A mung bean leaf spot disease recognition method based on RePMMS-Net
By constructing the RePMMS-Net network and utilizing parallel multi-branch multi-scale convolution kernel modules and global average pooling layers, the problems of low efficiency and low accuracy in mung bean leaf spot disease recognition in traditional methods were solved, and highly accurate automatic recognition was achieved.
Patent Information
- Application Number
- CN202210740190.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-28
- Publication Date
- 2025-10-10
- Estimated Expiration
- 2042-06-28
AI Technical Summary
Traditional mung bean leaf spot disease identification methods are inefficient and subjective. Traditional machine learning methods have limited ability to extract lesion features, resulting in low recognition rates and difficulty in achieving high-accuracy automated identification.
A mung bean leaf spot recognition method based on RePMMS-Net was adopted. Feature extraction was performed by constructing a parallel multi-branch multi-scale convolution kernel module PMMS. A deep learning network model was constructed by combining the global average pooling layer and the softmax classifier to perform image feature extraction and classification.
The accuracy of mung bean leaf spot disease recognition is improved, and automated, low-cost and efficient recognition is achieved with good generalization performance and robustness.
Smart Images
Figure CN114973005B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of image processing and plant leaf disease recognition, and particularly relates to a mung bean leaf spot disease recognition method based on RePMMS-Net. Background Art
[0002] Mung beans are a long-standing and widely cultivated minor grain crop in my country. With improvements in living standards and shifts in dietary patterns, mung beans are increasingly favored for their rich nutritional value and health-promoting health benefits. Furthermore, their wide adaptability, wide sowing period, short growing season, and high economic returns have made them a popular choice among farmers. However, due to unfavorable climatic conditions and poor cultivation management, mung bean pests and diseases are becoming increasingly severe. Leaf spot is a major mung bean disease. This disease primarily affects the leaves, causing them to wilt and fall off. In later stages, it can lead to premature aging of the plant. This disease can occur throughout the mung bean's growing season. Timely and effective analysis of crop lesion characteristics can help quickly determine the type and severity of the disease, provide appropriate disease prevention and control guidance, and thus reduce economic losses.
[0003] However, mung bean leaf spot disease of varying severity exhibits certain similarities. Traditional identification of mung bean leaf spot disease relies primarily on visual observation, a method that is not only inefficient but also often subjective. Some traditional machine learning methods require manual extraction of lesion features, but are limited in their ability to extract features for similar disease characteristics, and different diseases often require different features, making these methods inapplicable. Therefore, developing an automated, low-cost, and highly accurate method for accurately identifying mung bean leaf spot disease is of great practical significance. Summary of the Invention
[0004] Purpose of the invention: The invention proposes a mung bean leaf spot disease recognition method based on RePMMS-Net to solve the problem of low recognition rate due to the similarity of mung bean leaf spot disease spots of different degrees, realize the automation of mung bean leaf spot disease recognition, and improve the accuracy of identifying mung bean leaf spot disease of different degrees.
[0005] Technical solution: The present invention proposes a mung bean leaf spot disease identification method based on RePMMS-Net, which specifically includes the following steps:
[0006] (1) Acquire mung bean leaf spot disease images in advance and create a mung bean leaf spot disease dataset through manual annotation;
[0007] (2) preprocessing the mung bean leaf spot disease dataset obtained in step (1) to further expand the image sample set and increase the diversity of the samples;
[0008] (3) Constructing a parallel multi-branch multi-scale convolution kernel module PMMS: First, using multiple parallel branches, perform multi-branch multi-scale feature extraction on the image processed by step (2) to obtain the feature vector of the multi-branch multi-scale image; then, the convolution blocks corresponding to the feature vector of the obtained image are up-sampled and down-sampled respectively, and fused to obtain the feature vector of the fused image containing the convolution block information of different branches; finally, the feature vector of the image after the feature fusion is input into the convolution layer to reduce the number of channels of the output feature map;
[0009] (4) Repeatedly use the PMMS module to perform repeated feature learning to obtain the feature vector of the final image;
[0010] (5) Input the feature vector of the image obtained in step (4) into the global average pooling layer, and output the final feature vector as the input of the softmax classifier;
[0011] (6) Using the image feature vector obtained in step (5) as input, construct a softmax classifier to convert the input feature vector into a probability vector output about the category, where each probability component corresponds to a training corresponding category probability;
[0012] (7) Constructing the RePMMS-Net network model through steps (3) to (6) and training the network, that is, optimizing the network parameters through continuous iteration of the forward propagation process and the back propagation process;
[0013] (8) Use the trained RePMMS-Net network model to identify the target in the test image.
[0014] Furthermore, the step (2) includes the following steps:
[0015] (21) Flip transformation: flip the original image horizontally or vertically;
[0016] (22) Rotation: randomly rotate the original image by a certain angle;
[0017] (23) Scaling and deformation: enlarging or reducing the original image according to a certain ratio;
[0018] (24) Random cropping: randomly select a part from the original image sample and then enlarge the selected part to the original image size;
[0019] (25) Randomly select and superimpose the images processed by steps (21)-(24) and add them to the original image sample set.
[0020] Furthermore, the step (3) includes the following steps:
[0021] (31) Encoding format conversion:
[0022] Encoding the image into a tensor:
[0023] Batch RGB-Images→(batch, channel, height, width)
[0024] Where batch represents the number of image tensors input at one time; channel represents 3 for the RGB three-channel; height is the height of the input image, and width represents the width of the input image.
[0025] (32) Constructing a multi-path parallel branch module:
[0026] The multi-path parallel branch module is constructed by three branches, each of which is composed of one 3x3 convolution kernel, two 3x3 convolution kernels, and three 3x3 convolution kernels.
[0027] The feature map output by the first branch is down-sampled, and the feature map output by the third branch is up-sampled, so that the height and width of the feature maps output by the two branches become the same as the height and width of the feature map output by the second branch.
[0028] (33) Constructing a multi-branch fusion module:
[0029] The feature vectors of the three feature maps obtained from the three branches are connected and fused to obtain the feature vector of the fused image; the feature vector of the fused image is compressed in channels through a 1x1 convolution, and the added 1x1 convolution contains an activation function.
[0030] Further, the step (6) is implemented as follows:
[0031] The input feature vector is represented as: The parameter m represents the dimension size of the feature vector, and the weight matrix w is:
[0032] Where the parameter n represents the number of classes set by training;
[0033] The class y j The corresponding weight vector is:
[0034] The class vector is:
[0035] The class y j The corresponding softmax classification output probability is:
[0036] The final output of the softmax classifier is the probability vector corresponding to the category:
[0037] When applying the model, take the largest P j The corresponding category y j The final classification result.
[0038] Furthermore, the step (7) includes the following steps:
[0039] (71) Construct the RePMMS-Net network model, connect the four PMMS modules described in steps (3)-(4), then connect a global average pooling layer, and finally connect a softmax classifier;
[0040] (72) The convolutional neural network performs forward propagation as follows:
[0041]
[0042] in, represents the weight matrix between the m layer and the m-1 layer of the network, represents the offset parameter, p (m) Indicates the number of nodes in layer m, Represents the nonlinear activation function ReLU function;
[0043] (73) Based on the true label information of each image and the predicted information obtained by forward propagation, the loss function of the network is calculated;
[0044] (74) Perform the back propagation process, calculate the partial derivative of the loss function with respect to a certain node, and use the gradient descent method to update the weight parameters of the network as follows:
[0045]
[0046]
[0047] Where m represents the number of layers, i represents the number of neurons in the layer, j represents the number of a neuron input, and α represents the learning rate. After the backpropagation process finishes updating the network weight parameters, it jumps to the forward propagation.
[0048] (75) Save the trained RePMMS-Net model.
[0049] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects:
[0050] 1. Use convolutional neural networks in deep learning to extract image features instead of traditional manually defined features, which is more adaptable and has better detection effects;
[0051] 2. The use of parallel multi-branch and multi-scale convolution kernels for feature extraction and fusion improves the ability to extract lesion features similar to mung bean leaf spot disease. It is proposed to construct a multi-way parallel branch module with three branches. The three branches are composed of one 3×3 convolution kernel, two 3×3 convolution kernels, and three 3×3 convolution kernels respectively. This enriches the receptive field of the feature map and further improves the ability to extract features of different sizes. The use of such convolution kernels can minimize the number of model parameters while ensuring a certain level of accuracy.
[0052] 3. The PMMS module is reused to further extract features and improve the feature extraction capability of the network model;
[0053] 4. Using the global average pooling layer instead of the traditional fully connected layer greatly reduces the number of model parameters;
[0054] 5. Compared with the common convolutional neural network model, the network model proposed in this invention has a more reasonable network structure and a faster convergence speed. With a smaller number of parameters than other convolutional neural network models, it improves the classification accuracy and has good generalization performance and good robustness. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 is a flow chart of the present invention;
[0056] Figure 2 Schematic diagram of the image preprocessing effect adopted by the present invention;
[0057] Figure 3 Schematic diagram of the PMMS module structure proposed by the present invention;
[0058] Figure 4 This is a schematic diagram of the RePMMS-Net network structure proposed in the present invention. DETAILED DESCRIPTION
[0059] The present invention will be described in further detail below with reference to the accompanying drawings.
[0060] like Figure 1 As shown, the present invention proposes a mung bean leaf spot disease identification method based on RePMMS-Net, comprising the following steps:
[0061] Step 1: Obtain captured images of mung bean leaf spot and create a mung bean leaf spot dataset through manual annotation. Preprocess the images to further expand the image sample set and increase sample diversity.
[0062] Flip the original image horizontally or vertically, randomly rotate the original image by a certain angle; enlarge or reduce the original image according to a certain ratio; randomly select a part from the original image sample, and then enlarge the selected part to the original image size; the image preprocessing effect is as follows Figure 2 As shown, the processed images are randomly selected and superimposed and added to the original image sample set to increase the diversity of the samples.
[0063] Step 2: Build the PMMS module (Parallel multi-branch multi-scale convolution kernel module), such as Figure 3 As shown in the figure: First, multi-branch and multi-scale feature extraction is performed on the processed image using multiple parallel branches to obtain the feature vector of the multi-branch and multi-scale image; then, the convolution blocks corresponding to the feature vector of the obtained image are upsampled and downsampled respectively, and fused to obtain the feature vector of the fused image containing the convolution block information of different branches; finally, the feature vector of the image after feature fusion is input into the convolution layer to reduce the number of channels of the output feature map; the PMMS module is reused 3 times for repeated feature learning to obtain the feature vector of the final image.
[0064] Encoding format conversion:
[0065] Encode the image into a tensor as follows:
[0066] Batch RGB-Images→(batch, channel, height, width)
[0067] Among them, batch represents the number of images input at a time; channel represents 3 RGB channels; height is the height of the input image, and width represents the width of the input image.
[0068] Construct a multi-way parallel branch module: Use three branches to construct a multi-way parallel branch module. The three branches are composed of one 3×3 convolution kernel, two 3×3 convolution kernels, and three 3×3 convolution kernels respectively, to further improve the ability to extract features of different sizes. The use of such convolution kernels can minimize the number of model parameters while ensuring a certain accuracy; downsample the feature map output by the first branch, and upsample the feature map output by the third branch, so that the height and width of the feature maps output by the two branches become the same size as the height and width of the feature map output by the second branch.
[0069] The processed green bean leaf image with a size of 224x224x3 is subjected to a convolution operation with a convolution kernel number of 24, a size of 3x3, a step of 1, and padding=1, to obtain a feature vector (224, 224, 24); then a maximum pooling operation with a convolution kernel number of 2 and a step of 2 is performed on it, and a feature vector (112, 112, 24) is output as the first branch of the multi-parallel branch module.
[0070] The same green bean leaf image is subjected to a convolution operation with a convolution kernel number of 24, a size of 3x3, a step of 2, and padding=1, and then a ReLU activation function, to obtain a feature vector (112, 112, 24); then a convolution operation with a convolution kernel number of 24, a size of 3x3, a step of 1, and padding=1 is performed on it, and then a ReLU activation function, to obtain a feature vector (112, 112, 24) as the second branch of the multi-parallel branch module.
[0071] The same green bean leaf image is subjected to a convolution operation with a convolution kernel number of 48, a size of 3x3, a step of 2, and padding=1, and then a ReLU activation function, to obtain a feature vector (112, 112, 48); then a convolution operation with a convolution kernel number of 48, a size of 3x3, a step of 2, and padding=1 is performed on it, and then a ReLU activation function, to obtain a feature vector (56, 56, 48); then a convolution operation with a convolution kernel number of 48, a size of 3x3, a step of 1, and padding=1 is performed on it, to obtain a feature vector (56, 56, 48); then an up-sampling operation is performed on it, to output a feature vector (112, 112, 48) as the third branch of the multi-parallel branch module.
[0072] A multi-branch fusion module is constructed: the feature vectors of the three feature maps obtained from the three branches are connected and fused to obtain a feature vector of the feature-fused image; the feature vector of the feature-fused image is compressed through a 1x1 convolution, and the added 1x1 convolution contains an activation function.
[0073] The feature vectors of the three branches obtained are connected and fused to obtain a feature vector (112, 112, 96) of the feature-fused image; the obtained feature vector (112, 112, 96) is compressed through a convolution kernel with a convolution kernel number of 3 and a size of 1x1, to obtain a feature vector (112, 112, 3) as the input of the next PMMS module.
[0074] First repetition of the PMMS module: The feature vector obtained by the previous PMMS module is convolved with 24 kernels, a size of 3×3, a stride of 1, and padding=1 to obtain the feature vector (112, 112, 24). This is then subjected to a max pooling operation with 2 kernels and a stride of 2, outputting the feature vector (56, 56, 24), which serves as the first branch of the multi-way parallel branch module. The feature vector obtained by the previous PMMS module is convolved with 24 kernels, a size of 3×3, a stride of 2, and padding=1, and then activated by the ReLU function to obtain the feature vector (56, 56, 24). This is then convolved with 24 kernels, a size of 3×3, a stride of 1, and padding=1, and then activated by the ReLU function to obtain the feature vector (56, 56, 24), which serves as the second branch of the multi-way parallel branch module. The feature vector obtained by the previous PMMS module is convolved with 48 kernels, a size of 3×3, a stride of 2, and padding = 1, and then activated by the ReLU function to obtain the feature vector (56, 56, 48). The feature vector is then convolved with 48 kernels, a size of 3×3, a stride of 2, and padding = 1, and then activated by the ReLU function to obtain the feature vector (28, 28, 48). The feature vector is then convolved with 48 kernels, a size of 3×3, a stride of 1, and padding = 1 to obtain the feature vector (28, 28, 48). The feature vector is then upsampled and output as the feature vector (56, 56, 48), which serves as the third branch of the multi-way parallel branch module. The feature vectors of the three branches are concatenated and fused to obtain the feature vector (56, 56, 96) of the fused image. The feature vector (56, 56, 96) is channel-compressed by a convolution kernel with a kernel size of 1×1 and a kernel number of 3 to obtain a feature vector (56, 56, 3), which is used as the input of the next PMMS module.
[0075] The feature vector obtained by the previous PMMS module is subjected to a convolution operation with a convolution kernel number of 24, a size of 3x3, a step of 1, and padding=1, to obtain a feature vector (56, 56, 24); then a maximum pooling operation with a convolution kernel number of 2 and a step of 2 is performed on it, to output a feature vector (28, 28, 24) as the first branch of the multi-parallel branch module. The feature vector obtained by the previous PMMS module is subjected to a convolution operation with a convolution kernel number of 24, a size of 3x3, a step of 2, and padding=1, and then a ReLU activation function is performed, to obtain a feature vector (28, 28, 24); then a convolution operation with a convolution kernel number of 24, a size of 3x3, a step of 1, and padding=1 is performed on it, and then a ReLU activation function is performed, to obtain a feature vector (28, 28, 24) as the second branch of the multi-parallel branch module. The feature vector obtained by the previous PMMS module is subjected to a convolution operation with a convolution kernel number of 48, a size of 3x3, a step of 2, and padding=1, and then a ReLU activation function is performed, to obtain a feature vector (28, 28, 48); then a convolution operation with a convolution kernel number of 48, a size of 3x3, a step of 2, and padding=1 is performed on it, and then a ReLU activation function is performed, to obtain a feature vector (14, 14, 48); then a convolution operation with a convolution kernel number of 48, a size of 3x3, a step of 1, and padding=1 is performed on it, to obtain a feature vector (14, 14, 48); then an up-sampling operation is performed on it, to output a feature vector (28, 28, 48) as the third branch of the multi-parallel branch module. The feature vectors of the three branches obtained are subjected to a connection fusion operation, to obtain a feature vector (28, 28, 96) of the image after feature fusion. The feature vector (28, 28, 96) is subjected to a channel compression operation by a convolution kernel with a convolution kernel number of 3 and a size of 1x1, to obtain a feature vector (28, 28, 3) as the input of the next PMMS module.
[0076] The PMMS module is repeated for the third time: the feature vector obtained from the previous PMMS module is convolved with 24 kernels, a size of 3×3, a stride of 1, and padding=1, resulting in a feature vector of (28, 28, 24). This is then subjected to a max pooling operation with 2 kernels and a stride of 2, outputting a feature vector of (14, 14, 24), which serves as the first branch of the multi-way parallel branch module. The feature vector obtained from the previous PMMS module is convolved with 24 kernels, a size of 3×3, a stride of 2, and padding=1, followed by a ReLU activation function, resulting in a feature vector of (14, 14, 24). This is then convolved with 24 kernels, a size of 3×3, a stride of 1, and padding=1, followed by a ReLU activation function, resulting in a feature vector of (14, 14, 24), which serves as the second branch of the multi-way parallel branch module. The feature vector obtained by the previous PMMS module is subjected to a convolution operation with 48 convolution kernels, a size of 3×3, a stride of 2, and padding=1, and then subjected to a ReLU activation function to obtain a feature vector (14, 14, 48); then a convolution operation with 48 convolution kernels, a size of 3×3, a stride of 2, and padding=1 is performed on it, and then a ReLU activation function is performed to obtain a feature vector (7, 7, 48); then a convolution operation with 48 convolution kernels, a size of 3×3, a stride of 1, and padding=1 is performed on it to obtain a feature vector (7, 7, 48); then an upsampling operation is performed on it to output a feature vector (14, 4, 48) as the third branch of the multi-way parallel branch module; the feature vectors of the three branches are connected and fused to obtain the feature vector (14, 14, 96) of the image after feature fusion. The obtained feature vector (14,14,96) is channel compressed by a convolution kernel with a kernel size of 1×1 and a kernel number of 3 to obtain a feature vector (14,4,3), which is used as the input of the global average pooling layer.
[0077] Step 3: Input the feature vector of the image into the global average pooling layer, which is used to replace the commonly used fully connected structure, which can greatly reduce the parameters of the model and output the final feature vector as the input of the softmax classifier.
[0078] The obtained feature vector is input into the global average pooling layer so that the size of each channel after pooling is 1x1, that is, there is only one pixel on each channel, which is used as the input of the softmax classifier.
[0079] Step 4: Use the obtained image feature vector as input to build a softmax classifier, and convert the input feature vector into a probability vector output about the category, where each probability component corresponds to a training corresponding category probability.
[0080] The input feature vector is used express: The parameter m represents the dimension of the feature vector, and the weight matrix w is:
[0081]
[0082] The parameter n represents the number of categories in the training set; the category y j The corresponding weight vector is: Category vector for: Category y j The corresponding softmax classification output probability is:
[0083]
[0084] The final output of the softmax classifier is the probability vector corresponding to the category:
[0085]
[0086] When applying the model, take the largest P j The corresponding category y j The final classification result.
[0087] Step 5: Construct the RePMMS-Net (Repeat parallel multi-branch multi-scale convolution kernel module network) network model through steps (2)-(4), as shown in Figure 4 As shown, the network is trained by continuously iterating the forward propagation process and the back propagation process to optimize the parameters of the network.
[0088] Forward propagation is the process of calculating the output of each neural network iteration of the input image. A deep neural network consists of an input layer, multiple intermediate layers, and an output layer. Each layer can be represented as a nonlinear mapping process. The formula for calculating the input of sample x at layer m is as follows:
[0089]
[0090] in, represents the weight matrix between the m layer and the m-1 layer of the network, represents the offset parameter, p (m) Indicates the number of nodes in layer m, Represents the nonlinear activation function ReLU function.
[0091] After the forward propagation is completed, the network's loss function is calculated and the obtained loss function value is compared with the threshold. If the loss function value is less than the threshold or the number of iterations at this time is greater than the maximum number of iterations, the training ends. Otherwise, the backpropagation is performed. The backpropagation process calculates the partial derivative of the loss function with respect to a certain node and uses the gradient descent algorithm to update the network weight parameters. The updated parameters of the network weight are as follows:
[0092]
[0093]
[0094] Where m is the number of layers, i is the number of neurons in the layer, j is the number of a neuron's input, and α is the learning rate. After the backpropagation process completes updating the network weight parameters, it jumps to the forward propagation process.
[0095] Step 6: Use the trained RePMMS-Net model to detect objects in the test image.
[0096] The image to be detected is resized to 224×224, and the image is input into the trained RePMMS-Net model for convolution calculation, and then the recognition accuracy is output.
[0097] Table 1 is the experimental results of the present invention compared with AlexNet
[0098] Model name Accuracy Model size RePMMS-Net 95.50％ 52.42MB AlexNet 85.74％ 76.18MB
[0099] As shown in Table 1, based on the above method of the embodiment of the present invention, the accuracy rate of identifying mung bean leaf spot disease is about 95.50%.
Claims
1. A mung bean leaf spot disease identification method based on RePMMS-Net, characterized in that: The following steps are involved: (1) Acquire mung bean leaf spot disease images in advance and create a mung bean leaf spot disease dataset through manual annotation; (2) preprocessing the mung bean leaf spot disease dataset obtained in step (1) to further expand the image sample set and increase the diversity of the samples; (3) Constructing a parallel multi-branch multi-scale convolution kernel module PMMS, including a multi-way parallel branch module and a multi-branch fusion module; first, using the multi-way parallel branch, perform multi-branch multi-scale feature extraction on the image processed by step (2) to obtain a feature vector of the multi-branch multi-scale image; then, up-sampling and down-sampling processing are performed on the convolution blocks corresponding to the feature vector of the obtained image, and fusion is performed to obtain a feature vector of the fused image containing the convolution block information of different branches; finally, the feature vector of the image after the feature fusion is input into the convolution layer to reduce the number of channels of the output feature map; (4) Repeatedly use the PMMS module to perform repeated feature learning to obtain the feature vector of the final image; (5) Input the feature vector of the image obtained in step (4) into the global average pooling layer, and output the final feature vector as the input of the softmax classifier; (6) Using the image feature vector obtained in step (5) as input, construct a softmax classifier to convert the input feature vector into a probability vector output about the category, where each probability component corresponds to a training corresponding category probability; (7) Constructing the RePMMS-Net network model through steps (3) to (6) and training the network, that is, optimizing the network parameters through continuous iteration of the forward propagation process and the back propagation process; (8) Use the trained RePMMS-Net network model to identify the target in the test image; The multi-path parallel branch module is constructed using three branches; the three branches are respectively composed of one 3×3 convolution kernel, two 3×3 convolution kernels, and three 3×3 convolution kernels; the feature map output by the first branch is downsampled, and the feature map output by the third branch is upsampled, so that the height and width of the feature maps output by the two branches become the same size as the height and width of the feature map output by the second branch; The processed mung bean leaf image of size 224×224×3 is convolved with 24 convolution kernels, a size of 3×3, a stride of 1, and padding = 1, resulting in a feature vector of (224, 224, 24). A maximum pooling operation is then performed on the image with 2 convolution kernels and a stride of 2, outputting a feature vector of (112, 112, 24) as the first branch of the multi-way parallel branch module. The same mung bean leaf image is convolved with 24 convolution kernels, 3×3 size, stride 2, and padding 1, and then activated by the ReLU function to obtain the feature vector (112, 112, 24). The image is then convolved with 24 convolution kernels, 3×3 size, stride 1, and padding 1, and then activated by the ReLU function to obtain the feature vector (112, 112, 24), which is used as the second branch of the multi-way parallel branch module. The same mung bean leaf image is subjected to a convolution operation with 48 convolution kernels, a size of 3×3, a stride of 2, and padding=1, and then subjected to a ReLU activation function to obtain a feature vector (112, 112, 48); then a convolution operation with 48 convolution kernels, a size of 3×3, a stride of 2, and padding=1 is performed on it, and then a ReLU activation function is performed to obtain a feature vector (56, 56, 48); then a convolution operation with 48 convolution kernels, a size of 3×3, a stride of 1, and padding=1 is performed on it, and a feature vector (56, 56, 48) is obtained; finally, an upsampling operation is performed on it, and the feature vector (112, 112, 48) is output as the third branch of the multi-way parallel branch module; The step (4) of reusing the PMMS module is to repeat the PMMS module three times, specifically comprising: Repeat the PMMS module for the first time: perform a convolution operation on the feature vector obtained by the previous PMMS module with a convolution kernel number of 24, a size of 3×3, a step size of 1, and padding=1 to obtain a feature vector (112,112,24); then perform a maximum pooling operation on it with a convolution kernel number of 2 and a step size of 2, and output a feature vector (56,56,24) as the first branch of the multi-way parallel branch module; perform a convolution operation on the feature vector obtained by the previous PMMS module with a convolution kernel number of 24, a size of 3×3, a step size of 2, and padding=1, and then pass it through the ReLU activation function to obtain a feature vector (56,56,24); then perform a convolution operation on it with a convolution kernel number of 24, a size of 3×3, a step size of 1, and padding=1, and then pass it through the ReLU activation function to obtain a feature vector (56,56,24) as the first branch of the multi-way parallel branch module The second branch; the feature vector obtained by the previous PMMS module is subjected to a convolution operation with a convolution kernel number of 48, a size of 3×3, a step size of 2, and padding=1, and then subjected to a ReLU activation function to obtain a feature vector (56,56,48); then the feature vector is subjected to a convolution kernel number of 48, a size of 3×3, a step size of 2, and padding=1, and then subjected to a ReLU activation function to obtain a feature vector (28,28,48); then the feature vector is subjected to a convolution kernel number of 48, a size of 3×3, a step size of 1, and padding=1 to obtain a feature vector (28,28,48); then the feature vector is subjected to an upsampling operation to output a feature vector (56,56,48) as the third branch of the multi-way parallel branch module; the feature vectors of the three branches are connected and fused to obtain the feature vector (56,56,96) of the image after feature fusion; The feature vector (56, 56, 96) is channel-compressed by a convolution kernel with a kernel size of 1×1 and a kernel number of 3, resulting in a feature vector (56, 56, 3) as the input of the next PMMS module. Repeat the PMMS module for the second time; perform a convolution operation on the feature vector obtained by the previous PMMS module with a convolution kernel number of 24, a size of 3×3, a step size of 1, and padding=1 to obtain a feature vector (56,56,24); then perform a maximum pooling operation on it with a convolution kernel number of 2 and a step size of 2, and output a feature vector (28,28,24) as the first branch of the multi-way parallel branch module; perform a convolution operation on the feature vector obtained by the previous PMMS module with a convolution kernel number of 24, a size of 3×3, a step size of 2, and padding=1, and then pass it through the ReLU activation function to obtain a feature vector (28,28,24); then perform a convolution operation on it with a convolution kernel number of 24, a size of 3×3, a step size of 1, and padding=1, and then pass it through the ReLU activation function to obtain a feature vector (28,28,24) as the second branch of the multi-way parallel branch module; perform a convolution operation on the feature vector obtained by the previous PMMS module with a convolution kernel number of 48, a size of 3× 3. A convolution operation with a step size of 2 and padding = 1 is performed, and then a ReLU activation function is performed to obtain the feature vector (28, 28, 48); then a convolution operation with a kernel size of 48, a size of 3×3, a step size of 2, and padding = 1 is performed, and then a ReLU activation function is performed to obtain the feature vector (14, 14, 48); then a convolution operation with a kernel size of 48, a size of 3×3, a step size of 1, and padding = 1 is performed to obtain the feature vector (14,14,48); then perform an upsampling operation on it and output a feature vector (28,28,48), which is used as the third branch of the multi-way parallel branch module; connect and fuse the feature vectors of the three branches to obtain the feature vector (28,28,96) of the image after feature fusion; perform channel compression on the feature vector (28,28,96) through a convolution kernel with a convolution kernel number of 3 and a size of 1×1 to obtain a feature vector (28,28,3), which is used as the input of the next PMMS module; Repeat the PMMS module for the third time: perform a convolution operation on the feature vector obtained by the previous PMMS module with a convolution kernel number of 24, a size of 3×3, a step size of 1, and padding=1 to obtain a feature vector (28,28,24); then perform a maximum pooling operation on it with a convolution kernel number of 2 and a step size of 2, and output a feature vector (14,14,24) as the first branch of the multi-way parallel branch module; perform a convolution operation on the feature vector obtained by the previous PMMS module with a convolution kernel number of 24, a size of 3×3, a step size of 2, and padding=1, and then pass it through the ReLU activation function to obtain a feature vector (14,14,24); then perform a convolution operation on it with a convolution kernel number of 24, a size of 3×3, a step size of 1, and padding=1, and then pass it through the ReLU activation function to obtain a feature vector (14,14,24) as the second branch of the multi-way parallel branch module; perform a convolution operation on the feature vector obtained by the previous PMMS module with a convolution kernel number of 48, a size of The convolution operation with a kernel size of 3×3, a stride of 2, and padding=1 is performed, and then the ReLU activation function is performed to obtain the feature vector (14,14,48); the convolution operation with a kernel size of 48, a size of 3×3, a stride of 2, and padding=1 is performed, and then the ReLU activation function is performed to obtain the feature vector (7,7,48); the convolution operation with a kernel size of 48, a size of 3×3, a stride of 1, and padding=1 is performed to obtain the feature vector The image is taken as the vector (7,7,48); an upsampling operation is then performed on it to output a feature vector (14,4,48), which is used as the third branch of the multi-way parallel branch module; the feature vectors of the three branches are concatenated and fused to obtain a feature vector (14,14,96) of the image after feature fusion; the obtained feature vector (14,14,96) is channel-compressed by a convolution kernel with a convolution kernel size of 1×1 and a number of 3 convolution kernels to obtain a feature vector (14,4,3), which is used as the input of the global average pooling layer.
2. The mung bean leaf spot disease identification method based on RePMMS-Net according to claim 1, characterized in that: The step (2) comprises the following steps: (21) Flip transformation: flip the original image horizontally or vertically; (22) Rotation: randomly rotate the original image by a certain angle; (23) Scaling and deformation: enlarging or reducing the original image according to a certain ratio; (24) Random cropping: randomly select a part from the original image sample and then enlarge the selected part to the original image size; (25) Randomly select and superimpose the images processed by steps (21)-(24) and add them to the original image sample set.
3. The mung bean leaf spot disease identification method based on RePMMS-Net according to claim 1, characterized in that: The implementation process of step (6) is as follows: The input feature vector is used express: The parameter m represents the dimension of the feature vector, and the weight matrix w is: Where the parameter n represents the number of categories in the training setting; Category y j The corresponding weight vector is: Category vector for: Category y j The corresponding softmax classification output probability is: The final output of the softmax classifier is the probability vector corresponding to the category: When applying the model, take the largest P j The corresponding category y j The final classification result.
4. The mung bean leaf spot disease identification method based on RePMMS-Net according to claim 1, characterized in that: The step (7) comprises the following steps: (71) Construct the RePMMS-Net network model, connect the four PMMS modules described in steps (3)-(4), then connect a global average pooling layer, and finally connect a softmax classifier; (72) The convolutional neural network performs forward propagation as follows: in, represents the weight matrix between the m layer and the m-1 layer of the network, represents the offset parameter, p (m) Indicates the number of nodes in layer m, Represents the nonlinear activation function ReLU function; (73) Based on the true label information of each image and the predicted information obtained by forward propagation, the loss function of the network is calculated; (74) Perform the back propagation process, calculate the partial derivative of the loss function with respect to a certain node, and use the gradient descent method to update the weight parameters of the network as follows: Where m represents the number of layers, i represents the number of neurons in the layer, j represents the number of a neuron input, and α represents the learning rate. After the backpropagation process finishes updating the network weight parameters, it jumps to the forward propagation. (75) Save the trained RePMMS-Net model.