Plant disease detection method based on deep learning

Abstract

The invention provides a plant disease detection method based on deep learning, and relates to the technical field of image processing, and the method comprises the steps: constructing a plant disease detection model through employing a feature extraction layer, a neck network layer and a head network layer; wherein the feature extraction layer comprises a full-dimensional dynamic convolution module, the neck network layer comprises an efficient multi-scale attention mechanism, and the head network layer comprises a dynamic target detection module; training the plant disease detection model by using a mixed loss function to obtain a trained plant disease detection model; wherein the mixed loss function comprises bounding box loss, classification loss and distributed loss; and processing the plant leaf data set by using the trained plant disease detection model to obtain a plant disease detection result so as to complete the detection of the plant disease. The problems that existing plant disease detection is high in cost and insufficient in real-time performance are solved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a plant disease detection method based on deep learning. Background Art

[0002] There are currently several options for monitoring tomato leaf diseases: manual detection: identifying the current disease condition of tomato leaves with the naked eye; computer vision technology: taking pictures of tomato leaves, using image processing technology and machine learning models to automatically determine tomato leaf diseases; spectral analysis: capturing the reflectivity of plants under different spectra to identify early symptoms of diseases; sensor technology: installing sensors near the plant or on the surface of the leaves to monitor the humidity, temperature and other physical and chemical parameters of the leaves to predict or detect diseases; molecular biology methods: such as detecting disease-related DNA or RNA sequences through methods such as PCR, which can detect diseases at an early stage.

[0003] The existing technical solutions have the following problems: high labor costs: manual detection requires a lot of labor, especially in large-scale farmland, frequent disease monitoring will greatly increase labor costs; high equipment costs: methods such as spectral analysis and sensor technology usually rely on expensive equipment, such as multi-spectral or hyperspectral imaging equipment, environmental sensors, etc. These devices are not only costly, but also complex to maintain and update, making them difficult to be widely used; lack of real-time performance: many existing technologies, such as machine learning methods based on image processing, have slow processing speeds when facing large-scale data or complex environments, making it difficult to achieve real-time monitoring. This makes it impossible for farmers to take timely measures in the early stages of a disease outbreak. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a plant disease detection method based on deep learning, which solves the problems of high cost and lack of real-time performance of existing plant disease detection.

[0005] In order to achieve the above-mentioned purpose of the invention, the technical solution adopted by the present invention is: a plant disease detection method based on deep learning, comprising:

[0006] S1: constructing a plant disease detection model using a feature extraction layer, a neck network layer and a head network layer; wherein the feature extraction layer includes a full-dimensional dynamic convolution module, the neck network layer includes an efficient multi-scale attention mechanism, and the head network layer includes a dynamic target detection module;

[0007] S2: training the plant disease detection model using a mixed loss function to obtain a trained plant disease detection model; wherein the mixed loss function includes a bounding box loss, a classification loss, and a distributed loss;

[0008] S3: Using the trained plant disease detection model to process the plant leaf data set to obtain plant disease detection results, thereby completing the detection of plant diseases.

[0009] The beneficial effects of the present invention are as follows: a plant disease detection model is constructed by using a feature extraction layer including a full-dimensional dynamic convolution module, a neck network layer including an efficient multi-scale attention mechanism, and a head network layer including a dynamic target detection module, and a plant leaf data set is processed to obtain a plant disease detection result. (1) By constructing a plant disease detection model, the labor cost and equipment cost are greatly reduced; (2) By introducing a full-dimensional dynamic convolution module and a dynamic target detection module, the real-time detection is improved; (3) By introducing a dynamic target detection module and an efficient multi-scale attention mechanism, the detection accuracy and model complexity are balanced, the calculation amount is reduced, and the calculation accuracy is improved; (4) In this way, the efficient detection capability can be maintained to ensure wide application under different lighting conditions and complex farmland environments.

[0010] Furthermore, the full-dimensional dynamic convolution module is used to:

[0011] The input feature vector is compressed by global average pooling to obtain a compressed vector;

[0012] Using the fully connected unit and the rectified linear unit, the compressed vector is low-dimensionally mapped and negative values ​​are eliminated to obtain the attention feature vector:

[0013]

[0014] Among them, y represents the attention feature vector, α wi represents the spatial dimension attention weight, α fi represents the output channel attention weight, α ci represents the input channel attention weight, α si represents the convolution kernel attention weight, W i represents the i-th convolution kernel, x represents the input feature vector, i represents the convolution kernel index, n represents the total number of convolution kernels, ⊙ represents element-by-element multiplication, and * represents the convolution operation.

[0015] The ODConv convolution module allows the model to dynamically adjust the size of the convolution kernel to adapt to the complexity of different disease characteristics, greatly improving the flexibility and effectiveness of the feature extraction stage. This module effectively captures the key features of leaf diseases, reduces information loss, and ensures efficient recognition of the model in complex backgrounds.

[0016] Furthermore, the dynamic target detection module includes an attention mechanism having a scale perception unit, a space perception unit and a task perception unit:

[0017] The scale perception is used to dynamically fuse the semantic importance of different scales to obtain the semantic fusion features after scale perception:

[0018]

[0019] Among them, π L It represents the semantic fusion feature after scale perception, F represents the input feature map, σ represents the hard Sigmoid function, f represents the function acting on the feature map, which is used to extract features, S represents the fusion of features of different scales, and C represents the number of channels, which is usually the channel dimension size of the input feature F;

[0020] The spatial perception is used to analyze the shape, rotation and position information of the object in space to obtain spatial features;

[0021]

[0022] Among them, π S Indicates that spatial features are obtained through spatial perception, L indicates the total number of scales, l indicates the current scale index, K indicates the number of convolution kernels, and w l,k represents the weight of the kth convolution kernel at the lth scale, p k represents the discrimination area, Δp k represents the spatial offset of self-learning, c represents the feature channel, Δm k Represents the weight of each position;

[0023] The task perception is used to adjust the feature representation according to different task requirements to obtain task adjustment features:

[0024] π C (F)·F=max(α 1 (F)·F c +β 1 (F),α 2 (F)·F c +β 2 (F));

[0025] Among them, π C represents the task adjustment feature, max represents the maximum value function, α 1 ,α 2 represents the feature weighting functions corresponding to task 1 and task 2 respectively, β 1 , β 2 represents the bias term of the corresponding task, F c Represents the feature slice of the c-th channel.

[0026] By integrating the attention mechanism of scale perception, space perception and task perception, the detection accuracy of disease targets of different scales is significantly improved, especially the ability to identify small target diseases.

[0027] Furthermore, the bounding box loss is used to measure the difference between the predicted bounding box and the true bounding box, and its specific expression is:

[0028]

[0029] Among them, L CIoU represents the bounding box loss result, IoU represents the intersection over union ratio between the predicted box and the true box, ρ represents the Euclidean distance between the predicted box and the center point of the true box, b represents the predicted box, b g represents the real box, c represents the diagonal length of the minimum closure of the predicted box and the real box, α represents the penalty term related to the aspect ratio of the predicted box, and v represents the penalty term related to the aspect ratio of the real box.

[0030] Furthermore, the classification loss is used to measure the difference between the predicted category and the true category, and its specific expression is:

[0031]

[0032] Among them, L cls Represents the classification loss result, y i represents the actual category label, p i It represents the probability of the model predicting the category, i represents the actual category number, and C represents the number of categories.

[0033] Furthermore, the distributed loss is used to predict discrete probability distribution, and its specific expression is:

[0034]

[0035] Among them, L DFL Represents the distributed loss result, y k represents the actual discretized value, p k represents the probability predicted by the model, k represents the number of categories after discretization, and K represents the total number of categories after discretization. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] This specification will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0037] Figure 1 This is an exemplary flow chart of a plant disease detection method based on deep learning according to some embodiments of this specification. DETAILED DESCRIPTION

[0038] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0039] Example

[0040] Figure 1 is an exemplary flow chart of a plant disease detection method based on deep learning according to some embodiments of this specification. Figure 1 As shown, the process includes the following steps. In some embodiments, the process can be executed by a processor.

[0041] S1: Construct a plant disease detection model using a feature extraction layer, a neck network layer and a head network layer; wherein the feature extraction layer includes a full-dimensional dynamic convolution module, the neck network layer includes an efficient multi-scale attention mechanism, and the head network layer includes a dynamic target detection module.

[0042] The plant disease detection model is a neural network model for detecting plant diseases. The types of plant disease detection models can be various. For example, the types of plant disease detection models can include graph neural network models and YOLOv8 models, etc.

[0043] In some embodiments, the input of the plant disease detection model may be a plant leaf dataset, and the output of the plant disease detection model may be a plant disease detection result.

[0044] In some embodiments, the structure of the plant disease detection model is as follows:

[0045] The plant disease detection model includes a feature extraction layer, a neck network layer and a head network layer. The output of the feature extraction layer is used as the input of the neck network layer, the output of the neck network layer is used as the input of the head network layer, and the output of the head network layer is used as the final output of the plant disease detection model.

[0046] The feature extraction layer is used to extract multiple attention feature vectors. The input of the feature extraction layer may include a plant leaf dataset, and the output may include an attention feature vector.

[0047] The plant leaf dataset is a dataset constructed from plant leaf images. For example, the plant leaf dataset may include tomato leaf images, corn leaf images, and rice leaf images.

[0048] In some embodiments, the processor can use a camera and a drone to acquire corresponding plant leaf images and construct a plant leaf dataset.

[0049] The input feature vector is a feature vector reflecting the correlation between plant leaves and plant diseases.

[0050] In some embodiments, the feature extraction layer may perform preliminary feature extraction on the plant leaf dataset to obtain an input feature vector.

[0051] The attention feature vector is a vector constructed from multiple types of attention scalars. For example, the attention feature vector may include a barrier attention scalar.

[0052] In some embodiments, the feature extraction layer may include a full-dimensional dynamic convolution module.

[0053] The full-dimensional dynamic (ODConv) convolution module is used to: use the full-dimensional dynamic convolution module to compress the input feature vector through global average pooling to obtain a compressed vector; use the fully connected unit and the rectified linear unit to perform low-dimensional mapping and negative value elimination on the compressed vector to obtain the attention feature vector.

[0054] The compressed vector is the feature vector after length compression.

[0055] In some embodiments, the full-dimensional dynamic convolution module can perform global average pooling on the input feature vector to compress the length to C in , and get the compressed vector.

[0056] In some embodiments, the feature extraction layer may utilize a fully connected (FC) unit to map the compressed vector to a low-dimensional space with a reduction ratio of 1 / 16.

[0057] In this way, the complexity of the dynamic convolution module can be reduced and the detection efficiency can be improved.

[0058] In some embodiments, the feature extraction layer can use a rectified linear unit (ReLU) activation function to eliminate negative values ​​in the feature vector, construct four independent branches corresponding to the four types of attention scalars in the full-dimensional dynamic convolution module, and obtain an attention feature vector.

[0059] In some embodiments, the expression of the attention feature vector may be:

[0060]

[0061] Among them, y represents the attention feature vector, α wi represents the spatial dimension attention weight, α fi represents the output channel attention weight, α ci represents the input channel attention weight, αsi represents the convolution kernel attention weight, W i represents the i-th convolution kernel, x represents the input feature vector, i represents the convolution kernel index, n represents the total number of convolution kernels, ⊙ represents element-by-element multiplication, and * represents the convolution operation.

[0062] The neck network layer is used to fuse the attention feature vector. The input of the neck network layer may include the attention feature vector, and the output may include the fused feature vector.

[0063] The fused feature vector is the feature fusion of multiple attention feature vectors.

[0064] In some embodiments, the neck network layer can utilize efficient multi-scale attention (EMA attention mechanism) to fuse the attention feature vectors by calculating the feature vector path to obtain a fused feature vector.

[0065] In some embodiments, the neck network layer can use two parallel sub-networks to extract features from the attention feature vector to obtain a multi-scale feature vector; wherein the parallel sub-networks can be a 1x1 convolution branch and a 3x3 convolution branch.

[0066] The 1x1 convolution branch can be used to process local cross-channel information and capture the detailed features of a specific channel.

[0067] The 3x3 convolution branch can be used to expand the receptive field, capture more local spatial features, and enhance the model's perception of objects of different scales.

[0068] Through these two parallel branches, EMA can efficiently extract multi-scale features while reducing the depth of the network and avoiding excessive computational costs.

[0069] In some embodiments, the neck network layer can divide the multi-scale feature vector into multiple sub-feature groups along the channel dimension; wherein each sub-feature group contains different semantic information and is processed independently.

[0070] In this way, redundant calculations between channels can be effectively reduced while retaining more detailed information.

[0071] In some embodiments, the neck network layer can aggregate the global and local features of the sub-feature group through a cross-space learning method, by performing a global average pooling operation in the spatial dimension, extracting global context information, and calculating the spatial attention weight on the output of each sub-network to obtain an aggregated multi-scale feature vector.

[0072] In this way, attention to a specific spatial location can be enhanced.

[0073] In some embodiments, the neck network layer can also utilize a cross-dimensional interaction mechanism to further enhance the global representation capability of the aggregated features.

[0074] In this way, we ensure that the model not only focuses on local details but also enhances the understanding of the target in the global context.

[0075] In some embodiments, the neck network layer can perform cross-space learning aggregation on the output of the 1x1 convolution branch and the output of the 3x3 convolution branch to generate a feature map enhanced by global context, and superimpose these features together to form the final output to obtain a fused feature vector.

[0076] In this way, the spatial information of the input can be retained while improving the semantic understanding of specific targets; through multi-scale and cross-space learning methods, the feature representation ability of the model is significantly improved, especially while reducing computational overhead while still capturing global and local details. This makes EMA perform well in target detection.

[0077] Using efficient multi-scale attention can optimize the calculation path, reduce redundant operations, and thus reduce the computational complexity. In addition, the EMA attention mechanism divides the channel degree into multiple sub-feature groups, achieving effective interaction between different dimensions, thereby enhancing the understanding and representation of multi-dimensional features, thereby maximizing the efficiency of feature utilization, making the detection results more accurate and stable.

[0078] The head network layer is used to analyze the fused feature vector to obtain the plant disease detection result. The input of the head network layer may include the fused feature vector, and the output may include the plant disease detection result.

[0079] The plant disease detection results are the detection results that reflect the disease conditions of plant leaves.

[0080] In some embodiments, the head network layer can use the dynamic target detection module to analyze the fused feature vector to obtain scale perception results, space perception results and task perception results respectively; and obtain plant disease detection results by integrating the scale perception results, space perception results and task perception results.

[0081] The ODConv convolution module allows the model to dynamically adjust the size of the convolution kernel to adapt to the complexity of different disease characteristics, greatly improving the flexibility and effectiveness of the feature extraction stage. This module effectively captures the key features of leaf diseases, reduces information loss, and ensures efficient recognition of the model in complex backgrounds.

[0082] In some embodiments, the head network layer may include an attention mechanism having a scale-aware unit, a space-aware unit, and a task-aware unit; wherein the scale-aware unit is used to dynamically fuse the semantic importance of different scales to obtain semantic fusion features after scale perception; the space-aware unit is used to analyze the shape, rotation, and position information of the object in space to obtain spatial features; and the task-aware unit is used to adjust the feature representation according to different task requirements to obtain task-adjusted features.

[0083] The scale-aware semantic fusion features are features that reflect the comprehensive semantic information in the feature vector.

[0084] In some embodiments, the expression of the scale-aware semantic fusion feature can be:

[0085]

[0086] Among them, π L It represents the semantic fusion feature after scale perception, F represents the input feature map, σ represents the hard Sigmoid function, f represents the function acting on the feature map, which is used to extract features, S represents the fusion of features of different scales, and C represents the number of channels, which is usually the channel dimension size of the input feature F.

[0087] Spatial features are features that reflect the shape, rotation, and position of an object in space.

[0088] In some embodiments, the expression of the spatial feature may be:

[0089]

[0090] Among them, π S Indicates that spatial features are obtained through spatial perception, L indicates the total number of scales, l indicates the current scale index, K indicates the number of convolution kernels, and w l,k represents the weight of the kth convolution kernel at the lth scale, p k represents the discrimination area, Δp k represents the spatial offset of self-learning, c represents the feature channel, Δm k Represents the weight of each position.

[0091] Task adjustment characteristics are characteristics that reflect different task requirements.

[0092] In some embodiments, the expression of the task adjustment feature may be:

[0093] π C (F)·F=max(α 1 (F)·F c +β 1 (F),α 2(F)·F c +β 2 (F));

[0094] Among them, π C represents the task adjustment feature, max represents the maximum value function, α 1 ,α 2 represents the feature weighting functions corresponding to task 1 and task 2 respectively, β 1 , β 2 represents the bias term of the corresponding task, F c Represents the feature slice of the c-th channel.

[0095] By integrating the attention mechanism of scale perception, space perception and task perception, the detection accuracy of disease targets of different scales is significantly improved, especially the ability to identify small target diseases.

[0096] S2: Using a mixed loss function to train the plant disease detection model to obtain a trained plant disease detection model; wherein the mixed loss function includes a bounding box loss, a classification loss, and a distributed loss.

[0097] In some embodiments, the plant disease detection model can be obtained by training multiple labeled training samples. For example, multiple labeled training samples can be input into the initial plant disease detection model, and a loss function can be constructed by using the labels and the results of the initial plant disease detection model. The parameters of the initial plant disease detection model can be iteratively updated based on the loss function by gradient descent or other methods. When the preset conditions are met, the model training is completed and a trained plant disease detection model is obtained. The preset conditions can be that the loss function converges, the number of iterations reaches a threshold, etc.

[0098] In some embodiments, the training samples include historical plant leaf datasets. The labels are corresponding plant disease detection results. The labels can be manually annotated.

[0099] In some embodiments, a bounding box loss is used to measure the difference between the predicted bounding box and the true bounding box.

[0100] In some embodiments, the expression of the bounding box loss may be:

[0101]

[0102] Among them, L CIoU represents the bounding box loss result, IoU represents the intersection over union ratio between the predicted box and the true box, ρ represents the Euclidean distance between the predicted box and the center point of the true box, b represents the predicted box, b grepresents the real box, c represents the diagonal length of the minimum closure of the predicted box and the real box, α represents the penalty term related to the aspect ratio of the predicted box, and v represents the penalty term related to the aspect ratio of the real box.

[0103] In some embodiments, the classification loss is used to measure the difference between the predicted class and the true class.

[0104] In some embodiments, the expression of classification loss can be:

[0105]

[0106] Among them, L cls Represents the classification loss result, y i represents the actual category label, p i It represents the probability of the model predicting the category, i represents the actual category number, and C1 represents the number of categories.

[0107] In some embodiments, the distributed loss is used to predict a discrete probability distribution.

[0108] In some embodiments, the expression of the distributed loss may be:

[0109]

[0110] Among them, L DFL Represents the distributed loss result, y k represents the actual discretized value, p k represents the probability predicted by the model, k represents the number of categories after discretization, and K represents the total number of categories after discretization.

[0111] S3: Use the trained plant disease detection model to process the plant leaf data set to obtain scale perception results, space perception results and task perception results respectively; obtain the plant disease detection results by integrating the scale perception results, space perception results and task perception results to complete the detection of plant diseases.

[0112] In some embodiments of the present specification, a plant disease detection model is constructed using a feature extraction layer including a full-dimensional dynamic convolution module, a neck network layer including an efficient multi-scale attention mechanism, and a head network layer including a dynamic target detection module, and a plant leaf data set is processed to obtain plant disease detection results. (1) By constructing a plant disease detection model, labor costs and equipment costs are greatly reduced; (2) By introducing a full-dimensional dynamic convolution module and a dynamic target detection module, the real-time detection is improved; (3) By introducing a dynamic target detection module and an efficient multi-scale attention mechanism, the detection accuracy and model complexity are balanced, the amount of calculation is reduced, and the calculation accuracy is improved; (4) In this way, efficient detection capabilities can be maintained to ensure wide application under different lighting conditions and complex farmland environments.

Claims

1. A plant disease detection method based on deep learning, characterized in that: include: S1: constructing a plant disease detection model using a feature extraction layer, a neck network layer and a head network layer; wherein the feature extraction layer includes a full-dimensional dynamic convolution module, the neck network layer includes an efficient multi-scale attention mechanism, and the head network layer includes a dynamic target detection module; S2: training the plant disease detection model using a mixed loss function to obtain a trained plant disease detection model; wherein the mixed loss function includes a bounding box loss, a classification loss, and a distributed loss; S3: Use the trained plant disease detection model to process the plant leaf data set to obtain scale perception results, space perception results and task perception results respectively; obtain the plant disease detection results by integrating the scale perception results, space perception results and task perception results to complete the detection of plant diseases.

2. The plant disease detection method based on deep learning according to claim 1, characterized in that: The full-dimensional dynamic convolution module is used to: The input feature vector is compressed by global average pooling to obtain a compressed vector; Using the fully connected unit and the rectified linear unit, the compressed vector is low-dimensionally mapped and negative values ​​are eliminated to obtain the attention feature vector: Among them, y represents the attention feature vector, α wi represents the spatial dimension attention weight, α fi represents the output channel attention weight, α ci represents the input channel attention weight, α si represents the convolution kernel attention weight, W i represents the i-th convolution kernel, x represents the input feature vector, i represents the convolution kernel index, n represents the total number of convolution kernels, ⊙ represents element-by-element multiplication, and * represents the convolution operation.

3. The plant disease detection method based on deep learning according to claim 1, characterized in that: The dynamic target detection module includes an attention mechanism having a scale-aware unit, a space-aware unit, and a task-aware unit: The scale perception is used to dynamically fuse the semantic importance of different scales to obtain the semantic fusion features after scale perception: Among them, π L It represents the semantic fusion feature after scale perception, F represents the input feature map, σ represents the hard Sigmoid function, f represents the function acting on the feature map, which is used to extract features, S represents the fusion of features of different scales, and C represents the number of channels, which is usually the channel dimension size of the input feature F; The spatial perception is used to analyze the shape, rotation and position information of the object in space to obtain spatial features; Among them, π S Indicates that spatial features are obtained through spatial perception, L indicates the total number of scales, l indicates the current scale index, K indicates the number of convolution kernels, and w l,k represents the weight of the kth convolution kernel at the lth scale, p k represents the discrimination area, Δp k represents the spatial offset of self-learning, c represents the feature channel, Δm k Represents the weight of each position; The task perception is used to adjust the feature representation according to different task requirements to obtain task adjustment features: p C (F)·F=max(α 1 (F)·F c +b 1 (F),a 2 (F)·F c +b 2 (F)); Among them, π C represents the task adjustment feature, max represents the maximum value function, α 1 ,α 2 represents the feature weighting functions corresponding to task 1 and task 2 respectively, β 1 , β 2 represents the bias term of the corresponding task, F c Represents the feature slice of the c-th channel.

4. The plant disease detection method based on deep learning according to claim 1, characterized in that: The bounding box loss is used to measure the difference between the predicted bounding box and the true bounding box, and its specific expression is: Among them, L CIoU represents the bounding box loss result, IoU represents the intersection over union ratio between the predicted box and the true box, ρ represents the Euclidean distance between the predicted box and the center point of the true box, b represents the predicted box, b g represents the real box, c represents the diagonal length of the minimum closure of the predicted box and the real box, α represents the penalty term related to the aspect ratio of the predicted box, and v represents the penalty term related to the aspect ratio of the real box.

5. The plant disease detection method based on deep learning according to claim 1, characterized in that: The classification loss is used to measure the difference between the predicted category and the true category, and its specific expression is: Among them, L cls Represents the classification loss result, y i represents the actual category label, p i It represents the probability of the model predicting the category, i represents the actual category number, and C1 represents the number of categories.

6. The plant disease detection method based on deep learning according to claim 1, characterized in that: The distributed loss is used to predict discrete probability distribution, and its specific expression is: Among them, L DFL Represents the distributed loss result, y k represents the actual discretized value, p k represents the probability predicted by the model, k represents the number of categories after discretization, and K represents the total number of categories after discretization.

Images