Sugarcane disease recognition method based on fusion differential mechanism and double-gated coordinate attention

Abstract

This invention provides a sugarcane disease identification method based on a fusion difference mechanism and dual-gated coordinate attention. By constructing a sugarcane disease identification network with a difference extraction module, a collaborative feature fusion module, and a dual-gated coordinate attention module, the method effectively suppresses background redundancy in sugarcane images and enhances discriminative lesion features through a dual difference mechanism in both spatial and channel dimensions. The collaborative feature fusion module bridges semantic differences between shallow and deep features through multi-scale feature alignment and dual-branch complementary processing, improving the performance and robustness of the entire identification network. The dual-gated coordinate attention module, combined with dual-path parallel coding and a gated adaptive fusion mechanism, enhances the network's ability to capture lesion details, thereby effectively improving the accuracy of sugarcane disease identification and providing accurate data for subsequent sugarcane disease prevention and control.

Description

Technical Field

[0001] This invention relates to a plant disease identification method, and more particularly to a sugarcane disease identification method based on a fusion difference mechanism and dual-gated coordinate attention. Background Technology

[0002] Sugarcane is an important sugar crop. Throughout its growth period, sugarcane is often susceptible to various diseases such as rust, brown stripe, and smut, which can easily lead to a decline in yield and quality, resulting in economic losses. However, because the early symptoms of sugarcane are subtle and insidious, disease detection is often delayed. This delay often leads to the loss of the best window for prevention and control, and even to blind or excessive application of pesticides, posing a potential threat to the quality of agricultural products and the ecological environment.

[0003] In existing technologies, machine learning methods such as support vector machines, random forests, and convolutional neural networks are used to identify sugarcane diseases. However, these methods have the following drawbacks: First, the complex field background makes the features of lesions, which have weak discriminative power, easily interfered with by redundant information in the background image, resulting in inaccurate feature extraction. Second, sugarcane diseases exhibit diverse morphologies and significant scale differences, with small lesions coexisting with large diseased areas, thus demonstrating multi-scale features. Existing methods cannot accurately extract these multi-scale features. Finally, the edges of lesions and areas of abrupt texture changes are important discriminative regions for distinguishing sugarcane diseases. However, these key details are easily weakened during feature extraction and aggregation in existing methods, thus affecting the accurate identification of diseases.

[0004] Therefore, in order to solve the above-mentioned technical problems, it is urgent to propose a new technical approach. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a sugarcane disease identification method based on a fusion difference mechanism and dual-gated coordinate attention. By constructing a sugarcane disease identification network with a difference extraction module, a collaborative feature fusion module, and a dual-gated coordinate attention module, the method effectively suppresses background redundancy in sugarcane images and enhances discriminative lesion features through a dual difference mechanism in both spatial and channel dimensions. The collaborative feature fusion module bridges semantic differences between deep and shallow features through multi-scale feature alignment and dual-branch complementary processing, improving the performance and robustness of the entire identification network. The dual-gated coordinate attention module, combined with dual-path parallel coding and a gated adaptive fusion mechanism, enhances the network's ability to capture lesion details, thereby effectively improving the accuracy of sugarcane disease identification and providing accurate basis for subsequent sugarcane disease prevention and control.

[0006] This invention provides a sugarcane disease identification method based on a fusion difference mechanism and dual-gated coordinate attention, comprising the following steps:

[0007] S1. Obtain images of sugarcane disease samples and preprocess the images;

[0008] S2. Construct a sugarcane disease identification network based on fusion difference mechanism and dual-gated coordinate attention, and input the preprocessed sample images into the sugarcane disease identification network for training;

[0009] The sugarcane disease identification network includes an initial feature extraction module, a difference extraction module, a collaborative feature fusion module, a dual-gated coordinate attention module, a feedforward network, and a classifier.

[0010] The initial feature extraction module is used to extract features from the input sample image and input the output features into the collaborative feature fusion module for processing. The output features Fe of the collaborative feature fusion module are input into the dual-gated coordinate attention module for processing. The output features of the dual-gated coordinate attention module and the output features of the collaborative feature fusion module are added element-wise to form feature F1 and input into the feedforward network. The output features of the feedforward network and feature F1 are added element-wise to form feature F2 and input into the classifier. The classifier processes feature F2 and outputs the recognition result.

[0011] S3. Acquire real-time sugarcane images, preprocess the sugarcane images, and input them into the trained sugarcane disease recognition network to obtain the current sugarcane disease recognition results.

[0012] Furthermore, the differential extraction module includes four sequentially connected sub-modules with identical structures;

[0013] Each submodule includes a downsampling module I and a differential extraction unit. The input terminal of the downsampling module I serves as the input terminal of the submodule, and the output features of the downsampling module I are input into the differential extraction unit. The output terminal of the differential extraction unit serves as the output terminal of the submodule.

[0014] In the feature flow direction: the first submodule outputs feature X1, the second submodule outputs feature X2, the third submodule outputs feature X3, and the fourth submodule outputs feature X4;

[0015] Features X2, X3, and X4 are all input into the collaborative feature fusion module;

[0016] The scale of feature X1 is The scale of feature X2 is The scale of feature X3 is The scale of feature X1 is Where H, C, and W represent the height, number of channels, and width of the feature map, respectively.

[0017] Furthermore, the differential feature extraction unit includes a spatial mixing unit and a channel mixing unit. The input of the spatial mixing unit is the output feature of the downsampling module I. The output feature M of the spatial mixing unit is input into the channel mixing unit. The output of the channel mixing unit serves as the output of the differential feature extraction unit.

[0018] Furthermore, the spatial mixing unit includes a first convolution module, a second convolution module, a third convolution module, a global average pooling module I, a SiLU activation function module I, a first subtraction module, and a first depthwise convolution module;

[0019] The first convolutional module serves as the input of the spatial mixing unit. The output of the first convolutional module is connected to the input of the global average pooling module I. The output features of the global average pooling module I are input into the SiLU activation function module. The output features of both the SiLU activation function module I and the first convolutional module are input into the first subtraction module. The first subtraction module performs element-wise subtraction on the output features of the SiLU activation function module I and the first convolutional module. The output features of the first subtraction module are multiplied element-wise by the adaptive adjustment factor α to form feature M1. Feature M1 is added element-wise by the input features of the spatial mixing unit and then input into the second convolutional module. The output features of the second convolutional module are input into the first depth convolutional module. The output features of the first depth convolutional module are input into the third convolutional module. The output features of the third convolutional module are multiplied element-wise by the hierarchical scaling factor γ1 to form feature M2. Feature M2 is added element-wise by the input features of the spatial mixing unit to form feature M and then output to the channel mixing unit.

[0020] The scale of the first to the third convolutional modules is 1×1.

[0021] Furthermore, the channel mixing unit includes a fourth convolution module, a fifth convolution module, a sixth convolution module, a second depthwise convolution module, a second subtraction module, and a SiLU activation function module II;

[0022] The fourth convolutional module inputs feature M, and its output features are input to the second deep convolutional module. The output features of the second deep convolutional module are input to the fifth convolutional module, and their output features are input to the SiLU activation function module II. The output features of both the SiLU activation function module II and the second deep convolutional module are input to the second subtraction module. The second subtraction module performs element-wise subtraction on the output features of the SiLU activation function module II and the second deep convolutional module. The output features of the second subtraction module are multiplied element-wise by the adaptive adjustment factor β to form feature Y1. Feature Y1 is added element-wise by the output features of the second deep convolutional module and then input to the sixth convolutional module. The output features of the sixth convolutional module are multiplied element-wise by the layer scaling factor γ2 to form feature Y2. Feature Y2 is added element-wise by feature M and then used as the output feature of the channel mixing unit.

[0023] Among them, the scale of the third to sixth convolutional modules is 1×1 convolutional modules.

[0024] Furthermore, the collaborative feature fusion module includes a downsampling module II, a concatenation module I, a feature fusion module, a seventh convolution module, an eighth convolution module, a ninth convolution module, a tenth convolution module, an eleventh convolution module, a twelfth convolution module, a concatenation module II, a global average pooling module II, a multi-scale attention module, and a Sigmoid function module I;

[0025] Features X2 and X3 are both input to downsampling module II. The output features of downsampling module II and feature X4 are both input to concatenation module I for concatenation. The output features of concatenation module I are input to feature fusion module. The output features of feature fusion module are input to the seventh and eighth convolution modules respectively. The output features of the seventh convolution module are input to the ninth convolution module. The output features of the eighth convolution module are input to the tenth convolution module. The output features of the ninth and tenth convolution modules are both input to concatenation module II for concatenation. The output features of concatenation module II are input to the eleventh convolution module. The output features of the eleventh convolution module are input to global average pooling module II. The output features of global average pooling module II are input to multi-scale attention module. The output features of multi-scale attention module are input to Sigmoid function module I. The output features of Sigmoid function module I and the output features of the eleventh convolution module are multiplied element-wise and then input to the twelfth convolution module. The output features of the twelfth convolution module are the output features Fe of the collaborative feature fusion module.

[0026] The seventh, eleventh, and twelfth convolutional modules all have a scale of 1×1; the ninth and eighth convolutional modules both have a scale of 3×3; and the tenth convolutional module has a scale of 5×5.

[0027] Furthermore, the feature fusion module includes convolution module I and convolution module II;

[0028] The input of convolution module I serves as the input of the feature fusion module, the output features of convolution module I are input into convolution module II, and the output of convolution module II serves as the output of the feature fusion module.

[0029] Among them, the scale of convolution module II is 3×3, and the scale of convolution module I is 1×1.

[0030] Furthermore, the dual-gated coordinate attention module includes a width branch network and a height branch network, which have the same structure. The feature Fe is input to the height branch network and the width branch network, respectively, and the output features of the width branch network and the height branch network are Gw and Gh, respectively.

[0031] The width branch network includes an average pooling module, a max pooling module, a splicing module III, a convolution module III, a convolution module IV, a convolution module V, a convolution module VI, a Sigmoid function module II, a batch normalization module, a SiLU activation function module III, and a Sigmoid function module III.

[0032] Both the average pooling module and the max pooling module input features Fe. The output features of both the average pooling module and the max pooling module are input into the splicing module III for splicing operations. The output features of the splicing module III are input into the convolution module III and the convolution module IV, respectively. The output features of the convolution module IV are input into the Sigmoid function module II. The output features of the Sigmoid function module II are multiplied element-wise with the output features of the convolution module III and then input into the convolution module V. The output features of the convolution module V are input into the batch normalization module. The output features of the batch normalization module are input into the SiLU activation function module III. The output features of the SiLU activation function module III are input into the convolution module VI. The output features of the convolution module VI are input into the Sigmoid function module III. The output of the Sigmoid function module III serves as the output of the width branch network.

[0033] The features Gw, Gh, and Fe are multiplied element-wise and then used as the output features of the dual-gated coordinate attention module.

[0034] The beneficial effects of this invention are as follows: By constructing a sugarcane disease identification network with a differential extraction module, a collaborative feature fusion module, and a dual-gated coordinate attention module, the background redundancy of sugarcane images is effectively suppressed and the discriminative lesion features are enhanced through a dual differential mechanism of spatial and channel dimensions. The collaborative feature fusion module bridges the semantic differences between deep and shallow features through multi-scale feature alignment and dual-branch complementary processing, thereby improving the performance and robustness of the entire identification network. The dual-gated coordinate attention module, combined with dual-path parallel coding and a gated adaptive fusion mechanism, enhances the network's ability to capture lesion details, thus effectively improving the accuracy of sugarcane disease identification and providing an accurate basis for subsequent sugarcane disease prevention and control. Attached Figure Description

[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0036] Figure 1 This is a schematic diagram of the sugarcane disease identification network structure of the present invention.

[0037] Figure 2 This is a schematic diagram of the differential extraction module structure of the present invention.

[0038] Figure 3 This is a schematic diagram of the spatial hybrid unit structure of the present invention.

[0039] Figure 4 This is a schematic diagram of the channel mixing unit structure of the present invention.

[0040] Figure 5 This is a schematic diagram of the dual-gated coordinate attention module structure of the present invention.

[0041] Figure 6 This is a schematic diagram of the width branch network structure of the present invention.

[0042] Figure 7 This is a schematic diagram of the collaborative feature fusion module structure of the present invention. Detailed Implementation

[0043] The present invention will be further described in detail below:

[0044] This invention provides a sugarcane disease identification method based on a fusion difference mechanism and dual-gated coordinate attention, comprising the following steps:

[0045] S1. Obtain images of sugarcane disease samples and preprocess the images;

[0046] S2. Construct a sugarcane disease identification network based on fusion difference mechanism and dual-gated coordinate attention, and input the preprocessed sample images into the sugarcane disease identification network for training;

[0047] The sugarcane disease identification network includes an initial feature extraction module, a difference extraction module, a collaborative feature fusion module, a dual-gated coordinate attention module, a feedforward network, and a classifier.

[0048] The initial feature extraction module is used to extract features from the input sample image and input the output features into the collaborative feature fusion module for processing. The output features Fe of the collaborative feature fusion module are input into the dual-gated coordinate attention module for processing. The output features of the dual-gated coordinate attention module and the output features of the collaborative feature fusion module are added element-wise to form feature F1 and input into the feedforward network. The output features of the feedforward network and feature F1 are added element-wise to form feature F2 and input into the classifier. The classifier processes feature F2 and outputs the recognition result.

[0049] The initial feature extraction module consists of two 1×1 convolutional modules connected in series, which are used to extract features. The extracted features are then output to the difference extraction module for further processing. The feedforward network is an existing technology, and its structure generally includes an input layer, a hidden layer, and an output layer connected in series. The classifier is implemented using existing technology, such as the Softmax classifier.

[0050] S3. Acquire real-time sugarcane images, preprocess them, and input them into the trained sugarcane disease recognition network to obtain the current sugarcane disease recognition result. Preprocessing of sample images and real-time images, such as denoising and cropping, is existing technology and will not be elaborated here. In this invention, a sugarcane disease recognition network with a differential extraction module, a collaborative feature fusion module, and a dual-gated coordinate attention module is constructed. Through a dual differential mechanism in spatial and channel dimensions, background redundancy in sugarcane images is effectively suppressed, and discriminative lesion features are enhanced. The collaborative feature fusion module, through multi-scale feature alignment and dual-branch complementary processing, bridges the semantic differences between deep and shallow features, improving the performance and robustness of the entire recognition network. The dual-gated coordinate attention module, combined with dual-path parallel coding and gated adaptive fusion mechanism, enhances the network's ability to capture lesion details, thereby effectively improving the accuracy of sugarcane disease recognition and providing accurate basis for subsequent sugarcane disease prevention and control.

[0051] In this embodiment, the differential extraction module includes four sequentially connected sub-modules with the same structure;

[0052] Each submodule includes a downsampling module I and a differential extraction unit. The input terminal of the downsampling module I serves as the input terminal of the submodule, and the output features of the downsampling module I are input into the differential extraction unit. The output terminal of the differential extraction unit serves as the output terminal of the submodule.

[0053] In the feature flow direction: the first submodule outputs feature X1, the second submodule outputs feature X2, the third submodule outputs feature X3, and the fourth submodule outputs feature X4;

[0054] Features X2, X3, and X4 are all input into the collaborative feature fusion module;

[0055] The scale of feature X1 is The scale of feature X2 is The scale of feature X3 is The scale of feature X1 is Where H, C, and W represent the height, number of channels, and width of the feature map, respectively. For example... Figure 2 As shown, each submodule includes a downsampling module I, each with a different scale, thereby sampling to form the scales of features X1, X2, X3, and X4. These are then processed by a difference extraction unit to form the final features X1, X2, X3, and X4. Specifically:

[0056] The differential feature extraction unit includes a spatial mixing unit and a channel mixing unit. The output features of the input downsampling module I are used to the spatial mixing unit, and the output features M of the spatial mixing unit are input to the channel mixing unit. The output of the channel mixing unit serves as the output of the differential feature extraction unit.

[0057] The spatial mixing unit includes a first convolution module, a second convolution module, a third convolution module, a global average pooling module I, a SiLU activation function module I, a first subtraction module, and a first depthwise convolution module;

[0058] The first convolutional module serves as the input of the spatial mixing unit. The output of the first convolutional module is connected to the input of the global average pooling module I. The output features of the global average pooling module I are input into the SiLU activation function module. The output features of both the SiLU activation function module I and the first convolutional module are input into the first subtraction module. The first subtraction module performs element-wise subtraction on the output features of the SiLU activation function module I and the first convolutional module. The output features of the first subtraction module are multiplied element-wise by the adaptive adjustment factor α to form feature M1. Feature M1 is added element-wise by the input features of the spatial mixing unit and then input into the second convolutional module. The output features of the second convolutional module are input into the first depth convolutional module. The output features of the first depth convolutional module are input into the third convolutional module. The output features of the third convolutional module are multiplied element-wise by the hierarchical scaling factor γ1 to form feature M2. Feature M2 is added element-wise by the input features of the spatial mixing unit to form feature M and then output to the channel mixing unit.

[0059] The scale of the first to the third convolutional modules is 1×1.

[0060] The channel mixing unit includes a fourth convolution module, a fifth convolution module, a sixth convolution module, a second depthwise convolution module, a second subtraction module, and a SiLU activation function module II;

[0061] The fourth convolutional module inputs feature M, and its output features are input to the second deep convolutional module. The output features of the second deep convolutional module are input to the fifth convolutional module, and their output features are input to the SiLU activation function module II. The output features of both the SiLU activation function module II and the second deep convolutional module are input to the second subtraction module. The second subtraction module performs element-wise subtraction on the output features of the SiLU activation function module II and the second deep convolutional module. The output features of the second subtraction module are multiplied element-wise by the adaptive adjustment factor β to form feature Y1. Feature Y1 is added element-wise by the output features of the second deep convolutional module and then input to the sixth convolutional module. The output features of the sixth convolutional module are multiplied element-wise by the layer scaling factor γ2 to form feature Y2. Feature Y2 is added element-wise by feature M to become the output feature Y of the channel mixing unit. This feature corresponds to X1, X2, X3, and X4 in the first to fourth sub-modules, respectively.

[0062] In this invention, the scale of each convolutional module from the third to the sixth is 1×1. The scale of each depthwise convolutional module is set according to actual conditions, generally 3×3. Through the differential extraction module of this invention, the phase difference between local responses and global background commonalities can be extracted, and a learnable factor is used to superimpose this as a residual gain onto the original features, thereby explicitly suppressing background redundancy information and enhancing discriminative lesion features.

[0063] In this embodiment, the collaborative feature fusion module includes a downsampling module II, a splicing module I, a feature fusion module, a seventh convolution module, an eighth convolution module, a ninth convolution module, a tenth convolution module, an eleventh convolution module, a twelfth convolution module, a splicing module II, a global average pooling module II, a multi-scale attention module, and a Sigmoid function module I.

[0064] Features X2 and X3 are both input to downsampling module II. The output features of downsampling module II and feature X4 are both input to concatenation module I for concatenation. The output features of concatenation module I are input to feature fusion module. The output features of feature fusion module are input to the seventh and eighth convolution modules respectively. The output features of the seventh convolution module are input to the ninth convolution module. The output features of the eighth convolution module are input to the tenth convolution module. The output features of the ninth and tenth convolution modules are both input to concatenation module II for concatenation. The output features of concatenation module II are input to the eleventh convolution module. The output features of the eleventh convolution module are input to global average pooling module II. The output features of global average pooling module II are input to multi-scale attention module. The output features of multi-scale attention module are input to Sigmoid function module I. The output features of Sigmoid function module I and the output features of the eleventh convolution module are multiplied element-wise and then input to the twelfth convolution module. The output features of the twelfth convolution module are the output features Fe of the collaborative feature fusion module.

[0065] Among them, the scale of the seventh, eleventh and twelfth convolutional modules is 1×1; the scale of the ninth and eighth convolutional modules is 3×3; and the scale of the tenth convolutional module is 5×5.

[0066] The feature fusion module includes convolution module I and convolution module II;

[0067] The input of convolution module I serves as the input of the feature fusion module, the output features of convolution module I are input into convolution module II, and the output of convolution module II serves as the output of the feature fusion module.

[0068] In this model, convolutional module II has a scale of 3×3, while convolutional module I has a scale of 1×1. Although the differential extraction module can extract hierarchical features with significant heterogeneity, and its shallow features retain rich spatial details, its semantic abstraction ability is limited. Its deep features have strong semantic discriminative power, but low spatial resolution. By using the collaborative feature fusion module, these complementary features can be effectively integrated, which helps to bridge the semantic gap between shallow and deep layers, improves the network's ability to represent diseased areas, and avoids introducing redundant background noise.

[0069] In this embodiment, the dual-gated coordinate attention module includes a width branch network and a height branch network. The width branch network and the height branch network have the same structure. The feature Fe is input to the height branch network and the width branch network respectively. The output features of the width branch network and the height branch network are Gw and Gh respectively.

[0070] The width branch network includes an average pooling module, a max pooling module, a splicing module III, a convolution module III, a convolution module IV, a convolution module V, a convolution module VI, a Sigmoid function module II, a batch normalization module, a SiLU activation function module III, and a Sigmoid function module III.

[0071] Both the average pooling module and the max pooling module input feature Fe. The output features of both modules are input into the splicing module III for splicing. The output features of the splicing module III are input into convolution module III and convolution module IV, respectively. The output features of convolution module IV are input into sigmoid function module II. The output features of sigmoid function module II are multiplied element-wise with the output features of convolution module III and then input into convolution module V. The output features of convolution module V are input into the batch normalization module. The output features of the batch normalization module are input into SiLU activation function module III. The output features of SiLU activation function module III are input into convolution module VI. The output features of convolution module VI are input into Sigmoid function module III. The output of Sigmoid function module III serves as the output of the width branch network. The elements-wise multiplication of features Gw, Gh, and Fe serves as the output features of the dual-gated coordinate attention module. Since the height branch network and the width branch network are structurally identical, except that the height branch network focuses on the image height feature information and the width branch network focuses on the image width feature information, this implementation example only provides the structure of the width branch network, and the height branch network can be built according to this structure.

[0072] The average pooling module extracts stable global contextual representations by aggregating local features, while the max pooling module focuses on preserving details such as lesion edges and abrupt textures by responding to significant local changes. Then, the modules are fused together by the combined action of the stitching module III, convolution module III, convolution module IV, and sigmoid function module II, dynamically adjusting the ratio of high-frequency texture details to low-frequency background information.

[0073] The dual-gated coordinate attention module combines dual-path parallel encoding with a gated adaptive fusion mechanism to enhance the network's ability to capture lesion details. This avoids filtering out local details such as lesion edges and texture abrupt changes while suppressing noise, thereby preventing the loss of discriminative features and effectively improving the accuracy of sugarcane disease identification.

[0074] In the attached diagram, ⊙ represents element-wise multiplication, and ⊕ represents element-wise addition.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying sugarcane diseases based on a fusion difference mechanism and dual-gated coordinate attention, characterized in that: Includes the following steps: S1. Obtain images of sugarcane disease samples and preprocess the images; S2. Construct a sugarcane disease identification network based on fusion difference mechanism and dual-gated coordinate attention, and input the preprocessed sample images into the sugarcane disease identification network for training; The sugarcane disease identification network includes an initial feature extraction module, a difference extraction module, a collaborative feature fusion module, a dual-gated coordinate attention module, a feedforward network, and a classifier. The initial feature extraction module is used to extract features from the input sample image and input the output features into the collaborative feature fusion module for processing. The output features Fe of the collaborative feature fusion module are input into the dual-gated coordinate attention module for processing. The output features of the dual-gated coordinate attention module and the output features of the collaborative feature fusion module are added element-wise to form feature F1 and input into the feedforward network. The output features of the feedforward network and feature F1 are added element-wise to form feature F2 and input into the classifier. The classifier processes feature F2 and outputs the recognition result. S3. Acquire real-time sugarcane images, preprocess the sugarcane images, and input them into the trained sugarcane disease recognition network to obtain the current sugarcane disease recognition results.

2. The sugarcane disease identification method based on fusion difference mechanism and dual-gated coordinate attention as described in claim 1, characterized in that: The differential extraction module includes four sequentially connected sub-modules with identical structures; Each submodule includes a downsampling module I and a differential extraction unit. The input terminal of the downsampling module I serves as the input terminal of the submodule, and the output features of the downsampling module I are input into the differential extraction unit. The output terminal of the differential extraction unit serves as the output terminal of the submodule. In the feature flow direction: the first submodule outputs feature X1, the second submodule outputs feature X2, the third submodule outputs feature X3, and the fourth submodule outputs feature X4; Features X2, X3, and X4 are all input into the collaborative feature fusion module; The scale of feature X1 is The scale of feature X2 is The scale of feature X3 is The scale of feature X1 is Where H, C, and W represent the height, number of channels, and width of the feature map, respectively.

3. The sugarcane disease identification method based on fusion difference mechanism and dual-gated coordinate attention as described in claim 2, characterized in that: The differential feature extraction unit includes a spatial mixing unit and a channel mixing unit. The output features of the input downsampling module I are used to the spatial mixing unit, and the output features M of the spatial mixing unit are input to the channel mixing unit. The output of the channel mixing unit serves as the output of the differential feature extraction unit.

4. The sugarcane disease identification method based on fusion difference mechanism and dual-gated coordinate attention as described in claim 3, characterized in that: The spatial mixing unit includes a first convolution module, a second convolution module, a third convolution module, a global average pooling module I, a SiLU activation function module I, a first subtraction module, and a first depthwise convolution module; The first convolutional module serves as the input of the spatial mixing unit. The output of the first convolutional module is connected to the input of the global average pooling module I. The output features of the global average pooling module I are input into the SiLU activation function module. The output features of both the SiLU activation function module I and the first convolutional module are input into the first subtraction module. The first subtraction module performs element-wise subtraction on the output features of the SiLU activation function module I and the first convolutional module. The output features of the first subtraction module are multiplied element-wise by the adaptive adjustment factor α to form feature M1. Feature M1 is added element-wise by the input features of the spatial mixing unit and then input into the second convolutional module. The output features of the second convolutional module are input into the first depth convolutional module. The output features of the first depth convolutional module are input into the third convolutional module. The output features of the third convolutional module are multiplied element-wise by the hierarchical scaling factor γ1 to form feature M2. Feature M2 is added element-wise by the input features of the spatial mixing unit to form feature M and then output to the channel mixing unit. The scale of the first to the third convolutional modules is 1×1.

5. The sugarcane disease identification method based on fusion difference mechanism and dual-gated coordinate attention according to claim 3, characterized in that: The channel mixing unit includes a fourth convolution module, a fifth convolution module, a sixth convolution module, a second depthwise convolution module, a second subtraction module, and a SiLU activation function module II; The fourth convolutional module inputs feature M, and its output features are input to the second deep convolutional module. The output features of the second deep convolutional module are input to the fifth convolutional module, and their output features are input to the SiLU activation function module II. The output features of both the SiLU activation function module II and the second deep convolutional module are input to the second subtraction module. The second subtraction module performs element-wise subtraction on the output features of the SiLU activation function module II and the second deep convolutional module. The output features of the second subtraction module are multiplied element-wise by the adaptive adjustment factor β to form feature Y1. Feature Y1 is added element-wise by the output features of the second deep convolutional module and then input to the sixth convolutional module. The output features of the sixth convolutional module are multiplied element-wise by the layer scaling factor γ2 to form feature Y2. Feature Y2 is added element-wise by feature M and then used as the output feature of the channel mixing unit. Among them, the scale of the third to sixth convolutional modules is 1×1 convolutional modules.

6. The sugarcane disease identification method based on fusion difference mechanism and dual-gated coordinate attention as described in claim 2, characterized in that: The collaborative feature fusion module includes a downsampling module II, a concatenation module I, a feature fusion module, a seventh convolution module, an eighth convolution module, a ninth convolution module, a tenth convolution module, an eleventh convolution module, a twelfth convolution module, a concatenation module II, a global average pooling module II, a multi-scale attention module, and a Sigmoid function module I. Features X2 and X3 are both input into downsampling module II. The output features of downsampling module II and feature X4 are both input into concatenation module I for concatenation. The output features of concatenation module I are input into feature fusion module. The output features of feature fusion module are input into the seventh and eighth convolution modules respectively. The output features of the seventh convolution module are input into the ninth convolution module. The output features of the eighth convolution module are input into the tenth convolution module. The output features of the ninth and tenth convolution modules are both input into concatenation module II for concatenation. The output features of concatenation module II are input into the eleventh convolution module. The output features of the eleventh convolution module are input into global average pooling module II. The output features of global average pooling module II are input into multi-scale attention module. The output features of multi-scale attention module are input into Sigmoid function module I. The output features of Sigmoid function module I and the output features of the eleventh convolution module are multiplied element-wise and then input into the twelfth convolution module. The output features of the twelfth convolution module are the output features Fe of the collaborative feature fusion module. The seventh, eleventh, and twelfth convolutional modules all have a scale of 1×1; the ninth and eighth convolutional modules both have a scale of 3×3; and the tenth convolutional module has a scale of 5×5.

7. The sugarcane disease identification method based on fusion difference mechanism and dual-gated coordinate attention as described in claim 6, characterized in that: The feature fusion module includes convolution module I and convolution module II; The input of convolution module I serves as the input of the feature fusion module, the output features of convolution module I are input into convolution module II, and the output of convolution module II serves as the output of the feature fusion module. Among them, the scale of convolution module II is 3×3, and the scale of convolution module I is 1×1.

8. The sugarcane disease identification method based on fusion difference mechanism and dual-gated coordinate attention as described in claim 1, characterized in that: The dual-gated coordinate attention module includes a width branch network and a height branch network. The width branch network and the height branch network have the same structure. The feature Fe is input to the height branch network and the width branch network respectively. The output features of the width branch network and the height branch network are Gw and Gh respectively. The width branch network includes an average pooling module, a max pooling module, a splicing module III, a convolution module III, a convolution module IV, a convolution module V, a convolution module VI, a Sigmoid function module II, a batch normalization module, a SiLU activation function module III, and a Sigmoid function module III. Both the average pooling module and the max pooling module input features Fe. The output features of both the average pooling module and the max pooling module are input into the splicing module III for splicing operations. The output features of the splicing module III are input into the convolution module III and the convolution module IV, respectively. The output features of the convolution module IV are input into the Sigmoid function module II. The output features of the Sigmoid function module II are multiplied element-wise with the output features of the convolution module III and then input into the convolution module V. The output features of the convolution module V are input into the batch normalization module. The output features of the batch normalization module are input into the SiLU activation function module III. The output features of the SiLU activation function module III are input into the convolution module VI. The output features of the convolution module VI are input into the Sigmoid function module III. The output of the Sigmoid function module III serves as the output of the width branch network. The features Gw, Gh, and Fe are multiplied element-wise and then used as the output features of the dual-gated coordinate attention module.

Images