Dataset and Detail. This work uses a self-made agricul-
8 tural product packaging graphic design dataset. This dataset
< x, x>0
contains a total of 98,472 samples, of which 66,938 samples
PReLUðxÞ = x : ð4Þ
: , others are training sets, and the remaining 31,534 samples are test
a sets. Precision and recall are used as evaluation metrics for
this work. The experimental environment information is
4. Experiments illustrated in Table 2.

In this section, we define the dataset and detail, result of 4.2. Result of Backbone Network Improvement. As men-
backbone network improvement, result of multiscale tioned earlier, this work improves the backbone network.
 7396, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2022/6554371 by CAPES, Wiley Online Library on [27/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
8 Computational and Mathematical Methods in Medicine

100

98

96

94

Precision and recall

92

90

88

86

84

82

80
Precision Recall
No PBI
Have PBI

Figure 6: Result of priori box improvement.

100

98

96

94
Precision and recall

92

90

88

86

84

82

80
Precision Recall

No AFI
Have AFI

Figure 7: Result of activation function improvement.

To verify the effectiveness of this improvement strategy 4.3. Result of Multiscale Improvement. As mentioned earlier,
(BNI), this work conducts comparative experiments to com- this work improves the multiscale prediction. To verify
pare the defect detection performance of agricultural pack- the effectiveness of this improvement strategy (MSI), this
aging graphic design without and with BNI improvement, work conducts comparative experiments to compare the
respectively. The experimental results are illustrated in defect detection performance of agricultural packaging
Figure 4. graphic design without and with MSI improvement,
Compared with the non-BNI improvement strategy, respectively. The experimental results are illustrated in
after using the BNI strategy, the detection network can Figure 5.
obtain 2.1% precision and 1.5% recall improvement. This Compared with the non-MSI improvement strategy,
can prove the effectiveness and correctness of this work after using the MSI strategy, the detection network can
using the BNI improvement strategy. obtain 1.5% precision and 1.1% recall improvement. This
 7396, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2022/6554371 by CAPES, Wiley Online Library on [27/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Computational and Mathematical Methods in Medicine 9

can prove the effectiveness and correctness of this work Data Availability
using the MSI improvement strategy.
The datasets used during the current study are available
4.4. Result of Priori Box Improvement. As previously stated, from the corresponding author on reasonable request.
this effort enhances the priori box. This work performs com-
parison experiments to compare the defect detection perfor- Conflicts of Interest
mance of agricultural package graphic design without and
with PBI improvement, respectively, to validate the useful- The authors declare that they have no conflicts of interest.
ness of this improvement approach (PBI). The experimental
results are illustrated in Figure 6. Funding
Compared with the non-PBI improvement strategy, after
using the PBI strategy, the detection network can obtain This paper is part of the provincial “University Student
1.7% precision and 1.3% recall improvement. This can prove Innovation and Entrepreneurship Training Program” pro-
the effectiveness and correctness of this work using the PBI ject of Northwest A&F University: the research results of
improvement strategy. “Mintian Agricultural Products Marketing Planning Studio”.

4.5. Result of Activation Function Improvement. As men- References

tioned earlier, this work improves the activation function. To
verify the effectiveness of this improvement strategy (AFI), [1] T. C. Bagiritima, J. M. Tesha, and M. Kimani, “Investigation on
this work conducts comparative experiments to compare the the poor computer graphic design skills among art and design
defect detection performance of agricultural packaging students at university,” International Journal of Humanities
graphic design without and with AFI improvement, respec- Social Sciences and Education (IJHSSE), vol. 6, no. 10,
tively. The experimental results are illustrated in Figure 7. pp. 61–71, 2019.
Compared with the non-AFI improvement strategy, [2] P. B. Siti, M. Mustaji, B. Bachri, and F. Patricia, “Building
after using the AFI strategy, the detection network can empathy: exploring digital native characteristic to create learn-
obtain 1.0% precision and 0.8% recall improvement. This ing instruction for learning computer graphic design,” Inter-
national Journal of Emerging Technologies in Learning (iJET),
can prove the effectiveness and correctness of this work
vol. 15, no. 20, pp. 145–159, 2020.
using the AFI improvement strategy.
[3] B. Ji, Y. Li, D. Cao, C. Li, S. Mumtaz, and D. Wang, “Secrecy
performance analysis of UAV assisted relay transmission for
5. Conclusion cognitive network with energy harvesting,” IEEE Transactions
on Vehicular Technology, vol. 69, no. 7, pp. 7404–7415, 2020.
Computer graphics and image processing technology are [4] J. Li, Z. Zhou, J. Wu et al., “Decentralized on-demand energy
maturing in tandem with the rapid advancement of informa- supply for blockchain in internet of things: a microgrids
tion technology. In graphic design, the use of computer approach,” IEEE transactions on Computational Social Sys-
graphics and image processing technology can help to simplify tems, vol. 6, no. 6, pp. 1395–1406, 2019.
the process, enhance the design effect, and better the design [5] Y. Chen, W. Zou, and A. Sharma, “Graphic design method
expression. People’s expectations for graphic design quality based on 3D virtual vision technology,” Recent Advances in
are rising in tandem with the manufacturing industry’s rapid Electrical & Electronic Engineering (Formerly Recent Patents
growth; although, flaws are unavoidable in the design process. on Electrical & Electronic Engineering), vol. 14, no. 6,
Graphic design defects not only affect the appearance itself but pp. 627–637, 2021.
also affect the use; so, enterprises pay special attention to prod- [6] J. F. I. Nturambirwe and U. L. Opara, “Machine learning appli-
uct quality inspection to ensure the quality of products. At the cations to non-destructive defect detection in horticultural
same time, the quality inspection results are analyzed to fur- products,” Biosystems Engineering, vol. 189, pp. 60–83, 2020.
ther improve the graphic design process and reduce the occur- [7] Z. Ren, F. Fang, N. Yan, and Y. Wu, “State of the art in defect
rence of defects. As a result, this study examines the use of a detection based on machine vision,” International Journal of
Precision Engineering and Manufacturing-Green Technology,
convolutional neural network as a computer graphics and
vol. 9, pp. 661–691, 2021.
image processing technology in defect identification in graphic
[8] T. Wang, Y. Chen, M. Qiao, and H. Snoussi, “A fast and robust
design, as well as the current position. To complete the
convolutional neural network-based defect detection model in
enhancement of the YOLOv3 network model, this work uses product quality control,” The International Journal of
the packaging design of agricultural products as an example. Advanced Manufacturing Technology, vol. 94, no. 9-12,
The network model is presented to improve the direction of pp. 3465–3471, 2018.
this topic based on the experimental findings of the original [9] T. Bo, K. Jianyi, and W. Shiqian, “Review of surface defect
YOLOv3 and an in-depth analysis of the characteristics of var- detection based on machine vision,” Journal of Image and
ious agricultural product packaging graphic design faults. It Graphics, vol. 22, no. 12, pp. 1640–1663, 2017.
includes the improvement of the backbone network, the [10] J. Yang, S. Li, Z. Wang, and G. Yang, “Real-time tiny part
improvement of the feature map, the improvement of the defect detection system in manufacturing using deep learn-
prior frame, and the improvement of the activation function. ing,” IEEE Access, vol. 7, pp. 89278–89291, 2019.
Finally, the validity and feasibility of the model improvement [11] S. Ghorai, A. Mukherjee, M. Gangadaran, and P. K. Dutta,
are proved by experiments. “Automatic defect detection on hot-rolled flat steel products,”
 7396, 2022, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2022/6554371 by CAPES, Wiley Online Library on [27/08/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
10 Computational and Mathematical Methods in Medicine

IEEE Transactions on Instrumentation and Measurement, [28] T. W. Liao and K. Tang, “Automated extraction of welds from
vol. 62, no. 3, pp. 612–621, 2013. digitized radiographic images based on MLP neural networks,”
[12] D. Guyer and X. Yang, “Use of genetic artificial neural net- Applied Artificial Intelligence, vol. 11, no. 3, pp. 197–218, 1997.
works and spectral imaging for defect detection on cherries,” [29] X. Xie, “A review of recent advances in surface defect detection
Computers and Electronics in Agriculture, vol. 29, no. 3, using texture analysis techniques,” ELCVIA: electronic letters
pp. 179–194, 2000. on computer vision and image analysis, vol. 7, no. 3, pp. 1–
[13] Q. Jin and L. Chen, “A survey of surface defect detection of 22, 2008.
industrial products based on a small number of labeled data,” [30] D. Tabernik, S. Šela, J. Skvarč, and D. Skočaj, “Segmentation-
2022, https://arxiv.org/abs/2203.05733. based deep-learning approach for surface-defect detection,”
[14] N. L. Maskuri, M. H. Abu Bakar, and A. K. Ismail, Three- Journal of Intelligent Manufacturing, vol. 31, no. 3, pp. 759–
Dimensional Image Reconstruction for Automated Defect 776, 2020.
Detection at Artificial Metallic Surface Specimens, Progress in [31] P. Napoletano, F. Piccoli, and R. Schettini, “Anomaly detection
Engineering Technology III, [M.S. thesis], Springer, Cham, in nanofibrous materials by CNN-based self-similarity,” Sen-
2021. sors, vol. 18, no. 2, p. 209, 2018.
[15] X. Xu, J. Chen, H. Zhang, and W. W. Y. Ng, “D4Net: de- [32] T. S. Newman and A. K. Jain, “A survey of automated visual
deformation defect detection network for non-rigid products inspection,” Computer Vision and Image Understanding,
with large patterns,” Information Sciences, vol. 547, pp. 763– vol. 61, no. 2, pp. 231–262, 1995.
776, 2021. [33] K. D’huys, W. Saeys, and K. B. De, “Active infrared thermog-
[16] H. İ. Çelik, L. C. Dülger, and M. Topalbekiroğlu, “Develop- raphy for seal contamination detection in heat-sealed food
ment of a machine vision system: real-time fabric defect detec- packaging,” Journal of Imaging, vol. 2, no. 4, p. 33, 2016.
tion and classification with neural networks,” The Journal of
The Textile Institute, vol. 105, no. 6, pp. 575–585, 2014.
[17] X. Sun, X. Jiang, Y. Fu, C. Han, and M. Wen, “Surface defect
detection system for camshaft based on computer vision,”
Infrared and Laser Engineering, vol. 42, no. 6, pp. 1647–1653,
2013.
[18] X. Zhou, Y. N. Wang, Q. Zhu, C. Z. Wu, and Y. Peng,
“Research on defect detection method for bottle mouth based
on machine vision,” Journal of Electronic Measurement and
Instrument, vol. 30, no. 5, pp. 702–713, 2016.
[19] D. M. Tsai and C. T. Lin, “Fast normalized cross correlation for
defect detection,” Pattern Recognition Letters, vol. 24, no. 15,
pp. 2625–2631, 2003.
[20] S. Yin, H. Bao, B. Ji, J. Liu, and J. Zhang, “Defect detection of
engine cylinder based on small region template matching,”
Transducer and Microsystem Technologies, vol. 31, no. 6,
pp. 143–145, 2012.
[21] J. Liang, T. Hon, and D. Lin, “Design of smart meter visual
inspection system based on machine vision,” Electrical Mea-
surement & Instrumentation, vol. 10, pp. 64–68, 2013.
[22] Y. Hong and Y. Yang, “Keyboard defects detection based on
template matching method,” DEStech Transactions on Engi-
neering and Technology Research, vol. 24, no. iect, pp. 317–
320, 2016.
[23] F. Y. Zhang and R. J. Liu, “Study on the parts surface defect
detection method based on modifed SVM algorithm,” Applied
Mechanics and Materials, vol. 541-542, pp. 1447–1451, 2014.
[24] X. X. Wang, W. J. Zhou, and J. W. Xu, “Research on detection
system for TFT-LCD defects,” Journal of E-lectronic Measure-
ment and Instrument, vol. 28, no. 3, pp. 278–284, 2014.
[25] H. Abdellah, R. Ahmed, and O. Slimane, “Defect detection and
identification in textile fabric by SVM method,” IOSR Journal
of Engineering, vol. 4, no. 12, pp. 69–77, 2014.
[26] T. Ojala, M. Pietikainen, and T. Maenpaa, “Multiresolution
gray-scale and rotation invariant texture classification with
local binary patterns,” IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 24, no. 7, pp. 971–987, 2002.
[27] Y. Zhang, Y. Wang, and H. Jiang, “Edge detection of image
based on Sobel and wavelet transformation,” Computer Appli-
cations and software, vol. 24, no. 4, pp. 133-134, 2007.