Cotton similar disease detection method and device based on high and low frequency and attention fusion

Abstract

The invention discloses a cotton similar disease detection method and device based on high and low frequency and attention fusion, and the method comprises the steps: constructing a cotton leaf disease data set, collecting the cotton leaf disease data, and carrying out the preprocessing of the collected cotton leaf disease data, dividing a cotton leaf disease data set after preprocessing, and writing a data set label file; constructing a cotton similar disease intelligent detection model based on high and low frequency feature enhancement and attention guidance fusion, wherein the model comprises a high and low frequency feature adaptive enhancement module and an attention guidance fusion module; and disease monitoring is performed on agricultural cotton based on the intelligent detection model for similar diseases of cotton. The device comprises a processor and a memory. The method solves the problems that the existing algorithm has obvious defects in the aspect of small-scale lesion area feature representation, texture details and morphological features of the lesion area are difficult to effectively extract, and the detection precision is limited.

Description

Technical Field

[0001] The present invention relates to the field of agricultural cotton disease monitoring, and in particular to a cotton similar disease detection method and device based on high-low frequency and attention fusion. Background Art

[0002] Plant diseases are one of the key factors restricting healthy plant growth, and efficient and accurate detection technologies have become a core research focus in smart agriculture. Among existing disease detection technologies, computer vision-based methods, with their significant advantages such as non-contact, rapidity, and accuracy, have gradually become a research hotspot in the field of agricultural informatization. Deep learning-based computer vision detection technology has achieved breakthroughs. By constructing a deep convolutional neural network architecture, it can automatically extract the multi-level spatial features of plant disease spots and learn complex high-level features, significantly improving the accuracy of plant disease identification and detection.

[0003] This end-to-end feature learning mechanism not only breaks through the reliance of image processing technology on manually extracted feature engineering, but its deep network structure can also effectively capture the nonlinear relationship between disease phenotypes and pathological characteristics, providing technical support for the construction of an intelligent plant health monitoring system.

[0004] Although current research has made significant progress in the field of intelligent cotton disease detection, existing algorithms have significant deficiencies in characterizing small-scale lesion regions, making it difficult to effectively extract texture details and morphological features of lesion regions, resulting in limited detection accuracy.

[0005] In addition, the interference factors brought by the complex field environment exacerbate the difficulty of detection. The soil particles in the field background are small in size and densely distributed. Their spatial distribution characteristics form spectral domain aliasing with the morphological characteristics of the spreading edge of the lesion. This space-frequency domain coupling interference leads to the image signal-to-noise ratio, affecting the model's ability to identify effective disease symptoms. Summary of the Invention

[0006] The present invention provides a method and device for detecting similar cotton diseases based on high- and low-frequency and attention fusion. This invention solves the problem that existing algorithms have obvious deficiencies in characterizing the characteristics of small-scale lesion areas, making it difficult to effectively extract the texture details and morphological features of the lesion areas, resulting in limited detection accuracy. Detailed description is provided below:

[0007] In the first aspect, a method for detecting similar cotton diseases based on high-low frequency and attention fusion is provided, the method comprising:

[0008] Construct a cotton leaf disease dataset, collect cotton leaf disease data, and preprocess the collected cotton leaf disease data. After preprocessing, divide the cotton leaf disease dataset and write a dataset label file;

[0009] An intelligent cotton disease detection model based on high- and low-frequency feature enhancement and attention guidance fusion is constructed, the model comprising: a high- and low-frequency feature adaptive enhancement module and an attention guidance fusion module;

[0010] Agricultural cotton disease monitoring is carried out based on the intelligent detection model of similar cotton diseases.

[0011] Among them, in the high- and low-frequency feature adaptive enhancement module, the low-frequency features are enhanced by average pooling and multi-scale edge enhancement modules, and the high-frequency features are processed by the CMUNeXt block to obtain high-frequency features and low-frequency features respectively;

[0012] High-frequency features and low-frequency features are concatenated to obtain high- and low-frequency features. After fusion, the high- and low-frequency features are further extracted and enhanced through convolutional layers and attention mechanisms.

[0013] The extracted and enhanced features are added to the original input to achieve self-enhancement of high- and low-frequency features.

[0014] Among them, the attention-guided fusion module is used to effectively fuse feature maps from different levels and sources, and the Pixel Attention module weights the key areas in the image in the spatial dimension, so that the model focuses on the target area.

[0015] Among them, the high- and low-frequency feature adaptive enhancement module uses deep convolution of large convolution kernels, reverse bottleneck design and residual connection to efficiently extract these high-frequency features while being lightweight, and enhances the details of high-frequency information through a dense connection mechanism.

[0016] Furthermore, the high-frequency features and the low-frequency features are spliced ​​together to obtain high- and low-frequency features, and the high- and low-frequency features are fused as follows:

[0017] After concatenating the high-frequency features and the low-frequency features in the channel dimension, 1×1 convolution is used to compress the number of channels, and the high- and low-frequency features are fused. The channel attention mechanism CALayer is introduced to adjust the feature weights. Finally, 3x3 convolution is used to refine the fused features. The residual connection is used to add the fused features to the input x to retain the original information.

[0018] Among them, the attention-guided fusion module uses the PixelAttention attention module to calculate the final weighted coefficient of the cross-view feature interaction module, and weightedly fuses the left and right feature maps according to the coefficient to output the final fusion feature.

[0019] In the second aspect, a cotton similar disease detection device based on high-low frequency and attention fusion, the device includes: a processor and a memory, the memory stores program instructions, and the processor calls the program instructions stored in the memory to enable the device to execute any one of the methods described in the first aspect.

[0020] According to a third aspect, a computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes any one of the methods described in the first aspect.

[0021] The beneficial effects of the technical solution provided by the present invention are:

[0022] 1. This paper provides a high-low frequency feature enhancement module (HLAEM), which significantly improves the resolution accuracy of low-frequency boundary features while retaining high-frequency detail features by dynamically adjusting the convolution kernel expansion rate. HLAEM is applied to C3k2 to obtain the improved C3k2_HLAEM, which significantly improves C3k2's ability to capture multi-scale features by adaptively enhancing the receptive field. Secondly, an attention-guided fusion module (EAGFM) is provided to effectively suppress information redundancy while enhancing feature interaction.

[0023] 2. This paper proposes a high- and low-frequency feature adaptive enhancement module (HLAEM), which integrates a multi-branch parallel architecture, a serial skip-layer connection structure, and an attention mechanism to build an optimized system for multi-scale feature fusion. Through a ternary synergistic mechanism of multi-receptive field feature interaction, cross-level feature reuse, and adaptive weight allocation, it establishes a high-frequency feature enhancement channel while improving the refinement of low-frequency features, effectively improving the accuracy of key boundary positioning.

[0024] 3. The present invention proposes an Effective Attention Guided Fusion Module (EAGFM), which establishes a cross-modal feature interaction channel through a pixel-level attention mechanism; adopts a learnable dynamic weight allocation mechanism, uses Pixel Attention to model the spatial correlation of multi-source features, and realizes adaptive feature fusion through pixel-by-pixel feature importance evaluation, effectively enhancing the model's ability to distinguish target areas from non-target areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of the cotton similar disease detection method based on high-low frequency and attention fusion;

[0026] Figure 2 This is a schematic diagram of the overall structure of the cotton similar disease detection method based on high-low frequency and attention fusion;

[0027] Figure 3 Schematic diagram of HLAEM;

[0028] Figure 4 Schematic diagram of feature fusion method;

[0029] Figure 5 It is a schematic diagram of EAGFM;

[0030] Figure 6 Schematic diagram of a cotton similar disease detection method based on high-low frequency and attention fusion provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are described in further detail below.

[0032] Example 1

[0033] A cotton similar disease detection method based on high-low frequency and attention fusion, see Figure 1 and Figure 2 , the method comprising:

[0034] Step 101: Construct a cotton leaf disease dataset, collect cotton leaf disease data, preprocess the collected cotton leaf disease data, divide the cotton leaf disease dataset after preprocessing, and compile a dataset label file;

[0035] This step specifically includes:

[0036] Step (1.1): Collect image data of anthracnose, brown spot, target spot, and verticillium wilt; the data sources come from actual field photography, public datasets, and agricultural research databases;

[0037] Step (1.2): Labeling the cotton leaf disease image data; using agricultural experts to manually label the collected cotton leaf disease data, accurately labeling the collected image data; the labeling content includes: disease type and disease severity;

[0038] Step (1.3): Use data augmentation technology to expand the cotton leaf disease image data; use image rotation, flipping, blurring, adding noise, and changing brightness to augment the collected cotton leaf disease data to expand the dataset and increase the diversity and quantity of samples;

[0039] Step (1.4): Standardize the image, crop and resize the cotton leaf disease image;

[0040] Step (1.5): Classify the cotton leaf disease categories, save the cotton leaf disease data into folders of their respective disease categories according to the cotton leaf disease categories, and divide the cotton leaf disease dataset into a training set, a test set, and a validation set; randomly allocate the cotton leaf disease data obtained in steps (1.1) to (1.4) according to a ratio of 8:1:1;

[0041] Step (1.6): Write a dataset label file, which includes the dataset category to which the image belongs, the cotton leaf disease category to which the image belongs, and the cotton leaf disease image data path.

[0042] Step 102: Construct an intelligent cotton disease detection model based on the fusion of high- and low-frequency feature enhancement and attention-guided learning. First, a high- and low-frequency feature adaptive enhancement module (HLAEM) was designed. In this module, low-frequency features are enhanced through average pooling and the multi-scale edge enhancement module (MEEM), while high-frequency features are processed by the CMUNeXt block. After processing by the MEEM and CMUNeXt blocks, high- and low-frequency features are obtained, respectively. High- and low-frequency features are concatenated to obtain high- and low-frequency features. After fusion, these features are further extracted and enhanced through convolutional layers and an attention mechanism. Finally, the processed features are added to the original input to achieve self-enhancement of high- and low-frequency features. This mechanism not only enhances high- and low-frequency information in the image but also, through the introduction of an attention mechanism, increases the weight of key features, significantly enhancing the model's ability to learn features from different frequency bands. Next, the C3k2 module in the model architecture was improved, replacing it with C3k2_HLAEM, which integrates the HLAEM module. This design significantly improves C3k2's multi-scale feature extraction capabilities, enhancing its ability to recover detailed information and its understanding of global structure, thereby improving overall model performance. Finally, an efficient attention-guided fusion module (EAGFM) was designed to effectively fuse feature maps from different levels and sources. The PixelAttention module weights key areas in the image spatially, enabling the model to automatically focus on the target area, reduce background interference, and optimize the fusion of target features.

[0043] like Figure 2As shown in the figure, starting with the input image, the network sequentially passes through three main parts: the backbone network, the neck, and the head. In the backbone network, the input image first passes through two convolutional layers for preliminary feature extraction. It then passes through multiple C3K2 modules, such as C3K2(256, False, 0.25), which incorporate operations such as convolution and activation functions to further enhance feature extraction. In the middle and later stages of the backbone network, C3k2 is replaced with C3k2_HLAEM modules, which may have structural improvements for better feature processing. Finally, the network passes through modules such as convolution, C3k2_HLAEM, SPPF, and C2PSA. SPPF is used for multi-scale feature extraction, and C2PSA is used for feature aggregation. Entering the neck, the network primarily involves multiple feature fusion and processing operations. A splicing operation combines feature maps from different layers, combining features from the backbone network. Upsampling aligns the size of the feature maps with those of the shallow layers for easier splicing. Convolutional layers then further process the spliced ​​feature maps. The head part mainly includes an efficient attention-guided fusion module (EAGFM) and a detection head module. The efficient attention-guided fusion module (such as EAGFM[3,16], EAGFM[6,20], EAGFM[10,24]) is used to perform final processing and transformation of features in preparation for detection output. Finally, the detection head module converts the processed features into the final detection results, including the target category and location information.

[0044] This step specifically includes:

[0045] Step (2.1): The High-Low Frequency Feature Adaptive Enhancement Module (HLAEM) separates high and low frequencies through downsampling and interpolation, enhancing the features of different frequency bands respectively. The Multi-Scale Edge Enhancement Module (MEEM) and CMUNeXt Block (CMUNeXtBlock) are used to process high and low frequency information in a targeted manner, integrating the channel attention mechanism to achieve the purpose of dynamically adjusting features. Finally, the residual output is used to retain the original information and stabilize the training process; thereby achieving enhanced edge perception at each scale, highlighting the key boundaries of objects, and improving the refinement of low-frequency features;

[0046] Figure 3The path (a) in the figure processes high-frequency information, which typically reflects local details in an image and is crucial for various visual tasks. Therefore, through deep convolution with large convolutional kernels, a reverse bottleneck design, and residual connections, these high-frequency features are efficiently extracted while maintaining lightweightness. A dense connection mechanism is then used to enhance the details of high-frequency information. First, high-frequency details are separated through high = x - upsampling (low). Then, a CMUNeXt block is introduced into the high-frequency path. After processing by the CMUNeXt block, highfeat is obtained. The CMUNeXt block is the core module of the CMUNeXt network, focusing on efficiently extracting global contextual feature information while maintaining lightweightness and computational efficiency. The CMUNeXt block uses large convolution kernels and depthwise convolution to overcome the local receptive field limitations of ordinary convolution, enhance the network's perception of long-distance features, and adapt to the multi-scale and complex structures in medical images; uses depthwise separable convolution to decompose the convolution into intra-channel convolution and point convolution, significantly reducing the number of parameters and computational cost; uses reverse bottlenecks to expand the dimensions of the intermediate feature layer to achieve more comprehensive spatial and channel feature fusion; and ensures stability during model training and improves inference speed through residual connections and batch normalization.

[0047] Figure 3 The path b in the figure processes low-frequency information. Low-frequency information typically represents the overall structure of an image and is primarily used to recover large-scale image information, such as outlines and background. This addresses the problem of missing detailed edge features and improves the model's ability to capture object boundaries. First, 2x2 average pooling downsampling is used to extract low-frequency components. Then, a multi-scale edge enhancement module (MEEM) is introduced within the low-frequency processing structure. This module addresses the problem of missing detailed edges in object detection and improves the model's ability to capture object boundaries, thereby enhancing edge perception and highlighting key object boundaries. It uses average pooling and convolution to extract multi-scale edge information from the input image. The edge enhancer (EE) strengthens edge perception at each scale and highlights key object boundaries. The extracted multi-scale edge information is fused with the features of the main branch to improve the precision of the final prediction. Finally, the low-frequency features (lowf) are restored to the original image size to obtain lowfeat.

[0048] After concatenating the highfeat and lowfeat in the channel dimension, a 1×1 convolution is used to compress the number of channels and fuse the high- and low-frequency features. A channel attention mechanism, CALayer, is then introduced to adjust feature weights. Finally, a 3×3 convolution is used to further refine the fused features. Finally, a residual connection is used to add the fused features to the input x, preserving the original information.

[0049] Common feature fusion methods, such as Figure 4As shown in the figure. Additive fusion (a) directly adds elements, which is computationally simple but prone to feature loss; weighted addition (b) introduces learnable weights and requires balancing parameter complexity; gated fusion (c) uses a dynamic feature selection mechanism and has the highest computational complexity; multiplicative fusion (d) is sensitive to feature scale and has poor stability; splicing fusion (e) preserves the original features by splicing along the dimension, combining computational efficiency and scale adaptability. When designing the High-Low-Frequency Feature Adaptive Enhancement Module (HLAEM), after comprehensive consideration of high- and low-frequency fusion, the final choice of splicing fusion is to achieve a more comprehensive and information-rich feature representation, thereby improving the model's performance in detection.

[0050] Step (2.2): We propose an effective attention-guided fusion module (EAGFM) to enhance information sharing between feature maps from different sources. Through an attention-based weighted fusion method, the model can adaptively select the most useful features and avoid the introduction of redundant information. EAGFM is an attention-based feature interaction module that can effectively fuse feature maps from two different sources, enabling the model to better capture the relationship between them, thereby improving the performance of detection tasks.

[0051] Figure 5 The center box shows the Cross-View Feature Interaction Module (CVIM). EAGFM uses the Pixel Attention module to calculate the final weighting coefficient of the Cross-View Feature Interaction Module (CVIM). It then weights and fuses the left and right feature maps based on this coefficient, outputting the final fused features. The left and right feature maps interact through the attention mechanism. This allows the model to capture the underlying relationships and important information between the two feature maps, especially when complementary information exists between them. By calculating the attention coefficient, the module is able to assign different weights to each input feature map. The model automatically learns which input feature map is more important for the current task and increases its weight accordingly. This makes the feature fusion process more intelligent and can adapt to the characteristics of the data in different situations.

[0052] Step 103: Set evaluation indicators and experimental parameters, specifically:

[0053] This step specifically includes:

[0054] Step (3.1): Set evaluation indicators: use precision, recall, and F1 score as evaluation indicators to evaluate the performance of the model;

[0055] Step (3.2): Set the experimental parameters: set the batch size to 16, the initial value of the learning rate to 1e-5, the optimizer algorithm to SGD, the number of iterations to 300, and the learning rate decay rate to 0.0005.

[0056] Step 104: Design a comparative experiment to test the comprehensive performance advantages of the method proposed in the present invention.

[0057] This step specifically includes:

[0058] Step (4.1): Compare with other feature fusion strategy models, including additive fusion, weighted additive fusion, gated fusion, multiplicative fusion, and splicing fusion, to demonstrate the optimization effect of the proposed method.

[0059] Step (4.2): Ablation experiments were performed on the modules that make up high and low frequency processing respectively with HLAEM to demonstrate the good performance of HLAEM.

[0060] Step (4.3): Conduct comparative experiments on seven mainstream attention modules, including Pixel Attention, CA, ACmix, CBAM, ECA, Triplet Attention, and SE attention mechanisms, to verify the effectiveness of the attention mechanism in the EAGFM module. Step (4.4): Conduct ablation experiments on the modules used by EGAFM based on its performance in detecting similar diseases on cotton leaves.

[0061] Step (4.5): To systematically evaluate the performance of the model improvement strategy in detecting similar diseases on cotton leaves, ablation experiments were performed on the EGAFM and C3k2_HLAEM modules.

[0062] In summary, the embodiment of the present invention proposes a cotton similar disease detection method based on high-low frequency and attention fusion based on the existing target detection model YOLOv11 through the above steps (1.1) to (3.2), combined with the proposed high-low frequency feature adaptive enhancement module (HLAEM) and effective attention guidance fusion module (EAGFM). The present invention has achieved the best results on both self-built datasets and public cotton disease datasets.

[0063] Example 2

[0064] The feasibility of the solution in Example 1 is verified by combining specific experimental data and calculation formulas, as described below:

[0065] Step 4.1: Compare with other feature fusion strategy models, including additive fusion, weighted additive fusion, gated fusion, multiplicative fusion, and splicing fusion, to demonstrate the optimization effect of the proposed method.

[0066]

[0067] Step 4.2: Ablation experiments were performed on the modules that make up high and low frequency processing respectively with HLAEM to demonstrate the good performance of HLAEM.

[0068]

[0069]

[0070] Step 4.3: Conduct comparative experiments on seven mainstream attention modules, including: Pixel Attention, CA, ACmix, CBAM, ECA, Triplet Attention, and SE attention mechanism, to verify the effectiveness of the attention mechanism in the EAGFM module.

[0071]

[0072] Step 4.4: Based on the performance of EGAFM in detecting similar diseases on cotton leaves, an ablation experiment was conducted on its usage modules.

[0073]

[0074] Step 4.5: To systematically evaluate the effectiveness of the model improvement strategy in detecting similar diseases on cotton leaves, ablation experiments were performed on the EGAFM and C3k2_HLAEM modules.

[0075]

[0076] Example 3

[0077] In order to evaluate the overall performance of the model, the evaluation indicators used in this experiment are average precision, precision, recall rate, and F1-Score. Average precision, average precision and F1-Score indicators consider both precision and recall rate. The above indicators are defined as:

[0078] Precision = TP / (TP+FP)

[0079] Recall = TP / (TP+FN)

[0080]

[0081] F1 score=(2*Precision*Recall) / (Precision+Recall)

[0082] TP indicates that no true positive defects were detected, FP indicates a false positive defect, and FN indicates a false negative defect. Precision is the accuracy rate, which indicates the proportion of true positive samples among predicted positive samples. Recall is the recall rate, which indicates the proportion of true positive samples that were correctly identified. AP is the average precision. The precision-recall curve can reflect the relationship between the precision and recall of the classifier at different thresholds. F1-Score is a metric used to comprehensively evaluate the performance of classification models. It is the harmonic mean of precision and recall, and aims to achieve a balance between the two.

[0083] To illustrate the accuracy of the method, a comprehensive comparison was conducted on the public datasets A Training Dataset for Deep NeuralNetwork Model Recognition of Common Cotton Diseases, abbreviated as CCD (3453), SAR-CLD-2024: A Comprehensive Dataset for Cotton Leaf Disease Detection, abbreviated as SAR-CLD (2137), and our laboratory dataset Four Cotton Diseases (3751). The performance test results are shown in Table 1. On the self-built dataset, 11 types of detection models (SSD, Faster-RCNN, Retinanet, DETR, YOLOv5n, YOLOv6, YOLOv8n, YOLOv9t, YOLOv10n, YOLOv11n, and YOLOv12n) covering both dual-stage and single-stage detection were selected for horizontal comparison.

[0084] The results show that the proposed model performs well on the three datasets. In addition, the average detection accuracy and F1-score of the model on different datasets are improved, indicating that EHDM-YOLOV11n has good generalization ability on different datasets.

[0085] Table 1

[0086]

[0087]

[0088] Table 2

[0089]

[0090] A cotton similar disease detection device based on high-low frequency and attention fusion, the device comprising: a processor and a memory, wherein program instructions are stored in the memory, and the processor calls the program instructions stored in the memory to enable the device to execute the following method steps in Example 1:

[0091] Construct a cotton leaf disease dataset, collect cotton leaf disease data, and preprocess the collected cotton leaf disease data. After preprocessing, divide the cotton leaf disease dataset and write a dataset label file;

[0092] An intelligent cotton disease detection model based on high- and low-frequency feature enhancement and attention guidance fusion is constructed, the model comprising: a high- and low-frequency feature adaptive enhancement module and an attention guidance fusion module;

[0093] Agricultural cotton disease monitoring is carried out based on the intelligent detection model of similar cotton diseases.

[0094] In the high- and low-frequency feature adaptive enhancement module, low-frequency features are enhanced by average pooling and multi-scale edge enhancement modules, while high-frequency features are processed by the CMUNeXt block to obtain high-frequency features and low-frequency features respectively.

[0095] High-frequency features and low-frequency features are concatenated to obtain high- and low-frequency features. After fusion, the high- and low-frequency features are further extracted and enhanced through convolutional layers and attention mechanisms.

[0096] The extracted and enhanced features are added to the original input to achieve self-enhancement of high- and low-frequency features.

[0097] Among them, the attention-guided fusion module is used to effectively fuse feature maps from different levels and sources, and the Pixel Attention module weights the key areas in the image in the spatial dimension, so that the model focuses on the target area.

[0098] Among them, the high- and low-frequency feature adaptive enhancement module uses deep convolution of large convolution kernels, reverse bottleneck design and residual connection to efficiently extract these high-frequency features while being lightweight, and enhances the details of high-frequency information through a dense connection mechanism.

[0099] Furthermore, the high-frequency features and low-frequency features are spliced ​​together to obtain high- and low-frequency features, and the high- and low-frequency features are fused as follows:

[0100] After concatenating the high-frequency features and the low-frequency features in the channel dimension, 1×1 convolution is used to compress the number of channels, and the high- and low-frequency features are fused. The channel attention mechanism CALayer is introduced to adjust the feature weights. Finally, 3x3 convolution is used to refine the fused features. The residual connection is used to add the fused features to the input x to retain the original information.

[0101] Among them, the attention-guided fusion module uses the PixelAttention module to calculate the final weighted coefficient of the cross-view feature interaction module, and weightedly fuses the left and right feature maps according to the coefficient to output the final fusion feature.

[0102] It should be noted here that the device description in the above embodiment corresponds to the method description in the embodiment, and the embodiment of the present invention will not be described in detail here.

[0103] The execution subjects of the above-mentioned processor and memory can be computers, single-chip microcomputers, microcontrollers and other devices with computing functions. In specific implementation, the embodiment of the present invention does not limit the execution subject and it can be selected according to the needs of actual application.

[0104] Data signals are transmitted between the memory and the processor via a bus, which will not be described in detail in the embodiment of the present invention.

[0105] Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium, which includes a stored program, and when the program is running, controls the device where the storage medium is located to execute the method steps in the above embodiment.

[0106] The computer-readable storage medium includes but is not limited to a flash memory, a hard disk, a solid-state drive, and the like.

[0107] It should be noted here that the description of the readable storage medium in the above embodiment corresponds to the description of the method in the embodiment, and the embodiment of the present invention will not be described in detail here.

[0108] In the above embodiments, all or part of the embodiments may be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions according to the embodiments of the present invention are generated in whole or in part.

[0109] The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in or transmitted via computer-readable storage media. Computer-readable storage media can be any available medium that can be accessed by a computer, or a data storage device such as a server or data center that integrates one or more available media. Available media can include magnetic media or semiconductor media, etc.

[0110] Those skilled in the art will understand that the accompanying drawings are only a schematic diagram of a preferred embodiment, and the serial numbers of the embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A cotton similar disease detection method based on high-low frequency and attention fusion, characterized in that: The method comprises: Construct a cotton leaf disease dataset, collect cotton leaf disease data, and preprocess the collected cotton leaf disease data. After preprocessing, divide the cotton leaf disease dataset and write a dataset label file; An intelligent cotton disease detection model based on high- and low-frequency feature enhancement and attention guidance fusion is constructed, the model comprising: a high- and low-frequency feature adaptive enhancement module and an attention guidance fusion module; Agricultural cotton disease monitoring is carried out based on the intelligent detection model of similar cotton diseases.

2. The cotton similar disease detection method based on high-low frequency and attention fusion according to claim 1 is characterized in that: In the high- and low-frequency feature adaptive enhancement module, the low-frequency features are enhanced by the average pooling and multi-scale edge enhancement modules, and the high-frequency features are processed by the CMUNeXt block to obtain high-frequency features and low-frequency features respectively; High-frequency features and low-frequency features are concatenated to obtain high- and low-frequency features. After fusion, the high- and low-frequency features are further extracted and enhanced through convolutional layers and attention mechanisms. The extracted and enhanced features are added to the original input to achieve self-enhancement of high- and low-frequency features.

3. The cotton similar disease detection method based on high-low frequency and attention fusion according to claim 1 is characterized in that: The attention-guided fusion module is used to effectively fuse feature maps from different levels and sources. The Pixel Attention module weights the key areas in the image in the spatial dimension, so that the model focuses on the target area.

4. The cotton similar disease detection method based on high-low frequency and attention fusion according to claim 1, characterized in that: The high- and low-frequency feature adaptive enhancement module uses deep convolution of large convolution kernels, reverse bottleneck design and residual connection to efficiently extract these high-frequency features while being lightweight, and enhances the details of high-frequency information through a dense connection mechanism.

5. The cotton similar disease detection method based on high-low frequency and attention fusion according to claim 2, characterized in that: The high-frequency features and low-frequency features are spliced ​​together to obtain high- and low-frequency features, and the high- and low-frequency features are fused as follows: After concatenating the high-frequency features and the low-frequency features in the channel dimension, 1×1 convolution is used to compress the number of channels, and the high- and low-frequency features are fused. The channel attention mechanism CALayer is introduced to adjust the feature weights. Finally, 3x3 convolution is used to refine the fused features. The residual connection is used to add the fused features to the input x to retain the original information.

6. The cotton similar disease detection method based on high-low frequency and attention fusion according to claim 1, characterized in that: The attention-guided fusion module uses the Pixel Attention module to calculate the final weighted coefficient of the cross-view feature interaction module, and fuses the left and right feature maps according to the weighted coefficient to output the final fused feature.

7. A cotton similar disease detection device based on high-low frequency and attention fusion, characterized in that: The device includes: a processor and a memory, wherein program instructions are stored in the memory, and the processor calls the program instructions stored in the memory to enable the device to execute the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein the computer program includes program instructions. When the program instructions are executed by a processor, the processor is caused to perform the method according to any one of claims 1 to 6.

Images