UCV decomposition-based pest and disease classification method and system
By employing UCV decomposition and knowledge distillation techniques, the challenges of lightweighting and deploying pest and disease identification models have been addressed. This approach achieves a balance between model parameters and accuracy, improving identification accuracy and reducing inference latency. It is applicable to fields such as agriculture, autonomous driving, and aerospace.
Patent Information
- Application Number
- CN202511627556.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-07
- Publication Date
- 2026-03-03
AI Technical Summary
Existing pest and disease identification methods based on neural network architecture suffer from inference delays and insufficient identification accuracy in resource-constrained environments, leading to difficulties in deployment and economic losses.
We employ UCV decomposition and regularization techniques combined with the least squares method to decompose the model, and then use knowledge distillation for fine-grained fine-tuning to construct a lightweight pest and disease classification model.
It achieves a balance between model parameters and accuracy, improves recognition accuracy and reduces inference latency, and is applicable to various convolutional neural network architectures, suitable for fields such as agriculture, autonomous driving and aerospace.
Smart Images
Figure CN121598148A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and system for classifying pests and diseases based on UCV decomposition. Background Technology
[0002] With the increasing intensification of global agriculture and the intensification of climate change, pests and diseases pose a growing threat to crop production. Globally, crop losses due to pests and diseases reach 20%-40% annually, seriously threatening food security, farmers' livelihoods, and ecological balance. At the same time, frequent international trade exacerbates the risk of invasive species spreading; for example, invasive species such as the fall armyworm and red imported fire ants have already caused devastating damage to agriculture and ecosystems in many regions. Therefore, developing scientific and efficient pest and disease control methods is urgently needed.
[0003] Currently, methods based on computer vision and deep learning are being introduced to assist in the identification of pests and diseases, thereby achieving effective prevention and control. Compared to manual screening, these improved methods are more time-saving and labor-saving, and through large-scale application, they have increased the yield and quality of agricultural products.
[0004] However, existing methods based on neural network architectures suffer from severe inference latency and insufficient recognition accuracy in practical applications, leading to the inability to correctly distinguish between diseased and healthy plants and causing economic losses. Therefore, how to maintain high recognition accuracy under resource-constrained environments while reducing model complexity and inference latency is a problem that urgently needs to be solved in this field. Summary of the Invention
[0005] This invention aims to provide a pest and disease classification method and system based on UCV decomposition, which effectively solves the problems of lightweight models and difficult deployment, and achieves a balance between model parameters and accuracy.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a pest and disease classification method based on UCV decomposition, comprising the following steps: S1: Obtain the crop disease and pest dataset and divide the agricultural disease and pest dataset into training set and test set; S2: Pre-train the base model using the training set to obtain the initial model; S3: Decompose the convolution weight tensor of the initial model based on UCV and construct the optimization objective; S4: Decompose the optimization objective into multiple subproblems and solve them using the least squares method; S5: Construct new convolutional sub-layers based on the decomposed convolutional weight tensors to obtain intermediate models; S6: Use knowledge distillation to fine-tune the intermediate model to obtain the target model; S7: Test the target model using a test set, and deploy the target model to classify pests and diseases after the test is passed.
[0007] By adopting the above technical solutions, a novel decomposition scheme, namely UCV decomposition, was designed for convolution weights. Combined with regularization techniques, the least squares method was used for solving the problem, providing a good training foundation for the pest and disease model. In addition, knowledge distillation was introduced to fine-tune the model at a fine-grained level, achieving a significant performance improvement. This effectively solved the problems of lightweight models and difficult deployment, and achieved a balance between model parameters and accuracy.
[0008] Optionally, the base model in S2 can be any convolutional neural network model.
[0009] By adopting the above technical solution, the method of the present invention has good versatility and can adapt to a variety of recognition architectures based on convolutional neural networks, including but not limited to ResNet series, VGG series, etc.
[0010] Optionally, S3 includes: The fourth-order convolution weight tensor of the initial model Decomposed into a fourth-order core tensor and two third-order tensors , ; in, Represents the kernel size; and These represent the input and output channels, respectively. , , These represent the ranks in the decomposition process; Define the optimization goal: in, This represents minimizing the loss generated by the decomposition, i.e., constructing the optimization objective; The weights represent the initial model weights; Representative factor tensor , , The reconstructed tensor obtained by multiplication; , , Representing factor tensors respectively , , The balance coefficient; This represents the Frobenius norm.
[0011] Optionally, S4 includes: The optimization objective is broken down into three sub-problems: The factor tensor is obtained through multiple iterations. , , .
[0012] Optionally, S5 includes: Rewrite the convolution mapping as: in, Represents the output of the output layer; The convolutional layer is split into three smaller convolutional sub-layers: in, and These represent the outputs of the intermediate layers.
[0013] Optionally, S6 includes: Fine-tuning of the intermediate model is performed using the cross-loss function and the knowledge distillation loss function: in, Represents the cross-loss function; This represents the knowledge distillation loss function; Represents the distillation temperature; Represents the true label; and These represent the probability distributions output by the intermediate model and the target model, respectively. This represents the balance coefficient.
[0014] Optionally, S7 further includes: The test results of the target model are evaluated using the following formula: in, Represents compression ratio; and These represent the number of parameters in the initial model and the target model, respectively. This represents the proportion of the category with the highest probability in the target model's prediction results that matches the true label. This represents the proportion of the top five predicted categories in the target model's output that contain the true labels. and Represents respectively in and The total number of samples; The number of times the category with the highest probability matches the true label; This represents the number of times the target model's output contains the true label among the top five predicted categories.
[0015] Secondly, the present invention provides a pest and disease classification system based on UCV decomposition, used to implement a pest and disease classification method based on UCV decomposition as described in the first aspect, comprising: a data receiving unit, a model pre-training unit, a weight decomposition unit, a sub-problem splitting unit, a sub-layer reconstruction unit, a model fine-tuning unit, and a model testing unit connected in sequence.
[0016] In summary, the present invention has at least the following beneficial technical effects: 1. This method and system design a novel decomposition scheme for convolution weights, namely UCV decomposition, and combine it with regularization techniques, using the least squares method for solution, providing a good training base for pest and disease models.
[0017] 2. This method and system introduce knowledge distillation to fine-tune the model at a fine-grained level, thereby achieving a significant performance improvement.
[0018] 3. This method and system have good versatility and can be adapted to various recognition architectures based on convolutional neural networks, including but not limited to ResNet series, VGG series, etc.
[0019] 4. This method and system effectively solve the problems of lightweighting and difficult deployment of models, and achieve a balance between model parameters and accuracy. It can be widely used in agriculture, autonomous driving, aerospace and other fields. Attached Figure Description
[0020] Figure 1 This is a flowchart of a pest and disease classification method based on UCV decomposition in an embodiment of the present invention. Figure 2 This is a schematic diagram of a pest and disease classification system based on UCV decomposition in an embodiment of the present invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0022] The terminology used in the following embodiments of the present invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used in the specification and appended claims of the present invention, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in the present invention refers to and includes any or all possible combinations of one or more of the listed items. The terms “first” and “second” are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as “first” or “second” may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of the present invention, unless otherwise stated, “a plurality” means two or more.
[0023] This invention provides a pest and disease classification method based on UCV decomposition.
[0024] refer to Figure 1 and Figure 2 A pest and disease classification method based on UCV decomposition includes the following steps: S1: Obtain the crop disease and pest dataset, and divide the agricultural disease and pest dataset into training set and test set according to the preset ratio.
[0025] The data receiving unit acquires a crop disease and pest dataset (AI Challenger 2018, Agricultural Disease and Pest Research Library). The crop disease and pest dataset contains normal plants and diseased plants. Then, the target to be classified is labeled, thereby classifying diseased plants and assisting in the prevention and control of diseases and pests, which helps to improve the accuracy of decision-making.
[0026] The default ratio of the training set to the test set is 8:2.
[0027] S2: Use the training set to pre-train the base model to obtain the initial model.
[0028] The model training unit pre-trains the base model using the training set, calculates the gradient through backpropagation, and iteratively optimizes the initial model weights using the gradient information to obtain the initial model. After multiple rounds of pre-training, the initial model's final output is the classification accuracy, which is then displayed. This accuracy can then be used to identify diseased and pest-infested plants, demonstrating good feature representation capabilities and achieving the expected performance indicators in disease and pest classification. This provides a stable weight base for subsequent UCV decomposition and compression, contributing to improved decision-making accuracy.
[0029] The base model can be any convolutional neural network model.
[0030] S3: Decompose the convolution weight tensor of the initial model based on UCV and construct the optimization objective.
[0031] The weight decomposition unit decomposes the fourth-order convolution weight tensor of the initial model. Decomposed into a fourth-order core tensor and two third-order tensors , .
[0032] in, Represents the kernel size; and These represent the input and output channels, respectively. , , These represent the rank in the decomposition process.
[0033] Next, construct the optimization objective: in, This represents minimizing the loss generated by the decomposition, i.e., constructing the optimization objective; The weights represent the initial model weights; Representative factor tensor , , The reconstructed tensor obtained by multiplication; , , Representing factor tensors respectively , , The balance coefficient; This represents the Frobenius norm.
[0034] S4: Decompose the optimization objective into multiple subproblems and solve them using the least squares method.
[0035] The subproblem decomposition unit breaks down the optimization objective into three subproblems, which are solved alternately. The expressions for the three subproblems are: Through multiple iterations, the factor tensor can be obtained. , , .
[0036] S5: Construct new convolutional sub-layers based on the decomposed convolutional weight tensors to obtain intermediate models.
[0037] To accelerate the model, the sub-layer reconstruction unit uses a convolutional framework developed in PyTorch and constructs convolutional weights adapted to this framework from the decomposed tensors. For a convolutional layer, the mapping from input X to output Y is Y=WX. Using the above UCV decomposition, the convolutional mapping is rewritten as: in, This represents the output of the output layer.
[0038] At this point, a single convolutional layer can be split into three smaller convolutional sub-layers: in, and These represent the outputs of the intermediate layers.
[0039] According to UCV decomposition, the compression ratio is DDMN / (DDR1R2+MR1R3+NR2R3). For example, with 128 input channels, 64 output channels, a 3×3 convolution kernel size, and rank R1=R2=R3=1, the initial model layer has 128×64×3×3=73728 parameters. After compression, it is transformed into three smaller sub-layers with 1×1×128+3×3×1×1+64×1×1=201 parameters. The compression ratio of this layer is 73728 / 201≈367.
[0040] S6: Use knowledge distillation to fine-tune the intermediate model to obtain the target model.
[0041] To further enhance model performance, the model fine-tuning unit uses the cross-loss function as the classification loss function and introduces the knowledge distillation loss function to fine-tune the intermediate model to obtain the target model.
[0042] The expression for the total loss function is: in, Represents the cross-loss function; This represents the knowledge distillation loss function; Represents the distillation temperature; Represents the true label; and These represent the probability distributions output by the intermediate model and the target model, respectively. This represents the balance coefficient.
[0043] S7: Test the target model using a test set, and deploy the target model to classify pests and diseases after the test is passed.
[0044] The testing unit uses a test set to test the target model after UCV decomposition and evaluates the test results (including compression ratio, classification accuracy, training time, etc.). The evaluation formula is as follows: in, Represents compression ratio; and These represent the number of parameters in the initial model and the target model, respectively. This represents the proportion of the category with the highest probability in the target model's prediction results that matches the true label, and is used to reflect the accuracy of the target model under the preferred prediction. This represents the proportion of the top five predicted categories in the target model's output that contain the true labels. and Represents respectively in and The total number of samples; The number of times the category with the highest probability matches the true label; This represents the number of times the target model's output contains the true label among the top five predicted categories.
[0045] In general, even if the target model does not rank the correct category first, as long as that category appears in the first five predictions, it is considered a correct prediction.
[0046] Once the test is passed, the target model can be deployed to classify pests and diseases.
[0047] This invention also provides a pest and disease classification system based on UCV decomposition.
[0048] refer to Figure 2 A pest and disease classification system based on UCV decomposition includes: a data receiving unit, a model pre-training unit, a weight decomposition unit, a sub-problem splitting unit, a sub-layer reconstruction unit, a model fine-tuning unit, and a model testing unit connected in sequence.
[0049] The various variations and specific examples of the pest and disease classification method based on UCV decomposition provided in the above embodiments are also applicable to the pest and disease classification system based on UCV decomposition in this embodiment. Through the foregoing detailed description of the pest and disease classification method based on UCV decomposition, those skilled in the art can clearly understand the implementation method of the pest and disease classification system based on UCV decomposition in this embodiment. For the sake of brevity, it will not be described in detail here.
[0050] The above description of the embodiments is only used to provide a detailed introduction to the technical solution of the present invention. However, the description of the above embodiments is only for the purpose of helping to understand the method and core idea of the present invention, and should not be construed as a limitation of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be covered within the protection scope of the present invention.
Claims
1. A pest and disease classification method based on UCV decomposition, characterized in that, Includes the following steps: S1: Obtain the crop disease and pest dataset and divide the agricultural disease and pest dataset into training set and test set; S2: Pre-train the base model using the training set to obtain the initial model; S3: Decompose the convolution weight tensor of the initial model based on UCV and construct the optimization objective; S4: Decompose the optimization objective into multiple subproblems and solve them using the least squares method; S5: Construct new convolutional sub-layers based on the decomposed convolutional weight tensors to obtain intermediate models; S6: Use knowledge distillation to fine-tune the intermediate model to obtain the target model; S7: Test the target model using a test set, and deploy the target model to classify pests and diseases after the test is passed.
2. The pest and disease classification method based on UCV decomposition as described in claim 1, characterized in that, The base model in S2 is any convolutional neural network model.
3. The pest and disease classification method based on UCV decomposition as described in claim 1, characterized in that, S3 includes: The fourth-order convolution weight tensor of the initial model Decomposed into a fourth-order core tensor and two third-order tensors , ; in, Represents the kernel size; and These represent the input and output channels, respectively. , , These represent the ranks in the decomposition process; Define the optimization goal: in, This represents minimizing the loss generated by the decomposition, i.e., constructing the optimization objective; The weights represent the initial model weights; Representative factor tensor , , The reconstructed tensor obtained by multiplication; , , Representing factor tensors respectively , , The balance coefficient; This represents the Frobenius norm.
4. The pest and disease classification method based on UCV decomposition as described in claim 3, characterized in that, S4 includes: The optimization objective is broken down into three sub-problems: The factor tensor is obtained through multiple iterations. , , .
5. The pest and disease classification method based on UCV decomposition as described in claim 4, characterized in that, S5 includes: Rewrite the convolution mapping as: in, Represents the output of the output layer; The convolutional layer is split into three smaller convolutional sub-layers: in, and These represent the outputs of the intermediate layers.
6. The pest and disease classification method based on UCV decomposition as described in claim 5, characterized in that, S6 includes: Fine-tuning of the intermediate model is performed using the cross-loss function and the knowledge distillation loss function: in, Represents the cross-loss function; This represents the knowledge distillation loss function; Represents the distillation temperature; Represents the true label; and These represent the probability distributions output by the intermediate model and the target model, respectively. This represents the balance coefficient.
7. A pest and disease classification method based on UCV decomposition as described in any one of claims 1-6, characterized in that, The S7 also includes: The test results of the target model are evaluated using the following formula: in, Represents compression ratio; and These represent the number of parameters in the initial model and the target model, respectively. This represents the proportion of the category with the highest probability in the target model's prediction results that matches the true label. This represents the proportion of the top five predicted categories in the target model's output that contain the true labels. and Represents respectively in and The total number of samples; The number of times the category with the highest probability matches the true label; This represents the number of times the target model's output contains the true label among the top five predicted categories.
8. A pest and disease classification system based on UCV decomposition, characterized in that, A method for classifying pests and diseases based on UCV decomposition as described in any one of claims 1-7 includes: a data receiving unit, a model pre-training unit, a weight decomposition unit, a sub-problem splitting unit, a sub-layer reconstruction unit, a model fine-tuning unit, and a model testing unit connected in sequence.