Advances in Computer, Signals and Systems (2023) DOI: 10.23977/acss.2023.

070301
Clausius Scientific Press, Canada ISSN 2371-8838 Vol. 7 Num. 3

Research on an Improved YOLOV8 Image Segmentation

Model for Crop Pests
Jichang Kang, Lianjun Zhao, Kangtao Wang, Kunpeng Zhang
Shandong University of Technology, Zibo, Shangdong, 255000 China

Keywords: YOLOV8, Simulated annealing algorithm, Crop pests, Image segmentation

Abstract: With the change of ecosystem, there are more and more kinds of crop diseases
and insect pests, and the harm is becoming more and more serious. Preventing crop
diseases and insect pests is the premise to ensure crop yield. Image segmentation
technology is to divide a number of specific targets and regions with different
characteristics in the image according to the requirements through pixel-level classification
scheme, which is the first important link of image analysis. In this article, Simulated
Annealing (SA) algorithm is used to optimize YOLOV8. The main purpose is to randomly
find the optimal solution of the loss function in the last layer of convolutional neural
network (CNN) with SA algorithm, and then update the weights and offsets of the previous
layer with this solution. The CNN structure also uses the dropout regularization method to
effectively reduce the influence of over-fitting. The simulation results show that compared
with YOLOV7 algorithm, the average accuracy of disease identification of improved
YOLOV8 is obviously higher. The pest identification model based on the improved
YOLOV8 algorithm has more advantages than YOLOV7 algorithm in both accuracy and
efficiency. The proposed method achieves the best detection performance on large-scale
public data sets, and also performs well in the task of crop pest detection studied in this
article.

1. Introduction

Although the development of agriculture has reached an unprecedented height at present, it is

inevitable that China's agriculture, which is in the process of modernization transformation, has also
encountered some problems that restrict its development. Among these problems, the most
important one is the increasingly frequent crop pests [1]. With the change of ecosystem, there are
more and more kinds of crop diseases and insect pests, and the harm is becoming more and more
serious. Preventing crop diseases and insect pests is the premise to ensure crop yield. Deep learning
is a machine learning method, which is realized by deep data fitting with neural network. The
learning process is multi-layered and includes many nonlinear changing processes [2]. The
traditional identification method is inefficient and the identification effect is not good, mainly
through some simple machine learning models to judge the types of crop diseases in the actual
environment. Although this method is simple and time-consuming, it does not achieve the ideal
identification effect for crop diseases in actual natural conditions [3]. Through a certain
combination, the original fuzzy features can be more abstracted, and the complex data relations can

1
be more intuitively simulated, so that the original data relations that are not tight enough become
closer [4]. With the help of deep learning, this advantage is more prominent. With the development
of neural network and computer vision, image semantics are constantly improving, and deep
learning has become the development trend in the field of artificial intelligence.
Image segmentation technology is to divide a number of specific targets and regions with
different characteristics in the image according to the requirements through pixel-level
classification scheme, which is the first important link of image analysis [5]. Gwo-Jiun et al. put
forward a disease location diagnosis algorithm with wheat as the research object. After wavelet
decomposition transformation and grain matrix calculation, the binary image of crop disease area
was obtained by automatic threshold algorithm, so as to calculate the color feature value and
compare it with the image in the disease database, and make a disease detection judgment [6].
Wang et al. preprocessed the corn disease image first, then constructed CNN by using the
Tripletloss function to extract and learn the feature information of the disease image, then extracted
the texture features of the disease image by using the SIFT algorithm, and finally classified the
image by using the Softmax function [7]. At present, the automatic identification and diagnosis
method of crop diseases based on machine vision has become the mainstream technology. However,
in the application of related technologies in the actual production of agricultural field crops, there
are still many problems in the effective and high-speed segmentation and acquisition of disease
spots in crop disease images. In this article, an improved YOLOV8 based on SA algorithm is
proposed for the segmentation and extraction of disease spots in crop disease images under the
complicated background of agricultural field production. The SA algorithm is used to randomly find
the optimal solution of the loss function in the last layer of CNN, and then the weight and bias of
the previous layer are updated by this solution.

2. Disease Image Segmentation Based on Improved YOLOV8

The traditional dropout method forces a neural unit to work with other randomly selected neural
units to achieve good results. It weakens the joint adaptability between units, thus enhancing the
generalization performance of the structure [8]. In the standard neural network, the correlation
between nodes enables them to cooperate to trim the noise in other nodes, but these cooperations
cannot be generalized, so the over-fitting problem arises, and dropout destroys this correlation [9].
In this article, the pool layer and full connection layer in CNN structure use dropout regularization
to suppress some activation units. Using dropout in max-pooling layer can avoid the above
shortcomings, because dropout may suppress the values of some neurons, so the values obtained by
pooling are random, instead of always being the average or the maximum of all values in the pooled
area. The images of crop diseases collected under actual natural conditions are greatly influenced by
illumination, because illumination is a factor that has to be considered in the external natural
environment.
Border regression is used to adjust the regional suggestion box to make it closer to the real
border box, which will make the positioning more accurate. The measured image will generate two
boxes, one is to mark the true value, and the other is the target detection box generated by the target
detection algorithm. According to the overlapping ratio of the target detection frame and the real
frame, it is compared with the set threshold parameters. The detection window of diseased leaves is
shown in Figure 1.

2
 Fig.1 Detection Window of Diseased Leaves
It is known that the thermal point set of disease image is 𝑃, 𝑃𝑖𝑉 is the thermal value
information of thermal point 𝑃𝑖 , and 𝑃𝑖 is calculated by correlation coefficient function.
Calculating the thermal values of four vertices of the disease image;
2
|𝑄 𝑃 |
𝑉𝑖 = ∑𝑛𝑗=1 𝑛 𝑖 𝑗 2 ⋅ 𝑃𝑖𝑉 (1)
∑𝑘=1|𝑄𝑖 𝑃𝑘 |
2
Where |𝑄𝑖 𝑃𝑗 | is the square of the distance between vertex 𝑄𝑖 and thermal point 𝑃𝑖 .
Dropout will inhibit the values of some nerve cells, and the values obtained by pooling will be
random, which is no longer always the average of all values in the pooled area or the maximum
value [10]. In the training process, the input after each dropout is different, and more models can be
trained at the same time and with the same data. The main process of SA algorithm is to solve a
large range of combinations and randomly search the global optimal solution, which has the
characteristics of asymptotic convergence, parallelism, flexibility, easy implementation and fast
operation speed [11]. This algorithm is the process of searching for the optimal solution, that is, the
process of SA, in the process of reasonably adjusting the temperature drop. SA can not only find the
optimal solution, but also operate quickly.
SA is a heuristic method to find the optimal solution randomly. Because of adding appropriate
uncertain elements to the query, it will not only accept the good solution, but also receive the
solution which is actually worse than the current solution according to a probability value, and this
value will gradually decrease with the passage of time. This operation makes the value in the
neighborhood of the solution uncertain, so this method may avoid the interval optimal solution and
finally get the global optimal solution. SA is used to train CNN to get the optimal solution that
meets the conditions. The objective function is:
1
𝑄 = − ∑𝑁 𝑛 𝑛 𝑛 𝑛
𝑛=1(𝑦 𝑙𝑜𝑔(𝑜 ) + (1 − 𝑦 ) 𝑙𝑜𝑔(1 − 𝑜 )) (2)
𝑁

Where 𝑦 represents the expected output, 𝑜 represents the actual output, and 𝑁 represents the
number of samples in one training. There are two ways to terminate the training process: one is to
reach the set maximum number of iterations; Second, when the loss function is less than a certain
constant. Satisfying any criterion means reaching the optimal state.
Different light intensities on crop disease images have a great influence on feature extraction of
disease spots, so in order to make CNN have strong adaptability, namely generalization ability, the
crop disease images collected under actual natural conditions are processed and adjusted in two
aspects: image brightness and contrast [12]. Through this illumination transformation, the data
volume is enriched, so that CNN can obtain diversified images of the same disease and improve the
effect of subsequent image segmentation and recognition. Only when the feature intensity of a
certain region in the image reaches a specified value can the convolution kernel extract features in
this region, and the training of feature extraction methods will not be affected by other regions. In

3
reality, data is usually distributed nonlinearly, and the activation layer enables CNN to learn
nonlinear mapping. In an image, the relative position between features is more important than the
specific position of a specific feature in the whole image. Pool layer makes the data space smaller,
which indirectly prevents the occurrence of over-fitting phenomenon. For CNN, the last part is the
full connection layer. The fully connected layer includes three parts: input layer, hidden layer and
output layer. The output layer is usually the softmax layer. The output result of the full connection
layer is a probability value, which indicates the possibility that the input of CNN belongs to each
category. The CNN model of crop pest image segmentation is shown in Figure 2.

Fig.2 Cnn Model of Crop Pest Image Segmentation

In standard CNN, one of the reasons for the poor performance of small target detection comes
from the receptive field with limited convolution operation. Therefore, the spatial attention module
aims to make the network pay more attention to the position of small target pests by learning the
target weight of the global feature map in each spatial position. Contrary to the channel attention
module, the spatial attention module needs to eliminate the influence of channel information.
Because the image contains the labeling information of pest targets in spatial position, the spatial
attention module is trained through supervised learning. The convolution neural network function is
defined as:
𝑥𝑗𝑙 = 𝑓 (∑𝑖∈𝑀𝑗 𝑥𝑖𝑙−1 × 𝑘𝑖𝑗
𝑙
+ 𝑏𝑗𝑙 ) (3)
Where 𝑥𝑖 represents the input characteristic map, 𝑘 represents the convolution kernel, 𝑏
represents the deviation term, and the convolution output is the characteristic map 𝑥𝑗 . Assume that
the convolution layer convolves the input insect pest image with 𝑘 filters, and generates 𝑘 new
feature maps for subsequent processing. If the output feature map is represented in a layer, then:
(𝑛) (𝑛) (𝑛−1) (𝑛)
𝐹𝑗 = ∑𝑖 𝑤𝑖𝑗 ∗ 𝐹𝑖 + 𝑏𝑗 (4)
(𝑛) (𝑛)
Where: ∗ is a two-dimensional convolution; 𝑤𝑖𝑗 and 𝑏𝑗 are convolution filters and
(𝑛)
deviations, respectively; 𝐹𝑗 is the 𝑗 output characteristic map at the 𝑛 layer. The formula of
active layer after convolution is as follows:
(𝑛+1)
𝐹𝑗 = 𝑓(𝐹𝑗𝑛 ) (5)
Where: 𝑓 is a point-by-point activation function. Convert each data item 𝑥𝑖 in the small batch
𝐵 = {𝑥1 , 𝑥2 , 𝑥3 , … , 𝑥𝑚 } with size 𝑚 to 𝑦𝑖 :
𝑦𝑖 = 𝛾𝑥̑ 𝑖 + 𝛽 (6)
𝑥𝑖 −𝐸𝑀 (𝑥𝑖 )
𝑥̑ 𝑖 = (7)
√𝑉𝑎𝑟𝑀 (𝑥𝑖 )+𝜀

Where: 𝐸𝑀 (𝑥𝑖 ) and 𝑉𝑎𝑟𝑀 (𝑥𝑖 ) are the mean and variance.
When the same crop disease is collected at different time points in a day, the appearance of the

4
disease images is different, even sometimes very different, the biggest difference is reflected in the
color and brightness of the images. Dropout used in training is different every time, and multiple
models are trained at the same time and with the same data. In the test stage, a new model average
method is proposed to pool the probability of unit value and the influence of P value in the region to
solve the average value predicted by the model.

3. Result Analysis and Discussion

Before the target detection task, image processing technology can be used to preprocess the
image of crop disease leaves. Image preprocessing technology mainly weakens or even removes the
interference information in the picture through mathematical operation and image transformation,
strengthens the expression of important information, and makes the subsequent image feature
extraction more rapid and accurate, and enhances the detectability. The essence of pest detection
task is the subtask of general target detection. Aiming at AgriPest pest detection data set, this article
adopts the evaluation index of general target detection as the basis. In this article, the deep learning
pretreatment of crop disease images collected under actual natural conditions can increase the
number of data sets, for example, an image will generate at least two more images after illumination
transformation and normalization, which is reflected in the data set construction of this article. On
the other hand, the deep learning network can better extract the lesion features in crop disease
images during training, so that CNN with superior performance can be trained and the ability of
segmentation and recognition of crop diseases can be improved.
The leaves of crop diseases usually have the characteristics of uneven illumination and different
leaf sizes, so the selection of feature extraction, image segmentation and recognition methods is
very important, which directly affects the detection speed and accuracy of later image processing.
Compared with the high-level statistical features, the image model of crop pests constructed by
using the bottom features can get higher accuracy. The time-consuming of pest image segmentation
using different methods is compared and analyzed, as shown in Figure 3.

Fig.3 Time-Consuming Segmentation of Pest Images by Different Methods

As can be seen from Figure 3, the time-consuming of pest image segmentation processing based
on YOLOV7 algorithm increases with the increase of the number of pixel points of feature
information, which takes a long time. However, the time-consuming of pest image segmentation
based on improved YOLOV8 has an upward trend, and it has obvious advantages compared with
YOLOV7 algorithm. In computer operation, the time required for each matching is the same, so
reducing the number of matching times can reduce the time required for image recognition and

5
make the recognition results display faster.
The image feature information with differences is fused, so that the fused feature information can
be easily distinguished, and the expressive ability of the image is improved. The fused features are
washed by channels, so that each region of the feature map contains the features of different
channels, and the obtained image fusion information is more robust. 9610264 takes the average
accuracy of crop pest image segmentation as the test index, and selects YOLOV7 algorithm as the
comparison object. The experimental results are shown in Table 1 and Table 2.
Table 1 Average Accuracy of Pest Image Segmentation Based on Improved YOLOV8 Algorithm
Sample size Accuracy (%)
15 99.74
30 99.45
45 98.76
60 98.15
75 97.59
90 96.89
105 95.74
Table 2 Average Accuracy of Pest Image Segmentation Based on Yolov7 Algorithm
Sample size Accuracy (%)
15 97.75
30 96.69
45 95.48
60 94.37
75 92.77
90 91.71
105 90.52
From the experimental data, it can be seen that when the number of test samples begins to
increase, the average accuracy of disease identification of the two methods has a certain downward
trend. However, compared with YOLOV7 algorithm, the average accuracy of disease identification
of the improved YOLOV8 algorithm is obviously higher.
The classification effect of CNN on crop leaf images depends on the size of convolution kernel.
The larger the convolution kernel, the wider the receptive field of CNN, which is more convenient
for analyzing the global information of crop leaf images, but at the same time, the detailed features
of the images will be ignored. The smaller the convolution kernel is, the easier it is to obtain the
detailed features of the image, but the global information of the crop leaf image cannot be obtained.
Therefore, in order to obtain accurate feature information of images, multi-convolution kernels are
needed. The result of precision test using YOLOV7 algorithm is shown in Figure 4. The result of
precision test using the improved YOLOV8 algorithm is shown in Figure 5.

6
 Fig.4 Accuracy Test Results of YOLOV7 Algorithm

Fig.5 Accuracy Test Results of Improved YOLOV8 Algorithm

The points on the graph represent the ratio of the predicted value to the actual value. The closer
the predicted value is to the actual value, the closer the square point is to the straight line y=x. On
the contrary, the greater the difference between the predicted value and the actual value, the farther
the square point deviates from the straight line y=x. The comprehensive experimental results show
that the pest identification model based on the improved YOLOV8 algorithm is superior to
YOLOV7 algorithm in both accuracy and efficiency.

4. Conclusions

At present, the automatic identification and diagnosis method of crop diseases based on machine
vision has become the mainstream technology. However, in the application of related technologies
in the actual production of agricultural field crops, there are still many problems in the effective and
high-speed segmentation and acquisition of disease spots in crop disease images. Convolutional
network model has the characteristics of partial receptive field, hierarchical structure, combined
extraction process and classification operation, and has many achievements in image processing and
recognition research. In the traditional network, the learning rate is a global constant. Choosing a
large learning rate is not conducive to getting the minimum value of the loss function, while a small
learning rate will consume a lot of training time. In this article, an improved YOLOV8 based on SA

7
algorithm is proposed for the segmentation and extraction of disease spots in crop disease images
under the complicated background of agricultural field production. The comprehensive
experimental results show that the pest identification model based on the improved YOLOV8
algorithm is superior to YOLOV7 algorithm in both accuracy and efficiency. This method is
effective and more practical, and it can quickly converge and gradually reach the best, so it has
better feasibility and effectiveness. The CNN based on adaptive learning rate algorithm proposed in
this article is compared with the traditional CNN. In the future research, more adaptive learning rate
algorithms can be studied, and compared with the methods proposed in this article, we can find the
differences of various methods and get a more efficient model.

References
[1] Jiang X, Zhen J, Miao J, et al. Assessing mangrove leaf traits under different pest and disease severity with
hyperspectral imaging spectroscopy. Ecological Indicators, vol. 129, no. 2, pp. 107901, 202.
[2] MN Rodrí guez-García, F Garcí a-Sánchez, R Valencia-Garcí a. Knowledge-Based System for Crop Pests and
Diseases Recognition. Electronics, vol. 10, no. 8, pp. 905, 2021.
[3] Shajahan S, Sivarajan S, Maharlooei M, et al. Identification and Counting of Soybean Aphids from Digital Images
Using Shape Classification. Transactions of the ASABE (American Society of Agricultural and Biological Engineers),
vol. 60, no. 5, pp. 1467-1477, 2017.
[4] Zhang M, Zhang W, Liang X, et al. Detection of fatigue crack propagation through damage characteristic FWHM
using FBG sensors. Sensor Review, vol. 40, no. 6, pp. 665-673, 2020.
[5] Kim J Y, Bellotti A, Alapati P, et al. Use of a non-collinear wave mixing technique to image internal microscale
damage in concrete. Journal of Applied Physics, no. 14, pp. 131, 2022.
[6] Gwo-Jiun, Horng, Min-Xiang, et al. The Smart Image Recognition Mechanism for Crop Harvesting System in
Intelligent Agriculture. IEEE Sensors Journal, vol. 20, no. 5, pp. 2766-2781, 2019.
[7] Wang B. Identification of Crop Diseases and Insect Pests Based on Deep Learning. Scientific Programming, vol.
2022, pp. 1-10, 2022.
[8] Raja R, Slaughter D C, Fennimore S A, et al. Crop signalling: A novel crop recognition technique for robotic weed
control ScienceDirect. Biosystems Engineering, vol. 187, pp. 278-291, 2019.
[9] Zou W, Shen C, Yin G. Application of image recognition technology in agricultural production process.
International Agricultural Engineering Journal, vol. 27, no. 2, pp. 318-326, 2018.
[10] Yu D, Zhang B, Zhao C, et al. Scene classification of remote sensing image using ensemble convolutional neural
network. Journal of Remote Sensing, vol. 24, no. 6, pp. 717-727, 2020.
[11] Wang D, Mao K. Task-generic semantic convolutional neural network for web text-aided image classification.
Neurocomputing, vol. 329, no. 3, pp. 103-115, 2019.
[12] Zhao M, Hu C, Wei F, et al. Real-Time Underwater Image Recognition with FPGA Embedded System for
Convolutional Neural Network. Sensors, vol. 19, no. 2, pp. 350, 2019.