Light-weight empoasca flavescens insect attack detection method based on deep learning

Abstract

The invention discloses a light-weight empoasca flavescens insect attack detection method based on deep learning, relates to the technical field of agricultural disease detection, and is used for solving the problem that the disease condition cannot be quickly and comprehensively mastered when tea empoasca flavescens is manually detected. Comprising the following steps: S1, making a tea lesser leafhopper original data set containing different environment variables in a real tea garden environment; s2, dividing the marked image data into a training set, a verification set and a test set; and S3, constructing an improved tea lesser leafhopper detection model based on Yolov. And S4, training the Yov-based improved tea lesser leafhopper detection model constructed in the step S3 by using the training set in the step S2 to obtain a final optimization model. And S5, detecting the tea lesser leafhoppers trapped in the tea garden by adopting the optimization model finally obtained in the step S4 to obtain a detection result. And S6, deploying and putting into use. The problems of time and labor consumption and low efficiency of artificial pest detection are solved, labor force is liberated, cost input is reduced, and practicability is high.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural disease detection, in particular to a light-weight tea green leafhopper pest detection method based on deep learning. BACKGROUND

[0002] As one of the main pests in tea gardens, tea green leafhoppers have a wide distribution area, a long damage time, and cause huge losses. Tea green leafhoppers are usually only a few millimeters long, and their body color is similar to that of tender tea branches and leaves. They often hide in tender tea shoots and leaf back parts, and are extremely difficult to detect due to their small size and protective color. At the same time, tea green leafhoppers have strong migratory ability and can quickly fly away when slightly disturbed, making it extremely difficult to capture them manually. In addition, tea green leafhoppers often inhabit dispersedly in tea gardens, with a wide distribution range and not concentrated, so it takes a lot of time and effort to check each plant manually. In a large area of tea garden, manual detection cannot quickly and comprehensively grasp the occurrence of tea green leafhoppers, making it difficult to make effective prevention and control decisions in a timely manner. Therefore, accurately identifying tea green leafhoppers is crucial for timely prevention and control, ensuring the healthy growth of tea trees and the production of high-quality tea, and timely guiding the prevention and control of tea green leafhoppers to ensure the healthy growth of tea trees and the production of high-quality tea. SUMMARY

[0003] The purpose of the present application is to provide a light-weight tea green leafhopper pest detection method based on deep learning, which solves the problem that manual detection of tea green leafhoppers cannot quickly and comprehensively grasp the disease situation.

[0004] The technical solution adopted by the present application to solve the technical problem is: a light-weight tea green leafhopper pest detection method based on deep learning, comprising the following steps:

[0005] S1. Making a tea green leafhopper original data set containing different environmental variables in a real tea garden environment.

[0006] Deploying a trap plate inside the tea garden to capture tea green leafhoppers, collecting tea green leafhopper data images on the trap plate; further cropping the collected original images into multiple small images, and using an image labeling platform to label the images as two categories of tea green leafhoppers and other insects.

[0007] S2. Dividing the labeled image data into a training set, a validation set, and a test set.

[0008] S3. Constructing a tea green leafhopper detection model based on an improved Yolov.

[0009] In the Yolov model backbone network, a scale dynamic attention module SDAM is added, a P2 detection head is used to replace a P5 detection head, a separation and enhancement attention mechanism SEAM is introduced into the detection head, a WaveletPool is introduced to replace part of the convolution and up-sampling in the original Yolov model, and an improved tea green leafhopper detection model based on Yolov is obtained.

[0010] S4, training the improved tea green leafhopper detection model based on Yolov constructed in step S3 with the training set in step S2 to obtain a final optimized model; and verifying the model performance on the test set.

[0011] S5, using the optimized model finally obtained in step S4 to detect tea green leafhoppers trapped in the tea garden to obtain a detection result.

[0012] S6, deployment for use.

[0013] Further, the image data in step S1 is collected in the form of shooting, a pest shooting device is selected, and the pest trapping plate is shot to collect tea green leafhopper images under different light, different shielding and different weather.

[0014] Further, considering the influence of sunlight color temperature and the number of pests on the trapping plate on the feature information of the collected images, data is collected every hour.

[0015] Further, the original image is cropped into 1360 images, and the images with poor quality are removed, and the labelimg software is used to mark the rectangular frame of the image, and the label is divided into two categories of EPM and Others, EPM representing tea green leafhopper and Others representing other insects.

[0016] Further, the ratio of the training set, the verification set and the test set is 7:1.5:1.5.

[0017] Further, the scale dynamic attention module SDAM dynamically allocates weights for different convolution kernel branches by calculating the standard deviation of the input feature map to measure the complexity of the feature map. The algorithm of the scale dynamic attention module SDAM includes: (1); (2); (3). In the formula, F is the input feature map, C is the channel number, A is the adaptive average pooling, M is the adaptive maximum pooling, a is the calculation of the standard deviation of the feature map, sigma is the Sigmoid activation function, * represents element-wise multiplication, m is the average value in the channel dimension, x is the maximum value in the channel dimension, and C iis the i-th branch convolution; is the corresponding dynamic weight; concat is a concatenation operation; F' is the final attention feature after being activated by a sigmoid function and multiplied by the input feature map; F'' is the output feature.

[0018] Further, the WaveletPool specific algorithm includes: (4); (5); (6). In the formula, is an approximation function; is a detail function; is an approximation coefficient; is a detail coefficient; h ψ is a time-reversed scale; h ψ is a time-reversed wavelet vector; n represents a sample in the vector, j represents a resolution; and K represents a position index.

[0019] The beneficial effects of the present application are: a scale dynamic attention mechanism SDAM is proposed, the complexity of the input feature map is measured by calculating the standard deviation, and then the weights are dynamically allocated for different convolution kernel branches to capture the subtle features of the tea green leafhopper. The separation and enhancement attention mechanism SEAM is introduced to realize the reconstruction of the detection head and the strengthening of the multi-scale detection, and the detection ability of the model for the multi-scale distribution of the tea green leafhopper is optimized. The WaveletPool network is used to replace part of the traditional convolution and up-sampling of the model, based on the time-frequency localization characteristics of wavelet transform, the tea garden image is decomposed into high-frequency details and low-frequency background, the tea green leafhopper and the background are effectively separated, and the feature signal-to-noise ratio is improved, while the feature extraction capability is improved, the calculation amount is greatly reduced. The present application solves the problems of time-consuming, labor-intensive and low efficiency in artificial pest detection, liberates labor, reduces cost investment, and has strong practicality. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is the improved model network result graph of the present application.

[0021] Figure 2 is the method implementation step flowchart of the present application.

[0022] Figure 3 is the tea green leafhopper detection result graph of the present application. DETAILED DESCRIPTION

[0023] A kind of light-weighted green leafhopper pest detection method based on deep learning of the present application will be described in detail below in conjunction with the drawings. The light-weighted green leafhopper pest detection method based on deep learning includes the following steps:

[0024] S1, make tea green leafhopper original data set containing different environmental variables under real tea garden environment.

[0025] The tea green leafhoppers are captured by deploying the insect trapping plate in the tea garden, and the data images of the tea green leafhoppers on the insect trapping plate are collected. The collected original images are further cropped into multiple small images, and the image data is labeled as two categories of tea green leafhoppers and other insects using an image labeling platform to truly reflect the insect situation in the tea garden scene.

[0026] Specifically, the image data in step S1 is collected in the form of shooting. In order to meet the diversity and authenticity of the tea green leafhopper recognition samples in the complex environment of the tea garden, the insect situation shooting device is selected to shoot the insect trapping plate, collect tea green leafhopper images under different lightings, different occlusions and different weathers, and consider the influence of different time periods, sunlight color temperature and insect trapping plate insect quantity on the collected image feature information. Data is collected every hour. After the original image is cropped, a total of 1360 images are generated. The images with poor quality are removed, and the images are labeled with a rectangular box using the labelimg software. The labels are divided into two categories of EPM and Others. EPM represents tea green leafhopper, and Others represents other insects.

[0027] S2, the labeled image data is divided into training set, validation set and test set according to the ratio of 7:1.5:1.5.

[0028] S3, a tea green leafhopper detection model based on improved Yolov8 is constructed.

[0029] The scale dynamic attention module SDAM is added to the Yolov8 model backbone network, the P2 detection head is used to replace the P5 detection head, and the detection ability of the model for small scale targets is optimized. The separation and enhancement attention mechanism SEAM is introduced in the detection head to further optimize the detection performance of the model for different scale targets, and the rich features extracted by the SDAM module are combined to enable the model to use more accurate feature information when detecting different scale targets, thereby improving the overall detection performance. The WaveletPool is introduced to replace part of the convolution and up-sampling in the original Yolov8 model, and the improved tea green leafhopper detection model based on Yolov8 is obtained.

[0030] Step S3 is to enhance the recognition ability of the model for small targets and overlapping targets of tea green leafhoppers, a kind of small pests, and improve the detection accuracy. The scale dynamic attention mechanism is proposed, and the detection head resolution optimization method is designed to enhance the expression ability of the model for detailed features. On this basis, the separation and enhancement attention mechanism SEAM is integrated; finally, the WaveletPool module is introduced to replace part of the convolution and up-sampling operation, reduce the aliasing effect, and reduce the model calculation amount, realize the lightweight of the model. The scale dynamic attention module measures the complexity of the feature map by calculating the standard deviation of the input feature map, dynamically allocates weights to different convolution kernel branches, and focuses on key features: (1); (2). (3). In the formula, F is the input feature map; C is the number of channels; A is adaptive average pooling, M is adaptive maximum pooling; a is the standard deviation of the calculation feature map; σ is the Sigmoid activation function; * represents element-wise multiplication; m is the average value in the channel dimension, x is the maximum value in the channel dimension; C i is the i-th branch convolution; ωi is the corresponding dynamic weight; concat is the concatenation operation; F' is activated by the sigmoid function to produce the final attention feature, which is multiplied by the input feature map; F'' is the output feature.

[0031] In a complex tea garden background, WaveletPool is introduced to effectively separate the effective features of tea green leafhoppers through a unique transformation method, enhance the clarity and recognition of the features, and solve the problem of difficult feature extraction in complex background by traditional methods. While improving the feature extraction capability, it greatly reduces the computational complexity. When processing tea garden insect plate images, it can quickly complete the feature extraction task, significantly improve the detection efficiency, and achieve the balance between detection efficiency and accuracy. The specific algorithm is as follows:

[0032] (4).

[0033] (5).

[0034] (6). In the formula, φ is the approximation function; ψ is the detail function; Wφ is the approximation coefficient; W ψ is the detail coefficient; hφ[−n] represents the time-reversed scale; h ψ [−n] represents the time-reversed wavelet vector; n represents the sample in the vector, j represents the resolution; K represents the position index.

[0035] S4, use the training set in step S2 to train the improved tea green leafhopper detection model based on Yolov8n constructed in step S3; and verify the model performance on the test set.

[0036] S5, use the optimized model finally obtained in step S4 to detect the tea green leafhoppers trapped in the tea garden to obtain the detection results.

[0037] S6, deploy and put into use.

[0038] The advantages of the application are: (1) a plug-and-play scale dynamic attention mechanism SDAM is proposed, the standard deviation of the input feature map is calculated to measure its complexity, and then the weights are dynamically allocated for different convolution kernel branches to capture the subtle features of the tea green leafhopper; (2) the improved model detection head is introduced into the separation and enhanced attention mechanism SEAM, the detection head is reconstructed and the multi-scale detection is strengthened, and the detection ability of the model to the multi-scale distribution of the tea green leafhopper is optimized; (3) the WaveletPool network is used to replace part of the traditional convolution and up-sampling of the model, based on the time-frequency localization characteristics of wavelet transform, the tea garden image is decomposed into high-frequency details and low-frequency background, the tea green leafhopper is effectively separated from the background, and the feature signal-to-noise ratio is improved, while the feature extraction ability is improved, and the calculation amount is greatly reduced; (4) the problem of artificial pest detection, time-consuming, labor-intensive and low efficiency is solved, the labor is liberated, the cost investment is reduced, and the practicability is relatively strong.

Claims

1. A lightweight method for detecting small green leafhopper pests based on deep learning, characterized in that, Includes the following steps: S1. Create the original dataset of the tea green leafhopper containing different environmental variables in a real tea garden environment; Insect-attracting boards were deployed in the tea garden to capture the tea green leafhopper. Data images of the tea green leafhopper on the insect-attracting boards were collected and cropped into multiple small images. Based on the image data, the leafhoppers were labeled as either tea green leafhoppers or other insects. S2. Divide the labeled image data into training set, validation set and test set; S3. Construct an improved tea green leafhopper detection model based on Yolov; A scale-dynamic attention module (SDAM) is added to the backbone network of the Yolov model. The P2 detection head is used to replace the P5 detection head. The Separation and Enhancement Attention (SEAM) mechanism is introduced into the detection head. WaveletPool is introduced to replace some convolutions and upsampling in the original Yolov model, resulting in a Yolov-based improved tea green leafhopper detection model. S4. Use the training set from step S2 to train the improved tea green leafhopper detection model based on Yolov constructed in step S3 to obtain the final optimized model; and verify the model performance on the test set. S5. The optimized model obtained in step S4 is used to detect the tea green leafhoppers trapped in the tea garden, and the detection results are obtained. S6, deployed and put into use.

2. The lightweight green leafhopper pest detection method based on deep learning according to claim 1, characterized in that, In step S1, image data is collected by taking pictures. Insect photography equipment is selected to photograph the insect-attracting board and collect images of the tea green leafhopper under different lighting, different shading, and different weather conditions.

3. The lightweight green leafhopper pest detection method based on deep learning according to claim 2, characterized in that, Considering the impact of sunlight color temperature and the number of insects on the insect-attracting boards on the feature information of the collected images at different times, data was collected every hour.

4. The lightweight green leafhopper pest detection method based on deep learning according to claim 3, characterized in that, The original image was cropped into 1360 images, and images of poor quality were removed. The image was labeled with rectangular boxes using the labelimg software, and the labels were divided into two categories: EPM and Others. EPM represents the tea green leafhopper, and Others represents other insects.

5. A lightweight method for detecting small green leafhopper pests based on deep learning according to claim 4, characterized in that, The ratio of training set, validation set and test set is 7:1.5:1.

5.

6. A lightweight method for detecting small green leafhopper pests based on deep learning according to claim 5, characterized in that, The Scale Dynamic Attention Module (SDAM) measures the complexity of the input feature map by calculating the standard deviation of the input feature map and dynamically assigns weights to different convolutional kernel branches. The algorithm for Scale Dynamic Attention Module (SDAM) includes: (1); (2); (3); where F is the input feature map; C is the number of channels; A is adaptive average pooling, M is adaptive max pooling; α is the standard deviation of the feature map; σ is the Sigmoid activation function; * represents element-wise multiplication; m is the average value in the channel dimension, x is the maximum value in the channel dimension; C i For the i-th branch convolution; ωi represents the corresponding dynamic weights; concat is the concatenation operation; F' is the final attention feature generated after activation by the sigmoid function, which is then element-wise multiplied by the input feature mapping; F'' is the output feature.

7. A lightweight method for detecting small green leafhopper pests based on deep learning according to claim 6, characterized in that, The specific algorithm of WaveletPool includes: (4); (5); (6); where φ is the approximation function; ψ is the detail function; Wφ is the approximation coefficient; W ψ hφ[−n] represents the detail coefficients; hφ[−n] represents the scale of time reversal; h ψ [−n] represents the time-reversed wavelet vector; n represents the sample in the vector, j represents the resolution, and k represents the position index.

Images