Cotton leaf disease identification method and system based on multilayer context feature screening

Abstract

The invention relates to a cotton leaf disease identification method and system based on multilayer context feature screening, and belongs to the field of computer vision. According to the ConvNeXt-based cotton leaf disease identification method, a cotton leaf disease data set is constructed, and a system capable of identifying cotton physalospora piricola, brown blotch, angular leaf spot, verticillium wilt, fusarium wilt, anthracnose and powdery mildew and returning early warning information and prevention and treatment measures is established. According to the method provided by the invention, the ConvNeXt network is used as a backbone network, a feature enhancement module is added, the contrast among different regional features of the cotton leaves is improved, and on the basis, a feature screening attention mechanism of feature screening guided by multilayer context information is introduced, so that more effective feature screening is realized, and experiments show that the feature screening attention mechanism is more accurate. Compared with the recognition accuracy of the original ConvNeXt network, the recognition accuracy of the network model provided by the invention is improved by 4.1%.

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and in particular to a cotton leaf disease recognition method and system based on multi-layer context feature screening. Background Art

[0002] Cotton is susceptible to various diseases during its growth. Common cotton diseases include cotton ring spot, brown spot, angular spot, verticillium wilt, wilt, anthracnose, powdery mildew, etc. These diseases not only affect the normal growth of cotton, but may also lead to reduced cotton production or even total crop failure. Rapid and accurate identification of cotton leaf diseases is very important for farmers to formulate corresponding prevention and control strategies. However, the existing public cotton leaf disease dataset has the problem of data imbalance. The amount of data for some common diseases is much larger than that for other less common diseases. This imbalance may cause the model to be more inclined to identify common diseases during model training, while the recognition effect of a few diseases is poor, which ultimately affects the accuracy of overall disease recognition. Therefore, a cotton leaf disease dataset with a more balanced sample distribution is urgently needed.

[0003] In the traditional cotton industry, the identification of cotton leaf diseases often relies on the knowledge and practical experience of agricultural experts and cotton farmers. This method is not only inefficient and expensive, but also prone to misjudgment, which is not conducive to the prevention and control of cotton diseases. With the development of computer vision and image processing technology, computer vision technology has been widely used in the field of cotton leaf disease identification. Such as support vector machine algorithm, convolutional neural network algorithm, lightweight residual network recognition algorithm, and hybrid model of convolutional neural network and Transformer network, etc., although these algorithms can achieve a high accuracy rate, these methods are often in a controlled laboratory environment, and the background of cotton leaf disease images is relatively simple. When encountering minor diseases or the influence of real environmental factors such as lighting, shooting angle, background, etc., these algorithms are difficult to be well applied to data in real environments. Therefore, there is a need for a recognition method that can still have excellent results when facing complex backgrounds or minor diseases. Summary of the invention

[0004] The purpose of the present invention is to provide a cotton leaf disease recognition method and system based on multi-layer context feature screening, which solves the problem that it is difficult to accurately identify data with complex backgrounds or minor diseases affected by real environmental factors such as illumination, shooting angle, and background in the prior art. The present invention first collects data on cotton ring spot, brown spot, angular spot, verticillium wilt, wilt, anthracnose, powdery mildew, and healthy cotton leaf images, and combines public data sets to build a common cotton leaf disease data set with balanced samples. On this basis, a feature screening-feature enhancement normalized convolutional neural network recognition model based on multi-layer context information guidance is designed. The ConvNeXt network is used as the backbone network, and a feature enhancement module is added to improve the contrast between features in different regions of cotton leaves. On this basis, an attention mechanism for feature screening guided by multi-layer context information is introduced to achieve more effective feature screening, forming a cotton leaf disease recognition model with excellent performance. The model has the characteristics of fast recognition efficiency, high recognition accuracy, and excellent performance when processing complex backgrounds or minor diseases. On this basis, the present invention designs a cotton leaf disease identification system based on multi-layer context feature screening, which is convenient, efficient, easy to operate and visualized.

[0005] The above-mentioned object of the present invention is achieved by the following technical solutions:

[0006] A cotton leaf disease recognition method and system based on multi-layer context feature screening uses feature enhancement normalization methods to improve channel contrast and selectivity of feature data, enhance inter-channel feature competition of convolutional neural networks, and mine better feature data; at the same time, the attention mechanism of feature screening guided by context information is used to extract important information of high-level and low-level feature maps, screen effective features, and achieve efficient and accurate cotton leaf disease recognition, so as to quickly, conveniently and efficiently identify cotton leaf diseases for cotton growers, help farmers formulate prevention and control strategies as early as possible, and minimize farmers' losses. On this basis, a cotton leaf disease recognition system based on the screening of cotton leaf disease feature data guided by context information is designed and constructed, providing cotton practitioners with a convenient, efficient and easy-to-operate visualization system. The following steps are included:

[0007] Step 1: Construct a cotton leaf disease dataset, collect cotton leaf disease data, and preprocess the collected cotton leaf disease data. After preprocessing, divide the cotton leaf disease dataset and write the dataset label file.

[0008] Step 1.1: Collect image data of cotton ring spot, brown spot, angular spot, verticillium wilt, wilt, anthracnose, powdery mildew, and healthy cotton leaves; the data sources come from actual field photography, public datasets, agricultural research databases, and data shared by cooperative institutions; data collection covers different environmental conditions and different stages of disease development;

[0009] Step 1.2: Labeling the cotton leaf disease image data; using agricultural experts to manually label the collected cotton leaf disease data, and accurately label the collected image data; the labeling content includes the type of disease and the severity of the disease;

[0010] Step 1.3: Use data augmentation technology to expand cotton leaf disease image data; use image rotation, flipping, blurring, adding noise, and changing brightness to augment the collected cotton leaf disease data to increase the diversity and quantity of samples;

[0011] Step 1.4: Standardize the image, crop, resize, etc. the cotton leaf disease image. Perform center crop operation on the collected cotton leaf disease image, compare the long side and the short side, and then use the shortest side as the side length of the square to crop a square image from the center of the image; then perform resize operation to adjust the square image of cotton leaf disease obtained after center crop operation to a cotton leaf disease image of 224*224 size.

[0012] Step 1.5: Classify the cotton leaf disease categories, save the cotton leaf disease data into the folders of their respective disease categories according to the cotton leaf disease categories, and divide the cotton leaf disease data set into a training set and a test set; distribute the cotton leaf disease data obtained in steps 1.1-1.4 by random distribution according to the ratio of 8:2;

[0013] Step 1.6: Write a dataset label file, which includes the dataset category to which the image belongs, the cotton leaf disease category to which the image belongs, and the cotton leaf disease image data path.

[0014] Step 2: Design a context-guided feature selection attention mechanism-feature enhancer normalization convolutional neural network recognition model based on multi-layer context information guidance (CFSAM-FENConvNeXt). With the ConvNeXt network as the backbone network, a feature enhancement module is added to improve the contrast between the features of different regions of cotton leaves. On this basis, a multi-layer context information-guided feature selection attention mechanism is introduced to achieve more effective feature selection. Specifically:

[0015] Step 2.1: Use the ConvNeXt network as the backbone network to build the basic framework of the model. The present invention uses the currently more advanced ConvNeXt network model as the benchmark convolutional neural network model for improvement, and uses the stacking times of (3,3,27,3) to stack the blocks; the input dimensions of each stage are set to 128, 256, 512 and 1024 respectively.

[0016] Step 2.2: Add the feature enhancement normalization module (Feature Enhancer Normalization, FEN) to process the feature data. First, the feature map X of each channel i Summarize into a vector fx of a global function F(X), then calculate the normalization coefficient NC(F(X i )), then, a linear layer is used to perform a linear transformation on the input x, and a RELU activation function is applied to the output after the linear transformation to increase nonlinearity and enhance the contrast and selectivity of the feature data, as shown in the following formula:

[0017] X i =RELU(linear(X i ))

[0018] Finally, the features are calibrated using the calculated normalization coefficient NC(X), and a residual connection X is added between the input and output of the FEN module. i Finally, after being processed by the Feature Enhancement Normalization layer (FEN), the output data will be passed to the next layer in the neural network for further processing.

[0019] Step 2.3: Add a multi-layer context-guided feature selection attention mechanism (CFSAM). CFSAM is divided into two layers. The first layer is used to calculate the feature data of the feature selection guided by the context information. With Attention Searcher A t , the implementation formula is as follows:

[0020]

[0021]

[0022] in Represents low-level features The high-level features generated by the convolutional layers in a convolutional block containing N convolutional layers, W t and W A Indicates that W t Then, the weights of the multilayer perceptron activated by RELU are applied. Pool(·) represents the pooling operation. A t represents the attention searcher, W U is the weight of the multi-layer perceptron used for dynamic adjustment, ⊙ represents element-by-element multiplication, * represents point-by-point convolution, and σ is the sigmoid activation function.

[0023] The second layer uses the feature data of the first layer With Attention Searcher A t As input, As the initial low-level feature map, repeat the steps of the first layer; in this process, use the attention searcher A t Participate in guiding the new attention searcher A t+1 The generation of is expressed as: Finally, the characteristic data is obtained Then the final high-level feature data is generated using the feedforward operation

[0024] Among them, A t+1 represents the new attention searcher generated by the second layer, W U is the weight of the multi-layer perceptron used for dynamic adjustment.

[0025] Step 3: Set evaluation indicators and experimental parameters to implement model training, specifically:

[0026] Step 3.1: Set evaluation indicators. Use precision, recall, and F1 score as evaluation indicators to evaluate the performance of the model.

[0027] Step 3.2: Set the experimental parameters:

[0028] The batch size is set to 64. A too small batch size will slow down the convergence of the model during training and affect the training efficiency. However, increasing the batch size appropriately can improve the generalization ability of the model to a certain extent and accelerate the training process. The initial value of the learning rate is set to 0.0002. After multiple experimental verifications, it is found that 0.0002 is a more appropriate learning rate value, which can not only ensure the learning speed of the model, but also prevent overfitting. In addition, the optimizer algorithm is set to AdamW, the number of iterations is set to 150, and the learning rate decay rate is set to 0.05.

[0029] Step 4: Design a comparative experiment to test the usefulness of the method proposed in the present invention.

[0030] Step 4.1: Compare with other classic recognition models. The feature screening-feature enhancement normalized convolutional neural network recognition model guided by multi-layer context information is compared with ResNet50, ResNet101, VGG16, hornet, replknet, Swin_Transformer, and ConvNeXt networks to verify the usefulness of the present invention.

[0031] Step 4.2: Comparative experiments are conducted with other classic attention mechanisms to verify the effectiveness of the attention mechanism for feature screening guided by multi-layer context information of the present invention. CFSAM is compared with classic attention mechanisms such as SE, CA, ECA, and CBAM to verify the effectiveness of the present invention.

[0032] Step 4.3: Conducting a comparative experiment with other optimizers to verify the effectiveness of the AdamW optimizer used in the present invention. A comparative experiment was conducted between the Adam optimizer and the AdamW optimizer to verify that the AdamW optimizer is more suitable for the present invention.

[0033] Step 4.4: Conduct an ablation experiment to test the performance of the feature screening-feature enhancement normalized convolutional neural network recognition model guided by multi-layer context information.

[0034] Another object of the present invention is to provide a cotton leaf disease identification system based on multi-layer context feature screening, comprising the following modules:

[0035] The user module includes a common user submodule, a high-level user submodule, and an administrator submodule; the common user submodule is used by common users to upload cotton leaf disease data through mobile phones, tablet computers, or PC terminal devices; the high-level user submodule is used by high-level users to upload data manually, or to remotely control field shooting equipment through the system's preset automated process, so as to achieve real-time collection and uploading of disease data; the administrator submodule is used by administrators to manage various functions of the system. The administrator has the highest authority of the system and is responsible for the maintenance and management of the entire system; the administrator can create and update disease data sets, adjust the parameters of the disease identification model, and manage user permissions; the administrator can also monitor the system's operating status and generate system performance reports through the system's built-in analysis tools, so as to optimize and upgrade in a timely manner;

[0036] The data upload module receives and processes the cotton leaf disease image data uploaded by the user. It integrates image preprocessing functions, including automatic cropping and image resizing, to ensure data consistency and optimize the processing efficiency and accuracy of the recognition module.

[0037] The data set module includes a data preprocessing submodule and a data set construction submodule; the data preprocessing submodule preprocesses the uploaded cotton leaf disease image data. The preprocessing process not only includes conventional image adjustment operations, but also supports data enhancement generation. By various transformations of the original data, including rotation, flipping, and scaling, the training data set is enriched to further improve the generalization ability of the model; the data set construction submodule constructs a cotton leaf disease data set and a corresponding prevention and control measures database. These data sets are used for model training to improve the accuracy and efficiency of disease identification. When constructing the disease data set, the module supports multi-dimensional data annotation, including disease type, severity, and disease area;

[0038] The cotton leaf disease recognition module includes a model construction submodule, a model training submodule and a disease recognition submodule; the model construction submodule is responsible for building a disease recognition model for identifying cotton leaf diseases; the model training submodule uses the built disease data set to train the model; the disease recognition submodule receives the disease image data uploaded by the user end, and uses the trained model to perform recognition to obtain the corresponding recognition result;

[0039] The control measures query module includes a control measures database query submodule, an Internet query submodule, and an AI query submodule; the control measures database query submodule provides users with corresponding disease control suggestions and strategies by querying the internal control measures database. At the same time, the database is regularly updated to ensure that the information provided is the latest agricultural scientific research results; the Internet query submodule queries the latest disease control information from the Internet, supplements the missing content in the database, and provides users with more diverse control measures; the AI ​​query submodule calls a large language model, and the user directly asks the system about the control measures for specific diseases. The system generates personalized control plans based on the disease type and specific situation, and provides detailed execution steps and precautions to provide users with more personalized services;

[0040] The result output module includes an identification result output submodule and a control measure output submodule; the identification result output submodule outputs the result of disease identification to the user end for reference and use by the user; the control measure output submodule pushes the queried control measure information to the user end to help the user take effective control actions. In addition to traditional text output, it also supports the generation of an implementation plan for control measures, including a timetable and a list of required materials, to help users implement control measures efficiently; in addition, the system can also generate a QR code or link to facilitate users to share control suggestions with other agricultural experts or partners;

[0041] The early warning module is used to automatically detect the occurrence of diseases by real-time monitoring of uploaded data; when the system detects a potential major cotton disease threat, in addition to sending an alarm notification, it will automatically initiate an emergency response procedure.

[0042] The beneficial effects of the present invention are as follows: the cotton leaf disease identification method based on feature screening guided by multi-layer context information provided by the present invention adopts a feature enhancement normalization module to optimize the model's processing of feature data, processes the feature data initially obtained through the convolution layer, and mines out feature representations with more contrast and selectivity, effectively improving the feature's performance. By introducing the attention mechanism of feature screening guided by multi-layer context information, the information contained in high-level and low-level feature data is fully utilized, and the full utilization of high-level and low-level feature information enables the feature data to be utilized and mined more thoroughly, thereby achieving more effective feature screening. At the same time, the cotton leaf disease identification system based on feature screening guided by multi-layer context information provided by the present invention has the characteristics of high efficiency, accuracy, convenience and practicality. The system realizes high automation in various links such as data collection, preprocessing, model training, disease identification, and result output, ensuring the processing efficiency of large-scale data, and also facilitating the daily use of cotton farmers. Cotton farmers can use the visual interface of the system to conveniently and quickly identify cotton leaf diseases, thereby adjusting the prevention and control measures for cotton diseases in cotton fields. Cotton farmers can also use the AI ​​inquiry function to provide cotton farmers with a more comprehensive information and prevention and control measures for cotton diseases by the large model. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The drawings described herein are used to provide further understanding of the present invention and constitute a part of this application. The illustrative examples of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0044] Figure 1 It is a flow chart of the cotton leaf disease identification method based on multi-layer context feature screening of the present invention;

[0045] Figure 2 It is a structural diagram of a multi-layer context-guided feature screening-feature enhancement normalized convolutional neural network recognition model (CFSAM-FENConvNeXt) of the present invention;

[0046] Figure 3 The structure diagram of the multi-layer context-guided feature screening attention mechanism (CFSAM) of the present invention;

[0047] Figure 4 It is a schematic diagram of the structure of the cotton leaf disease identification system based on multi-layer context feature screening of the present invention;

[0048] Figure 5 A comparison chart of the effects of the cotton leaf disease identification method based on multi-layer context feature screening provided in an embodiment of the present invention.

[0049] Figure 6A flowchart of a cotton leaf disease identification system based on multi-layer context feature screening provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0051] See also Figures 1 to 6 As shown, the cotton leaf disease recognition method and system based on multi-layer context feature screening of the present invention constructs a cotton leaf disease data set, collects data on cotton ring spot, brown spot, angular spot, Verticillium wilt, wilt, anthracnose, powdery mildew and healthy cotton leaf images, and constructs a sample-balanced cotton leaf disease data set in combination with public data sets. On this basis, a feature screening-feature enhancement normalized convolutional neural network recognition model guided by multi-layer context information is designed. The ConvNeXt network is used as the backbone network, and a feature enhancement module is added to improve the contrast between the features of different regions of the cotton leaf. On this basis, the attention mechanism of feature screening guided by multi-layer context information is introduced to achieve more effective feature screening, forming a cotton leaf disease recognition model with excellent performance. The model has the characteristics of fast recognition efficiency, high recognition accuracy and excellent performance when processing complex backgrounds or minor diseases. On this basis, the present invention designs and constructs a feature screening-feature enhancement normalized convolutional neural network recognition system guided by multi-layer context information, which is convenient, efficient, easy to operate and visualized. Specifically, it includes the following steps:

[0052] Step 1: Construct a cotton leaf disease dataset. Collect cotton leaf disease data and preprocess the collected cotton leaf disease data. After preprocessing, divide the cotton leaf disease dataset and write the dataset label file.

[0053] Step 1.1: Mainly collect image data of cotton ring spot, brown spot, angular spot, verticillium wilt, wilt, anthracnose, powdery mildew, and healthy cotton leaf image data. The data sources come from actual field shooting, public data sets, agricultural research databases, and data shared by cooperative institutions. For field shooting data, high-resolution imaging equipment is used to ensure the clarity of disease characteristics. In addition, data collection covers different environmental conditions (such as light, humidity, soil conditions, etc.) and different stages of disease development to ensure the generalization ability of the model.

[0054] Step 1.2: Label the cotton leaf disease image data. Agricultural experts manually label the collected cotton leaf disease data and accurately label the collected image data. The labeling content includes the type of disease and the severity of the disease.

[0055] Step 1.3: Use data augmentation technology to expand cotton leaf disease image data. Use image rotation, flipping, blurring, adding noise, and changing brightness to augment the collected cotton leaf disease data and expand the dataset to increase the diversity and quantity of samples.

[0056] Step 1.4: Standardize the image, crop, resize, etc. the cotton leaf disease image. Perform center crop operation on the collected cotton leaf disease image, compare the long side and the short side, and then use the shortest side as the side length of the square to crop a square image from the center of the image; then perform resize operation to adjust the square image of cotton leaf disease obtained after center crop operation to a cotton leaf disease image of 224*224 size.

[0057] Step 1.5: Classify the cotton leaf disease categories, save the cotton leaf disease data into the folders of their respective disease categories according to the cotton leaf disease categories, and divide the cotton leaf disease data set into a training set and a test set. According to the ratio of 8:2, the cotton leaf disease data obtained after steps 1.1-1.4 are randomly allocated. The training set is used for model training, and the test set is used for model performance testing to ensure that the disease types in each set are evenly distributed, and the data set structure is well organized to ensure that the data set is clear and easy to use.

[0058] Step 1.6: Write a dataset label file, including the dataset category to which the image belongs, the cotton leaf disease category to which the image belongs, and the cotton leaf disease image data path. Create a label file, which records the detailed information of the cotton leaf disease dataset obtained after steps 1.1-1.5, including the dataset category to which the image belongs, the cotton leaf disease category to which the image belongs, and the image path information.

[0059] Step 2: Designed a feature screening-feature enhancement normalized convolutional neural network recognition model (CFSAM-FENConvNeXt) guided by multi-layer context information.

[0060] A feature screening-feature enhancement normalized convolutional neural network recognition model (CFSAM-FENConvNeXt) based on multi-layer contextual information guidance is proposed. It is characterized by taking the ConvNeXt network as the backbone network, adding a feature enhancement module, improving the contrast between the features of different regions of cotton leaves, and on this basis, introducing the attention mechanism of feature screening guided by multi-layer contextual information to achieve more effective feature screening.

[0061] Step 2.1: Use ConvNeXt network as the backbone network to build the basic framework of the model.

[0062] The present invention adopts the currently more advanced ConvNeXt network model as the benchmark convolutional neural network model for improvement, and uses the stacking times of (3,3,27,3) to stack the blocks; the input dimension size of each stage is set to 128, 256, 512 and 1024 respectively.

[0063] Each Block is a residual module. It includes a depthwise convolution and an inverted bottleneck layer. Each bottleneck layer, where the number of hidden channels is set to four times larger than the number of input channels, contains two point-wise convolutions. The normalization layer is selected as the layer normalization to avoid the problem of being affected by small batch data distribution. The Gaussian Error Linear Unit (GELU) is selected as the activation layer to avoid overfitting. The parameters related to the ConvNeXt network model are shown in the following table.

[0064] Table 1 ConvNeXt network model parameters

[0065]

[0066]

[0067] Based on this Block, a feature enhancement normalization module (FEN) is added after the dimension expansion layer in the inverted bottleneck of each Block to process the feature data obtained through the convolution layer, mine out more contrasting and selective feature representations, and obtain better feature data; then, a context-guided feature screening attention mechanism (CFSAM) is added after the inverted bottleneck layer of the Block in stage three and stage four. Through the iteratively generated attention searcher, the generation of channel and spatial attention weights is guided to output feature data with more contextual relevance.

[0068] Step 2.2: Add a feature enhancement normalization module (FEN) to process the feature data.

[0069] The input feature is represented as X∈R H*W*C , where H, W, C are the height, width, and number of channels.

[0070] First, capture the statistical information between channels and convert the feature map X of each channel i Summarize into a vector fx of a global function F(X), then calculate the normalization coefficient NC(x), which is the importance of the i-th channel relative to all other channels, expressed as:

[0071] Here, eps is a small constant that prevents the denominator from converging to zero.

[0072] Next, a linear layer is used to transform the input x linearly, and a RELU activation function is applied to the output after the linear transformation to increase nonlinearity, which is expressed as:

[0073] X i =RELU(linear(X i ))

[0074] Finally, the features are calibrated using the calculated normalization coefficient NC(x), and a residual connection X is added between the input and output of the FEN module. i Finally, after being processed by the Feature Enhancement Normalization layer (FEN), the output data will be passed to the next layer in the neural network for further processing.

[0075] Step 2.3: Add a multi-layer context-guided feature screening attention mechanism (CFSAM).

[0076] CFSAM is divided into two layers. The first layer is used to calculate the feature data filtered by the context information. With Attention Searcher A t , expressed as:

[0077]

[0078] in Represents low-level features The high-level features generated by the convolutional layers in a convolutional block containing N convolutional layers, W t and W A Indicates that W t Then, the weights of the multilayer perceptron activated by RELU are applied. Pool(·) represents the pooling operation. A t represents the attention searcher, W Uis the weight of the multi-layer perceptron used for dynamic adjustment, ⊙ represents element-by-element multiplication, * represents point-by-point convolution, and σ is the sigmoid activation function.

[0079] Next, the existing feature representation is transformed into With A t Multiply to get the channel scaled feature representation It is expressed as: Next, A s With X t ′0 Perform point-by-point convolution to obtain a two-dimensional spatial graph S map , expressed as: Then, S is multiplied by element-wise multiplication. map The sigmoid activation value is Multiply them together to get the spatially scaled feature representation, which is used as the input for the next layer of calculation and is expressed as:

[0080]

[0081] Where ⊙ represents element-wise multiplication, * represents point-wise convolution, and σ is the sigmoid activation function.

[0082] The second layer uses the feature data of the first layer With Attention Searcher A t As input, As the initial low-level feature map, repeat the steps of the first layer; in this process, use the attention searcher A t Participate in guiding the new attention searcher A t+1 The generation of is expressed as:

[0083]

[0084] Among them, A t+1 represents the new attention searcher generated by the second layer, W U is the weight of the multi-layer perceptron used for dynamic adjustment.

[0085] Finally, the characteristic data is obtained Then the final feature data is generated using the feedforward operation

[0086] Step 3: Set evaluation indicators and experimental parameters to implement model training.

[0087] Step 3.1: Set evaluation indicators. The performance of the model is evaluated mainly using accuracy, recall, and F1 score.

[0088] TP (True Positive): indicates a positive example that is predicted correctly.

[0089] FP (False Positive): indicates a counterexample of incorrect prediction.

[0090] FN (False Negative): indicates a positive example with incorrect prediction.

[0091] TN (True Negative): indicates a counterexample that is predicted correctly.

[0092] Accuracy: The ratio of the number of correct predictions in positive and negative examples to the total number, expressed as:

[0093]

[0094] Recall rate: The ratio of the predicted correct positive samples to the total positive positive samples among the actual positive samples, based on the actual samples. Among the actual positive samples, either the predicted correct TP or the predicted incorrect FN is expressed as:

[0095]

[0096] Precision: The proportion of samples predicted as positive examples that are correctly predicted based on the prediction results. There are two types of results predicted as positive examples, either actually positive examples TP or actually negative examples FP, expressed as:

[0097]

[0098] F1 value (F1 score):

[0099]

[0100] Step 3.2: Set the experimental parameters:

[0101] The present invention sets the batch size to 64. Too small a batch size will cause the convergence speed of the model to slow down during the training process, affecting the training efficiency. By appropriately increasing the batch size, the generalization ability of the model can be improved to a certain extent, while accelerating the training process. In the present invention, the initial value of the learning rate is set to 0.0002. Through multiple experimental verifications, it is found that 0.0002 is a more suitable learning rate value, which can not only ensure the learning speed of the model, but also prevent the occurrence of overfitting. In addition, the present invention sets the optimizer algorithm to AdamW, the number of iterations is set to 150, and the learning rate decay rate is set to 0.05.

[0102] Step 4: Conduct various comparative experiments to verify the usefulness of the method proposed in the present invention.

[0103] Step 4.1: Design a comparative experiment with other classic recognition models. The feature screening-feature enhancement normalized convolutional neural network recognition model guided by multi-layer context information is compared with ResNet50, ResNet101, VGG16, hornet, replknet, Swin_Transformer, and ConvNeXt networks to verify the usefulness of the present invention.

[0104] (1) ResNet (Residual Network): Proposed by Microsoft Research, ResNet mainly solves the gradient vanishing problem in deep neural networks. Its core idea is the residual block, which introduces skip connections to enable information to be directly transmitted between previous and next layers. The residual connections used in the ResNet network solve the gradient vanishing problem that is prone to occur in deep networks, allowing the network to be deeper without causing a decrease in model performance.

[0105] Among them, ResNet50 has a depth of 50 layers and is a relatively balanced deep neural network that maintains the depth advantage of the network without being too complex.

[0106] Among them, ResNet101 has a depth of 101 layers, which is suitable for more complex tasks and can extract more detailed features, but it requires higher computing resources than ResNet50.

[0107] (2) VGG16: The VGG network was proposed by researchers at Oxford University. It uses a very deep convolutional network for feature extraction. The VGG network is simple in design, with convolutional layers and pooling layers stacked alternately. It is easy to understand and implement, but the computational overhead is high, especially the memory consumption. VGG16 has a depth of 16 layers, including 13 convolutional layers and 3 fully connected layers. The design of the VGG network adopts a very regular convolutional structure. All convolution kernels are 3x3 in size and the pooling layers are 2x2.

[0108] (3) Hornet: Hornet is a newer neural network architecture that focuses on performance on images and other perception tasks, combining the advantages of convolution and multi-head self-attention mechanisms. It uses a new technology called dynamic convolution, which enables the network to process local and global information of the image more effectively. Hornet excels at maintaining high performance without relying on a large amount of computing resources, and is particularly suitable for scenarios that require efficient computing and real-time processing.

[0109] (4) RepLKNet: RepLKNet is a network architecture that introduces very large convolution kernels. Unlike traditional convolutional networks, RepLKNet uses ultra-large convolution kernels (such as 31x31 or 61x61) to better capture global information, while reducing the computational burden through re-parameterization technology. By using large convolution kernels, RepLKNet can effectively capture global contextual information and reduce the computational burden with the help of re-parameterization technology, which is particularly suitable for large-scale image and video processing tasks.

[0110] (5) Swin Transformer: Swin Transformer is a network that combines the self-attention mechanism and the hierarchical structure. It was originally designed to solve the efficiency problem of Transformer in visual tasks. Swin Transformer uses a shifted window attention mechanism to divide the image into small windows and perform local self-attention calculations in each window, thereby reducing computational complexity. The hierarchical structure of Swin Transformer enables it to adapt to input images of different sizes and extract multi-scale features. Compared with traditional convolutional neural networks (CNNs), Swin Transformer has significant advantages in capturing long-distance dependencies and global information, and is suitable for image classification, object detection, and segmentation tasks.

[0111] (6) ConvNeXt: ConvNeXt is a convolutional neural network architecture proposed in 2022, which aims to re-examine and optimize the traditional convolutional network structure. It draws on the design concepts of Transformer, such as LayerNorm, GELU activation function, etc., while retaining the advantages of convolutional networks to form a more modern convolutional network architecture. The efficiency and performance of the traditional CNN architecture are improved through modern convolution operations, making convolutional networks competitive again in image classification tasks. ConvNeXt not only has efficient computing power, but also is simple and easy to implement.

[0112] Table 2 Main experimental results, the best results are shown in bold

[0113]

[0114] According to the data in the table, the CFSAM-FENConvNeXt model performed best in the disease recognition task and had the highest recognition accuracy. Its accuracy, F1 value, and regression rate in the cotton leaf disease recognition task all reached the highest level, with an accuracy of 94.4%. This result shows that the CFSAM-FENConvNeXt model has strong performance and robustness when dealing with disease recognition tasks, and has a clear advantage in accuracy compared to other models such as ConvNeXt and Swin Transformer.

[0115] Step 4.2: Conduct comparative experiments with other classic attention mechanisms to verify the effectiveness of the attention mechanism for feature screening guided by multi-layer context information of the present invention.

[0116] (1) SE (Squeeze-and-Excitation) attention mechanism. The SE module automatically learns which feature channels are more important through a global channel attention mechanism. The SE attention mechanism has a simple structure and can effectively improve the performance of the network.

[0117] (2) CA (Coordinate Attention) mechanism. The CA module is a new coordinate attention mechanism that combines spatial and channel attention to combine position information and channel information. Compared with the traditional SE mechanism, CA can retain more position information and is suitable for mobile networks and other lightweight network architectures.

[0118] (3) ECA (Efficient Channel Attention) mechanism. ECA improves the SE mechanism by removing the complex fully connected layer and realizing local interaction between channels through convolution operations. The ECA attention mechanism is more efficient and is particularly suitable for tasks and models that are sensitive to parameters.

[0119] (4) CBAM (Convolutional Block Attention Module) attention mechanism. The CBAM module combines channel attention and spatial attention by gradually focusing on different important information. CBAM combines channel and spatial attention while maintaining a small computational overhead, making it suitable for various visual tasks.

[0120] Table 3 Comparison experiment with other classic attention mechanisms

[0121]

[0122] According to the data in the table, the comprehensive accuracy, F1 value and regression rate show that the ConvNeXt network model with the CFSAM attention mechanism performs best. This shows that CFSAM can more accurately identify targets and minimize false detections in the cotton leaf disease recognition task.

[0123] Step 4.3: Conduct comparative experiments with other optimizers to verify the effectiveness of the AdamW optimizer used in the present invention.

[0124] (1) Adam Optimizer: Adam is a momentum-based adaptive learning rate optimizer that combines two classic optimization methods: momentum and RMSProp. It accelerates gradient descent through momentum and dynamically adjusts the learning rate for each parameter.

[0125] (2) AdamW optimizer: AdamW is an improvement on the Adam optimizer. It proposes the decoupling of weight decay (WeightDecay) from gradient updates, thus solving the shortcomings of the Adam optimizer in regularization. AdamW separates weight decay from Adam's momentum calculation during gradient updates, and directly subtracts the proportional value of the weight at each update. This separation reduces the bias problem of the Adam optimizer in regularization.

[0126] Table 4 Comparison experiment between Adam and AdamW optimizers

[0127]

[0128] According to the data in the table, it can be concluded that the AdamW optimizer is excellent in all indicators. By adopting the AdamW optimizer, the performance of the network can be effectively enhanced.

[0129] Step 4.4: Conduct an ablation experiment to test the performance of the feature screening-feature enhancement normalized convolutional neural network recognition model guided by multi-layer context information.

[0130] Table 5 Ablation experiment

[0131]

[0132] Adding CFSAM and FEN feature configurations significantly improves the performance of the ConvNeXt model. Specifically, when using CFSAM and FEN at the same time, the model's accuracy, F1 score, and recall rate all reach the best level, with an accuracy of 94.4%, an F1 score of 94.8%, and a recall rate of 94.8%. This shows that by introducing these two feature configurations, the overall performance and accuracy of the model can be significantly improved.

[0133] Step 5: A cotton leaf disease recognition system based on feature screening guided by multi-layer context information was designed. The system includes the following modules:

[0134] User module: The user module includes a common user submodule, a senior user submodule, and an administrator submodule. The common user submodule is used for common users to upload cotton leaf disease data through terminal devices such as mobile phones, tablet computers or PCs. This module supports uploading of multiple data formats, such as images, videos, etc., and integrates a guiding interface for data uploading to ensure the standardization of data quality. The senior user submodule is used for senior users to not only manually upload data, but also remotely control field shooting equipment through the system's preset automated process to achieve real-time collection and uploading of disease data. In addition, this submodule also integrates a data annotation tool, allowing senior users to make detailed annotations on the uploaded data and provide high-quality data sets for subsequent model training. Senior users have higher permissions and can access advanced functions and specific data sets. The administrator submodule is used for administrators to manage various functions of the system through this submodule. The administrator has the highest authority of the system and is responsible for the maintenance and management of the entire system. The administrator can create and update disease data sets, adjust the parameters of the disease recognition model, and manage user permissions to ensure that the operating permissions of common users and senior users are properly configured. Administrators can also use the system's built-in analysis tools to monitor the system's operating status and generate system performance reports for timely optimization and upgrades.

[0135] Data upload module: The data upload module is responsible for receiving and processing the cotton leaf disease image data uploaded from the user end. This module integrates image preprocessing functions such as automatic cropping, automatic image resizing, color space normalization, etc. to ensure data consistency and optimize the processing efficiency and accuracy of the recognition module. The processed data will be transmitted to the subsequent modules for disease identification and analysis.

[0136] Dataset module: The dataset module includes a data preprocessing submodule and a dataset construction submodule. The data preprocessing submodule is used to preprocess the uploaded cotton leaf disease image data. The preprocessing process not only includes conventional image adjustment operations, but also integrates data cleaning functions, which can identify and eliminate low-quality or redundant data. This submodule also supports enhanced data generation. Through various transformations of the original data (such as rotation, flipping, scaling, etc.), the training data set is enriched to further enhance the generalization ability of the model. The dataset construction submodule is used to construct a cotton leaf disease dataset and a corresponding prevention and control measures database. These datasets will be used for model training to improve the accuracy and efficiency of disease identification. When constructing the disease dataset, the module supports multi-dimensional data annotation, such as disease type, severity, disease area, etc.

[0137] Cotton leaf disease identification module: The cotton leaf disease identification module includes a model construction submodule, a model training submodule and a disease identification submodule. The model construction submodule is responsible for constructing a disease identification model for identifying cotton leaf diseases, and is combined with a context-guided feature screening mechanism to ensure that the model can effectively capture the key features of cotton leaf diseases during the training process. The model training submodule is used to train the model using the constructed disease data set. During the continuous optimization process, the model gradually improves the recognition accuracy of cotton leaf diseases. During the training process, the system supports distributed training acceleration and can perform parallel computing on multiple GPUs or the cloud, greatly shortening the training time. The disease identification submodule is used to receive disease image data uploaded by the user end, and use the trained model for identification, output the corresponding disease category and feature information, and obtain the corresponding identification result.

[0138] Prevention and control measures query module: The prevention and control measures query module includes a prevention and control measures database query submodule, an Internet query submodule, and an AI query submodule. The prevention and control measures database query submodule is used to provide users with corresponding disease prevention and control suggestions and strategies by querying the internal prevention and control measures database. At the same time, the database will be updated regularly to ensure that the information provided is the latest agricultural scientific research results. The Internet query submodule queries the latest disease prevention and control information from the Internet, supplements the content that may be missing from the database, and provides users with more diverse prevention and control measures. The AI ​​inquiry submodule is used to call large language models (such as GPT, Wenxin Yiyan, etc.). Users can directly ask the system about prevention and control measures for specific diseases. The system will generate personalized prevention and control plans based on the type of disease and specific circumstances, and provide detailed implementation steps and precautions to provide users with more personalized services.

[0139] Result output module: The result output module includes an identification result output submodule and a prevention and control measures output submodule. The identification result output submodule is used to output the results of disease identification to the user end for user reference and use. The identification result output is not limited to simple text descriptions. The system also supports multiple forms of output, such as disease table generation, data report generation (including statistical analysis and trend forecasting), etc., to provide users with a variety of result presentation methods. The prevention and control measures output submodule is used to push the queried prevention and control measures information to the user end to help users take effective prevention and control actions. In addition to traditional text output, this submodule also supports the generation of implementation plans for prevention and control measures, including timetables, lists of required materials, etc., to help users implement prevention and control measures efficiently. In addition, the system can also generate QR codes or links to facilitate users to share prevention and control suggestions with other agricultural experts or partners.

[0140] Early warning module: The early warning module is used to automatically detect the occurrence of diseases by real-time monitoring of uploaded data. When the system detects a potential threat of a major cotton disease, in addition to sending an alarm notification, it will automatically initiate emergency response procedures, such as sending an emergency report to the administrator, automatically updating the prevention and control measures database, etc., to ensure that the problem is handled in a timely manner.

[0141] Embodiment 1:

[0142] In order to verify the effectiveness of the cotton leaf disease recognition method based on feature screening of multi-layer context information proposed in this invention, a visualization analysis of the final effect of using the traditional cotton disease recognition model is performed. Figure 5 As shown, through in-depth analysis, it is found that the traditional cotton disease recognition model can obtain better recognition results when identifying cotton leaf disease data with simple background and obvious disease traits. However, when faced with cotton leaf disease data with milder disease traits and less obvious disease traits, its recognition performance is lower. The experimental results show that the traditional cotton disease recognition model is sometimes not effective when dealing with complex scenes or mild diseases. The method of the present invention can obtain better recognition results in both situations, reflecting the effectiveness of the cotton leaf disease recognition method based on multi-layer context feature screening proposed by the present invention.

[0143] Embodiment 2:

[0144] In order to intuitively demonstrate the convenience and effectiveness of the cotton leaf disease recognition system based on multi-layer context feature screening in daily use by cotton farmers, the cotton leaf disease recognition system based on multi-layer context feature screening was demonstrated to cotton farmers.

[0145] See also Figure 6 As shown, a flowchart of the cotton leaf disease recognition system based on multi-layer context feature screening is shown. First, the user logs in to the cotton leaf disease recognition system page. Then, the user selects the image file that he wants to identify and submits it. The image file submitted by the user will be directly displayed on the page for the user to check whether the submitted file is correct. If the submission is wrong, you can reselect a new image file to submit. If the selection is correct, the selected image enters the system for disease recognition. If the recognition result is not diseased, the result is directly output. If the recognition result is diseased, the disease warning information is submitted, and the prevention and control measures of the disease are queried at the same time. The prevention and control measures are a collection of cotton disease prevention and control measures database, Internet query and AI large model query. Finally, the recognition results, warning information and disease prevention and control measures are output.

[0146] From the above examples, it can be seen that the cotton leaf disease recognition method based on multi-layer context feature screening of the present invention processes the feature data initially obtained through the convolution layer by adopting feature enhancement normalization, and mines out more contrasting and selective feature representations, which effectively improves the performance of the features; at the same time, the attention mechanism of multi-layer context feature screening is adopted to make full use of the information contained in the high-level and low-level feature data, and the full use of high-level and low-level feature information enables the feature data to be used and mined more thoroughly, and the data of complex backgrounds or minor diseases can also be effectively identified. The cotton leaf disease recognition system based on multi-layer context feature screening of the present invention has the characteristics of high efficiency, accuracy, convenience and practicality. The system realizes high automation in various links such as data acquisition, preprocessing, feature extraction, model training, disease recognition, and result output, ensuring the processing efficiency and recognition accuracy of large-scale data. The CFSAM-FENConvNeXt model adopted integrates the context-guided feature screening mechanism and the feature enhancement normalization module, which can effectively identify and distinguish different types of cotton leaf diseases.

[0147] The present invention proposes a cotton leaf disease recognition method based on ConvNeXt, and constructs a cotton leaf disease data set, clarifies the characteristics of various diseases, establishes a system that can identify cotton ring spot, brown spot, angular spot, verticillium wilt, wilt, anthracnose, powdery mildew and return warning information and prevention and control measures, and also performs data annotation. In the annotation process, the annotator mainly focuses on the different characteristics of cotton leaf diseases, including the shape and color change of the lesions and the expansion of the diseases. The data set not only marks the occurrence area of ​​the disease, but also marks the environmental factors that may affect the expansion of the disease, so as to carry out further analysis and optimization. The method proposed by the present invention uses the ConvNeXt network as the backbone network, adds a feature enhancement module, improves the contrast between the features of different regions of the cotton leaf, and on this basis, introduces the attention mechanism of feature screening guided by multi-layer context information, and realizes more effective feature screening. Experiments show that the recognition accuracy of the network model proposed by the present invention is improved by 4.1% compared with the original ConvNeXt network.

[0148] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made to the present invention shall be included in the protection scope of the present invention.

Claims

1. A cotton leaf disease identification method based on multi-layer context feature screening, characterized in that: The following steps are involved: Step (1): Constructing a cotton leaf disease dataset: collecting cotton leaf disease data and preprocessing the collected cotton leaf disease data. After preprocessing, the cotton leaf disease dataset is divided and a dataset label file is written; Step (2): Design a feature screening-feature enhancement normalized convolutional neural network recognition model based on multi-layer context information guidance: Using the ConvNeXt network as the backbone network, the feature enhancement module and the attention mechanism of feature screening guided by multi-layer context information are adopted to form a more effective feature screening cotton leaf disease recognition model; Step (3): Set evaluation indicators and experimental parameters to implement model training.

2. The cotton leaf disease identification method based on multi-layer context feature screening according to claim 1, characterized in that: The step (1) of constructing a cotton leaf disease dataset comprises: Step (1.1): Collect image data of cotton ring spot, brown spot, angular spot, verticillium wilt, wilt, anthracnose, powdery mildew, and healthy cotton leaves; the data sources come from actual field photography, public data sets, agricultural research databases, and data shared by cooperative institutions; data collection covers different environmental conditions and different stages of disease development; Step (1.2): labeling the cotton leaf disease image data; using agricultural experts to manually label the collected cotton leaf disease data, and accurately labeling the collected image data; the labeling content includes the type of disease and the severity of the disease; Step (1.3): Use data enhancement technology to expand the cotton leaf disease image data; use image rotation, flipping, blurring, adding noise, and changing brightness to augment the collected cotton leaf disease data to expand the data set and increase the diversity and quantity of samples; Step (1.4): standardize the image, crop and resize the cotton leaf disease image; Step (1.5): Classify the cotton leaf disease categories, save the cotton leaf disease data into the folders of their respective disease categories according to the cotton leaf disease categories, and divide the cotton leaf disease data set into a training set and a test set; distribute the cotton leaf disease data obtained through steps (1.1)-(1.4) by random allocation according to a ratio of 8:2; Step (1.6): Write a dataset label file, including the dataset category to which the image belongs, the cotton leaf disease category to which the image belongs, and the cotton leaf disease image data path.

3. The cotton leaf disease identification method based on multi-layer context feature screening according to claim 1, characterized in that: The design described in step (2) is based on a feature screening-feature enhancement normalized convolutional neural network recognition model guided by multi-layer context information, specifically: Step (2.1): Use the ConvNeXt network as the backbone network to build the basic framework of the model; use the ConvNeXt network model as the baseline convolutional neural network model for improvement, and use the stacking times of (3,3,27,3) to stack the blocks; the input dimensions of each stage are set to 128, 256, 512 and 1024 respectively; Step (2.2): Add the feature enhancement normalization module FEN to process the feature data; first, transform the feature map X of each channel i Summarize into a vector fx of a global function F(X), and then calculate the normalization coefficient NC(F(X i )), then, a linear layer is used to perform a linear transformation on the input x, and a RELU activation function is applied to the output after the linear transformation to increase nonlinearity and enhance the contrast and selectivity of the feature data, as shown in the following formula: X i =RELU(linear(X i )) The features are calibrated using the calculated normalization coefficient NC(X), and a residual connection X is added between the input and output of the FEN module i ;Finally, after being processed by the feature enhancement normalization layer FEN, the output data is passed to the next layer in the neural network for further processing; Step (2.3): Add a multi-layer context-guided feature selection attention mechanism CFSAM. CFSAM is divided into two layers. The first layer calculates the feature data of feature selection guided by context information. With Attention Searcher A t , the implementation formula is as follows: in, Represents low-level features The high-level features generated by the convolutional layers in a convolutional block containing N convolutional layers, W t and W A Indicates that W t Then, the weights of the multilayer perceptron activated by RELU are applied. Pool(·) represents the pooling operation. A t represents the attention searcher, W U is the weight of the multilayer perceptron used for dynamic adjustment, ⊙ represents element-by-element multiplication, * represents point-by-point convolution, and σ is the sigmoid activation function; The second layer uses the feature data of the first layer With Attention Searcher A t As input, As the initial low-level feature map, repeat the steps of the first layer; in this process, use the attention searcher A t Participate in guiding the new attention searcher A t+1 The generation of is expressed as: Finally, the characteristic data is obtained Then the final high-level feature data is generated using the feedforward operation Among them, A t+1 represents the new attention searcher generated by the second layer, W U are the weights of the MLP used for dynamic adjustment.

4. The cotton leaf disease identification method based on multi-layer context feature screening according to claim 1, characterized in that: The evaluation index and experimental parameters set in step (3) are specifically: Step (3.1): Set evaluation indicators: Use accuracy, recall, and F1 score as evaluation indicators to evaluate the performance of the model; Step (3.2): Set the experimental parameters: set the batch size to 64, the initial value of the learning rate to 0.0002, the optimizer algorithm to AdamW, the number of iterations to 150, and the learning rate decay rate to 0.

05.

5. A cotton leaf disease recognition system based on multi-layer context feature screening, characterized in that: The system includes the following modules: The user module includes a common user submodule, a high-level user submodule, and an administrator submodule; the common user submodule is used for common users to upload cotton leaf disease data through mobile phones, tablet computers, or PC terminal devices; the high-level user submodule is used for high-level users to upload data manually, or remotely control field shooting equipment through the system's preset automated process, to achieve real-time collection and uploading of disease data; The administrator submodule is used by the administrator to manage the various functions of the system. The administrator has the highest authority in the system and is responsible for the maintenance and management of the entire system; The data upload module receives and processes the cotton leaf disease image data uploaded by the user. It integrates image preprocessing functions, including automatic cropping and image resizing, to ensure data consistency and optimize the processing efficiency and accuracy of the recognition module. The dataset module includes a data preprocessing submodule and a dataset construction submodule; The data preprocessing submodule preprocesses the uploaded cotton leaf disease image data. The preprocessing process not only includes conventional image adjustment operations, but also supports data enhancement generation. Through various transformations of the original data, including rotation, flipping, and scaling, the training data set is enriched to further improve the generalization ability of the model. The dataset construction submodule constructs the cotton leaf disease dataset and the corresponding prevention and control measures database. These datasets are used for model training to improve the accuracy and efficiency of disease identification. When constructing the disease dataset, the module supports multi-dimensional data annotation, including disease type, severity, and disease area. The cotton leaf disease recognition module includes a model construction submodule, a model training submodule and a disease recognition submodule; the model construction submodule is responsible for building a disease recognition model for identifying cotton leaf diseases; the model training submodule uses the built disease data set to train the model; the disease recognition submodule receives the disease image data uploaded by the user end, and uses the trained model to perform recognition to obtain the corresponding recognition result; The control measures query module includes a control measures database query submodule, an Internet query submodule, and an AI query submodule; the control measures database query submodule provides users with disease control suggestions and strategies by querying the internal control measures database; the Internet query submodule queries the latest disease control information from the Internet, supplements the missing content in the database, and provides users with more diverse control measures; the AI ​​query submodule calls a large language model, and the user directly asks the system about the control measures for specific diseases. The system generates a personalized control plan based on the disease type and specific situation, and provides detailed execution steps and precautions to provide users with more personalized services; A result output module, including an identification result output submodule and a control measure output submodule; The identification result output submodule outputs the result of disease identification to the user end; the control measure output submodule pushes the queried control measure information to the user end; the system can also generate a QR code or link; The early warning module automatically detects the occurrence of diseases by real-time monitoring of uploaded data; when the system detects a potential major cotton disease threat, in addition to sending an alarm notification, it will automatically initiate an emergency response program.

Images