Mung bean leaf spot disease recognition method and device based on MS-PLNet model
By adopting a multi-scale feature extraction and feature fusion method based on the MS-PLNet model in the identification of mung bean leaf spot disease, the problems of large parameters and high computational volume of traditional convolutional neural network models are solved, and efficient and accurate leaf spot disease recognition is achieved.
Patent Information
- Application Number
- CN202210598574.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-30
- Publication Date
- 2025-03-28
- Estimated Expiration
- 2042-05-30
AI Technical Summary
Traditional convolutional neural network models have problems with large amounts of parameters and high calculations in the identification of mung bean leaf spot disease, resulting in large resource consumption and low recognition efficiency.
A method for identifying mung bean leaf spot disease based on MS-PLNet model is proposed. Through a multi-scale feature extraction module, feature fusion module and classifier, combined with ordinary convolution and improved depth separation convolution, the channel attention mechanism is adopted to reduce network parameters and improve feature extraction capabilities.
The high-accuracy mung bean leaf spot disease recognition was achieved, with the recognition rate reaching 99.96%, and the model size was only 7.59M, which has good robustness and wide application value.
Smart Images
Figure CN114972264B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of agricultural plant protection, and in particular to a mung bean leaf spot disease recognition method based on an MS-PLNet model. Background Art
[0002] Crops are various plants cultivated by humans in agricultural production activities. They are divided into two categories: food crops and cash crops. No matter which type of crops, they are easily affected by various aspects such as water, soil, diseases and pests during the planting process. This requires the use of effective cultivation techniques to reduce the impact of these factors on crop planting.
[0003] Identifying diseases is a critical step in the agricultural sector because it affects the planting and harvesting of crops. Identification of these diseases requires a lot of time and resources, and often requires expert knowledge. Farmers will have to manually check the symptoms of large areas of crops. This process is repetitive and inefficient. Mung bean leaf spot is the main disease of mung bean. The harmful part is the leaves, which are severely affected during the flowering and pod-setting period. In the early stage of the disease, small brown spots with water stains appear on the leaves, and after expansion, they form nearly circular lesions with reddish brown to reddish brown edges and light gray to light brown in the middle. When the disease is serious, the lesions merge into pieces and dry up quickly. In mild cases, the yield is reduced by 20% to 50%, and in severe cases, it is as high as 90%. Traditional identification of mung bean leaf spot has many shortcomings in actual agricultural production. Naked eye observation cannot accurately and objectively evaluate mung bean leaf spot, and it is difficult to identify on a large scale. Although optical microscopy technology, bioassay technology, PCR technology, etc. can perform accurate identification, they require professional operation, are time-consuming and labor-intensive, and these technologies are difficult to apply to real-time online identification of mung bean leaf spot. Using computer vision techniques, the identification and grading of diseases has improved greatly, and today, smart agriculture is focused on providing the infrastructure to leverage advanced technologies such as IoT and big data. With the recent advent of deep learning, it is now possible to automatically identify these diseases with great accuracy.
[0004] At present, convolutional neural networks have been widely used in agricultural engineering related fields, and have also played an important role in the identification of mung bean leaf diseases. In order to obtain better performance, the number of network layers has been increasing in recent years, from 7-layer AlexNet to 16-layer VGGNet, to 22-layer GoogleNet, 152-layer ResNet, and even thousands of layers of ResNet and DenseNet. However, these traditional convolutional neural network recognition systems have problems of large parameters and high computational complexity. The purpose of studying lightweight convolutional neural networks is to significantly reduce the amount of computation and parameters on the basis of ensuring that the accuracy of large convolutional neural networks is comparable, so as to reduce training and deployment resources. Summary of the invention
[0005] Purpose of the invention: The present invention overcomes the shortcomings of traditional convolutional neural network models, such as large number of parameters and high amount of model calculation, and proposes a method and device for identifying mung bean leaf spot disease based on the MS-PLNet model, which can not only have a good leaf spot disease detection accuracy, but also enable the model to have a smaller number of parameters, thereby realizing rapid identification and statistical analysis of leaf spot disease.
[0006] Technical solution: The present invention proposes a method for identifying mung bean leaf spot disease based on the MS-PLNet model, comprising the following steps:
[0007] (1) Preprocessing plant leaf disease images containing multiple disease types, and dividing them into a training set, a validation set, and a test set;
[0008] (2) Construct the MS-PLNet leaf spot disease recognition network;
[0009] (3) Training the MS-PLNet leaf spot disease recognition network, passing the preprocessed training set through the built MS-PLNet leaf spot disease recognition network, comparing the obtained prediction results with the true labels to calculate the classification loss, and updating the model parameters through the momentum gradient descent algorithm to obtain the MS-PLNet model for leaf spot disease recognition;
[0010] (4) Identification of mung bean leaf spot disease: The images in the validation set are input into the MS-PLNet model trained in step (3) to evaluate the performance of the model, and the images in the test set are input into the trained model to obtain the leaf disease type of the image through forward propagation.
[0011] Furthermore, the image in step (1) is an image of a mung bean leaf taken with an ordinary camera, and data enhancement is performed on all leaf images, including random illumination transformation, left-right flipping, up-down flipping, diagonal flipping, and random cropping.
[0012] Furthermore, the MS-PLNet leaf spot disease recognition network includes a feature extraction module, a feature fusion module and a classifier; inputting images of three scales into the feature extraction module; constructing feature extraction modules of three scales, wherein the feature extraction modules are constructed using ordinary convolution and improved depthwise separable convolution, during which the first feature extraction module increases the number of channels, the second feature extraction module reduces the number of channels, and the third feature extraction module increases the number of channels; inputting feature maps obtained by the feature extraction modules of the three scales into a feature fusion module using a channel attention mechanism; and inputting the feature maps output by feature fusion into the classifier to obtain classification results.
[0013] Furthermore, the step (2) comprises the following steps:
[0014] (21) In the image input stage, the input 3-channel image is resized to 128×128 pixels, marked as IM0; the first feature extraction module inputs the 256×256 pixel image after the UnSample deconvolution operation, marked as IM1; the second feature extraction module inputs the 128×128 pixel image, marked as IM2; the third feature extraction module inputs the 64×64 pixel image after the MaxPool2d maximum pooling operation, marked as IM3;
[0015] (22) In the feature extraction module of the first scale, IM1 is used as the input of the multi-convolution network to perform convolution with a kernel of 4×4 to obtain C11, C11 is subjected to point convolution with a kernel of 1×1 to increase the number of channels and a depth convolution of 7×7 to obtain C12, C12 is subjected to the same convolution calculation to obtain C13, C13 is subjected to the same convolution calculation again to obtain C14, C14 is subjected to point convolution with a kernel of 1×1 to reduce the number of channels and a depth convolution of 7×7 to obtain C15, C15 is convolved with a 1×1 point convolution kernel to increase the number of channels and a 7×7 depth convolution calculation to obtain C16, and C16 is convolved with the same convolution calculation to obtain C17; in the feature extraction module of the second scale, IM2 is used as the input of the multi-convolution network to perform a 4×4 convolution calculation to obtain C21, and C21 is convolved with a 1×1 point convolution kernel to increase the number of channels and a 5×5 depth convolution calculation to obtain C22, and C22 is convolved with the same convolution calculation to obtain C23, and C23 is further The same convolution calculation is performed to obtain C24, and C24 is subjected to a point convolution with a convolution kernel of 1×1 to reduce the number of channels and a depth convolution of 7×7 to obtain C25. C25 is subjected to a point convolution with a convolution kernel of 1×1 to increase the number of channels and a depth convolution of 7×7 to obtain C26. C26 is subjected to the same convolution calculation to obtain C27. In the feature extraction module of the third scale, IM3 is used as the input of the multi-convolution network to perform a convolution kernel of 4×4 to obtain C31. C31 is subjected to a convolution kernel of Perform a 1×1 point convolution to increase the number of channels and a 3×3 depth convolution to obtain C32, perform the same convolution calculation on C32 to obtain C33, perform the same convolution calculation on C33 to obtain C34, perform a 1×1 point convolution on C34 to reduce the number of channels and a 3×3 depth convolution to obtain C35, perform a 1×1 point convolution on C35 to increase the number of channels and a 3×3 depth convolution to obtain C36, and perform the same convolution calculation on C36 to obtain C37;
[0016] (23) The feature maps C17, C27 and C37 obtained by the feature extraction modules of the three scales are concatenated to obtain R1, and the weights of each feature channel of R1 are readjusted through the channel attention mechanism, and the high-dimensional features are reasonably reorganized to obtain R2;
[0017] (24) R2 is input into the average pooling layer, the fully connected layer and the Softmax activation function, and the classification result is output.
[0018] Based on the same inventive concept, the present invention also proposes a mung bean leaf spot disease identification device based on the MS-PLNet model, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the mung bean leaf spot disease identification method based on the MS-PLNet model is implemented.
[0019] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention constructs an MS-PLNet model for mung bean leaf spot disease recognition. In the data input stage, multi-scale images are input to increase the diversity of model extraction features; in the feature extraction stage, ordinary convolution and improved depthwise separable convolution are used for construction, and a three-stage strategy of first increasing the number of channels, then decreasing them, and finally increasing them is adopted to reduce network parameters; finally, in the feature fusion stage, the channel attention mechanism is continued to be used to focus on important channels and suppress redundant channels according to the importance of the channels; the recognition rate of the present invention is 99.96%, and the model size is 7.59M. It can be used for the detection of mung bean leaf spot disease and leaf spot disease of other crops, and has good robustness and wide application value. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 is a flow chart of the present invention;
[0021] Figure 2 Schematic diagram of the MS-PLNet model structure;
[0022] Figure 3 This is a schematic diagram of the feature extraction network structure of scale 1 in the MS-PLNet model;
[0023] Figure 4 This is a schematic diagram of the feature extraction network structure of scale 2 in the MS-PLNet model;
[0024] Figure 5 This is a schematic diagram of the feature extraction network structure of scale three in the MS-PLNet model;
[0025] Figure 6 The figure is a comparison chart of the experimental results of the present invention and the performance of the lightweight model. DETAILED DESCRIPTION
[0026] The present invention will be further described in detail below in conjunction with the accompanying drawings.
[0027] This embodiment involves a large number of variables, and each variable is described as follows, as shown in Table 1.
[0028] Table 1 Variable description table
[0029]
[0030]
[0031] The present invention proposes a mung bean leaf spot disease recognition method based on the MS-PLNet model. Figure 1 As shown, the specific steps include:
[0032] Step 1: Preprocess the plant leaf disease images containing various disease types acquired in advance and divide them into training set, validation set and test set.
[0033] The mung bean leaves were photographed with a Canon EOS 550D camera, and data enhancement was performed on all leaf images, including random illumination transformation, left-right flipping, up-down flipping, diagonal flipping, and random cropping.
[0034] The data sets used in the present invention are the PlantVillage public data set and the mung bean leaf data set taken by the Jiangsu Academy of Agricultural Sciences in natural scenes for experiments. The PlantVillage public data set includes 20,636 images of diseased and healthy leaves of 15 different categories; the data set provided by the Jiangsu Academy of Agricultural Sciences includes 1,006 images of diseased and healthy mung bean leaves. In order to make up for the impact of the uneven distribution of PlantVillage data samples on the model recognition performance and avoid network overfitting, the present invention performs data enhancement operations before training. The data enhancement methods used are: 1) Random illumination transformation: PCA (principal component analysis) is performed on the pixels of all images to obtain the eigenvalues and eigenvectors in the RGB space, and some random additions and subtractions are made to the pixel values of the picture along the direction of the eigenvector to simulate images with different illumination. 2) Random flipping: including left-right flipping, up-down flipping and diagonal flipping. 3) Random cropping: cropping is performed at random positions of the picture according to the specified size. This data enhancement method can move the positions of various areas on the picture on the basis of retaining the image ratio. Finally, the size of all images was unified to 128 × 128 pixels. After preprocessing, a total of 31,117 image samples were obtained, which served as the mung bean leaf spot disease dataset for the final experiment.
[0035] Step 2: Construct the MS-PLNet leaf spot disease recognition network, such as Figure 2As shown, it includes a feature extraction module, a feature fusion module and a classifier; input images of three scales into the feature extraction module; construct feature extraction modules of three scales, and the feature extraction modules are constructed using ordinary convolution and improved depthwise separable convolution. During the construction process, the first feature extraction module increases the number of channels, the second feature extraction module reduces the number of channels, and the third feature extraction module increases the number of channels; the feature maps obtained by the feature extraction modules of the three scales are input into the feature fusion module using the channel attention mechanism. Based on the obtained feature fusion, the disease image is classified and identified by a classifier composed of average pooling and linear layers. This method can extract image features from three different angles in a multi-scale feature extraction manner without increasing model parameters, and is suitable for the classification of leaf spot diseases of crops. Specifically, it includes the following steps:
[0036] (1) The multi-scale feature extraction network is composed of ordinary 4×4 convolution plus improved depth-wise separable convolution, and adopts a three-stage strategy, that is, the number of channels is first increased, then decreased, and then increased again to reduce the number of channels accordingly without losing the feature map of the effective channel. The process of the multi-scale feature extraction network is as follows:
[0037] 1) Feature extraction network of scale 1 such as Figure 3As shown in the figure: an 18-layer multi-combination network is constructed, with a total of 16 convolutional layers and 2 deconvolutional layers; the input image is a 3-channel color image with a resolution of 128×128; a convolution with a convolution kernel of 1×1 is used to maintain the number of channels at 3; a deconvolution operation with a convolution kernel of 2×2 is used to increase the image resolution to 256×256; a convolution with a convolution kernel of 4×4 is used to reduce the feature map resolution to 64×64, and the number of channels is 96; a convolution with a convolution kernel of 1×1 is used, and the number of channels is increased to 192; a convolution with a convolution kernel of 7×7 and groups of 192 is used to reduce the image resolution to 32×32; a convolution with a convolution kernel of 1×1 is used, and the number of channels is increased to 384; a convolution with a convolution kernel of 7×7 and groups of 384 is used to reduce the image resolution to 16×16; a convolution with a convolution kernel of 1×1 is used, and the channel The number of channels increases to 768; the image resolution is reduced to 8×8 by using a convolution kernel of 7×7 and groups of 768; the number of channels is reduced to 384 by using a convolution kernel of 1×1; the image resolution is reduced to 4×4 by using a convolution kernel of 7×7 and groups of 384; the number of channels increases to 768 by using a convolution kernel of 1×1; the image resolution is reduced to 2×2 by using a convolution kernel of 7×7 and groups of 768; the number of channels remains at 768 by using a convolution kernel of 1×1; the image resolution is increased to 4×4 by using a deconvolution operation with a convolution kernel of 2×2; the image resolution is reduced to 2×2 by using a convolution kernel of 7×7 and groups of 768; the image resolution is reduced to 1×1 by using a convolution kernel of 7×7 and groups of 768.
[0038] 2) Feature extraction network of scale 2 is as follows Figure 4As shown in the figure: a 17-layer multi-combination network is constructed, with a total of 15 convolutional layers and 2 deconvolutional layers; the input image is a 3-channel color image with a resolution of 128×128; the convolution kernel of 4×4 is used to reduce the feature map resolution to 32×32, and the number of channels is 96; the convolution kernel of 1×1 is used to increase the number of channels to 192; the convolution kernel of 5×5 and groups of 192 is used to reduce the image resolution to 16×16; the convolution kernel of 1×1 is used to increase the number of channels to 384; the convolution kernel of 5×5 and groups of 384 is used to reduce the image resolution to 8×8; the convolution kernel of 1×1 is used to increase the number of channels to 768; the convolution kernel of 5×5 and groups of 768 is used to reduce the image resolution to 4×4; the convolution kernel of 1 ×1 convolution, the number of channels is reduced to 384; using a convolution kernel of 5×5 and a groups of 384, the image resolution is reduced to 2×2; using a convolution kernel of 1×1, the number of channels is increased to 768; using a convolution kernel of 5×5 and a groups of 768, the image resolution is reduced to 1×1; using a convolution kernel of 1×1, the number of channels is maintained at 768; using a deconvolution with a convolution kernel of 2×2, the image resolution is increased to 2×2; using a convolution kernel of 5×5 and a groups of 768, the image resolution is reduced to 1×1; using a convolution kernel of 1×1, the number of channels is maintained at 768; using a deconvolution with a convolution kernel of 2×2, the image resolution is increased to 2×2;
[0039] 3) Feature extraction network of scale 3 Figure 5As shown in the figure: a 20-layer multi-combination network is constructed, with a total of 16 convolutional layers, 3 deconvolutional layers, and 1 maximum pooling layer; the input image is a 3-channel color image with a resolution of 128×128; a convolution with a convolution kernel of 1×1 is used to maintain the number of channels to 3; a maximum pooling operation with a convolution kernel of 2×2 is used to reduce the image resolution to 64×64; a convolution with a convolution kernel of 4×4 is used to reduce the feature map resolution to 16×16, with a number of channels of 96; a convolution kernel of 1 ×1 convolution, the number of channels increases to 192; using a convolution kernel of 3×3 and a convolution of 192 groups reduces the image resolution to 8×8; using a convolution kernel of 1×1, the number of channels increases to 384; using a convolution kernel of 3×3 and a convolution of 384 groups reduces the image resolution to 4×4; using a convolution kernel of 1×1, the number of channels increases to 768; using a convolution kernel of 3×3 and a convolution of 768 groups reduces the image resolution to 4×4. The resolution is reduced to 2×2; the number of channels is reduced to 384 using a convolution with a convolution kernel of 1×1; the image resolution is reduced to 1×1 using a convolution with a convolution kernel of 3×3 and groups of 384; the number of channels is increased to 768 using a convolution with a convolution kernel of 1×1; the image resolution is increased to 2×2 using a deconvolution with a convolution kernel of 2×2; the image resolution is reduced to 1×1 using a convolution with a convolution kernel of 3×3 and groups of 768; the image resolution is reduced to 1×1 using a convolution with a convolution kernel of 1× 1, the number of channels remains at 768; using a 2×2 convolution kernel for deconvolution, the image resolution is increased to 2×2; using a 3×3 convolution kernel and a groups of 768 convolutions reduces the image resolution to 1×1; using a 1×1 convolution kernel, the number of channels remains at 768; using a 2×2 convolution kernel for deconvolution, the image resolution is increased to 2×2; using a 3×3 convolution kernel and a groups of 768 convolutions reduces the image resolution to 1×1.
[0040] (2) The feature fusion network fuses the feature maps obtained by the feature extraction network through the channel attention mechanism. The feature maps obtained in the three feature extraction modules in step 1 are accumulated to obtain 2304 feature maps, which are input into the channel that is paid attention to in the SqEx module.
[0041] (3) The classification module is further input into the average pooling layer with a convolution kernel of 1×1; the 2304 vectors are input into the Linear layer, and the output of the Linear layer is the number of categories of the samples.
[0042] Step 3: Train the MS-PLNet leaf spot disease recognition network. Pass the preprocessed training set through the built MS-PLNet leaf spot disease recognition network, compare the predicted results with the true labels, calculate the classification loss, update the model parameters through the momentum gradient descent algorithm, and obtain the MS-PLNet model for leaf spot disease recognition.
[0043] Step 4: Identification of mung bean leaf spot disease. Input the images in the validation set into the MS-PLNet model trained in step 3 to evaluate the performance of the model. Input the images in the test set into the trained model to obtain the leaf disease type of the image through forward propagation.
[0044] Based on the same inventive concept, the present invention also proposes a mung bean leaf spot disease identification device based on the MS-PLNet model, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the mung bean leaf spot disease identification method based on the MS-PLNet model is implemented.
[0045] The experimental results of the MS-PLNet model are shown in Figure 6 As shown, this embodiment is verified on a validation set and tested on a test set. It can be seen that the algorithm of this patent can bring a leaf spot disease detection accuracy of 99.96%, which is close to 100%. At the same time, the algorithm of this patent has better generalization and faster convergence speed during training.
[0046] The performance of the MS-PLNet model for mung bean leaf spot disease recognition was compared with that of the currently popular lightweight algorithm. This embodiment achieved a leaf spot disease detection accuracy of 99.96% with a model parameter of only 7.59M, which is far superior to the currently popular lightweight algorithm and can better detect crop leaf spot disease.
Claims
1. A method for identifying mung bean leaf spot disease based on the MS-PLNet model, characterized in that: The following steps are involved: (1) Preprocessing plant leaf disease images containing multiple disease types, and dividing them into a training set, a validation set, and a test set; (2) Constructing an MS-PLNet leaf spot disease recognition network; the MS-PLNet leaf spot disease recognition network includes three feature extraction modules, a feature fusion module and a classifier; the three feature extraction modules are constructed using ordinary convolution and improved depthwise separable convolution, during which the first feature extraction module increases the number of channels, the second feature extraction module reduces the number of channels, and the third feature extraction module increases the number of channels; the feature maps obtained by the feature extraction modules of the three scales are input into the feature fusion module using the channel attention mechanism; the feature maps output by the feature fusion are input into the classifier to obtain the classification result; the specific implementation process is as follows: (21) In the image input stage, the input 3-channel image is resized to 128×128 pixels, marked as IM0; the first feature extraction module inputs the 256×256 pixel image after the UnSample deconvolution operation, marked as IM1; The second feature extraction module inputs a 128×128 pixel image, marked as IM2; the third feature extraction module inputs a 64×64 pixel image after the MaxPool2d maximum pooling operation, marked as IM3; (22) In the feature extraction module of the first scale, IM1 is used as the input of the multi-convolution network to perform convolution with a kernel of 4×4 to obtain C11, C11 is subjected to point convolution with a kernel of 1×1 to increase the number of channels and a depth convolution of 7×7 to obtain C12, C12 is subjected to the same convolution calculation to obtain C13, C13 is subjected to the same convolution calculation again to obtain C14, C14 is subjected to point convolution with a kernel of 1×1 to reduce the number of channels and a depth convolution of 7×7 to obtain C15, C15 is convolved with a 1×1 point convolution kernel to increase the number of channels and a 7×7 depth convolution calculation to obtain C16, and C16 is convolved with the same convolution calculation to obtain C17; in the feature extraction module of the second scale, IM2 is used as the input of the multi-convolution network to perform a 4×4 convolution calculation to obtain C21, and C21 is convolved with a 1×1 point convolution kernel to increase the number of channels and a 5×5 depth convolution calculation to obtain C22, and C22 is convolved with the same convolution calculation to obtain C23, and C23 is further The same convolution calculation is performed to obtain C24, and C24 is subjected to a point convolution with a convolution kernel of 1×1 to reduce the number of channels and a depth convolution of 7×7 to obtain C25. C25 is subjected to a point convolution with a convolution kernel of 1×1 to increase the number of channels and a depth convolution of 7×7 to obtain C26. C26 is subjected to the same convolution calculation to obtain C27. In the feature extraction module of the third scale, IM3 is used as the input of the multi-convolution network to perform a convolution kernel of 4×4 to obtain C31. C31 is subjected to a convolution kernel of Perform a 1×1 point convolution to increase the number of channels and a 3×3 depth convolution to obtain C32, perform the same convolution calculation on C32 to obtain C33, perform the same convolution calculation on C33 to obtain C34, perform a 1×1 point convolution on C34 to reduce the number of channels and a 3×3 depth convolution to obtain C35, perform a 1×1 point convolution on C35 to increase the number of channels and a 3×3 depth convolution to obtain C36, and perform the same convolution calculation on C36 to obtain C37; (23) The feature maps C17, C27 and C37 obtained by the feature extraction modules of the three scales are concatenated to obtain R1, and the weights of each feature channel of R1 are readjusted through the channel attention mechanism, and the high-dimensional features are reasonably reorganized to obtain R2; (24) Input R2 into the average pooling layer, the fully connected layer and the Softmax activation function, and output the classification result; (3) Training the MS-PLNet leaf spot disease recognition network, passing the preprocessed training set through the built MS-PLNet leaf spot disease recognition network, comparing the obtained prediction results with the true labels to calculate the classification loss, and updating the model parameters through the momentum gradient descent algorithm to obtain the MS-PLNet model for leaf spot disease recognition; (4) Identification of mung bean leaf spot disease: The images in the validation set are input into the MS-PLNet model trained in step (3) to evaluate the performance of the model, and the images in the test set are input into the trained model to obtain the leaf disease type of the image through forward propagation.
2. The mung bean leaf spot disease identification method based on the MS-PLNet model according to claim 1, characterized in that: The image in step (1) is an image of a mung bean leaf taken with an ordinary camera. Data enhancement is performed on all leaf images, including random illumination transformation, left-right flipping, up-down flipping, diagonal flipping, and random cropping.
3. A mung bean leaf spot disease identification device based on the MS-PLNet model, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into a processor, the method for identifying mung bean leaf spot based on the MS-PLNet model according to any one of claims 1 to 2 is implemented.