A small target detection method for rice leaf blast image recognition
By introducing an attention mechanism and a feature fusion module, the CenterNet network solves the problems of difficult small target detection and slow inference speed in rice leaf blast detection, and achieves high-precision and fast lesion identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2024-01-30
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies lack fine-grained classification of lesions and early detection of small targets in real farmland backgrounds for rice leaf blast detection, and the model inference speed is slow.
We adopted CenterNet, an anchorless object detection algorithm, as the baseline network, introduced an attention mechanism and a feature fusion module, designed a loss function suitable for small object detection, and improved model performance through data augmentation and feature fusion.
It achieves high accuracy and high inference speed for small target detection, with a detection accuracy of 95.4% and an inference speed of 122.1 FPS.
Smart Images
Figure CN117994656B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to small target detection, and more particularly to a method for small target detection in rice leaf blast image recognition. Background Technology
[0002] Crop pest and disease detection has always been a research hotspot in the agricultural field. Accurately identifying and effectively controlling various pests and diseases can effectively prevent yield reductions caused by large-scale crop infection. With the development of artificial intelligence technology, more and more advanced identification methods are being applied to the detection of pests and diseases in rice and other crops.
[0003] With the in-depth development of computer theory research, machine learning techniques combined with traditional feature extraction algorithms have been gradually applied to the field of crop disease identification, achieving significant results. Islam et al. artificially designed four color features by calculating the RGB values of the infected areas of rice leaves. These features were then used as input to a Naive Bayes algorithm to classify rice leaf spot, bacterial blight, and rice blast. Xiao et al. manually extracted 21 key features of rice lesions, then used Principal Components Analysis (PCA) to reduce the dimensionality of these disease features, and finally used a Back Propagation Neural Network (BP) to classify rice leaf blast. Pothen et al. achieved a classification accuracy of up to 94.6% for images of bacterial blight, rice blast, and sesame spot disease by segmenting the disease-infected area, using Local Binary Patterns (LBP) and Histogram of Oriented Gradient (HOG) to separate various features, and finally combining Support Vector Machine (SVM) and HOG algorithm to train a classifier.
[0004] While existing methods achieve high accuracy in crop disease diagnosis, their reliance on manual feature extraction limits model expressive power, robustness, and generalization ability. Unlike traditional feature extraction, Chen et al. trained Inception-VGGNet through transfer learning to classify various rice leaf diseases with good accuracy. Liang et al. designed a classification model for rice blast based on a Convolutional Neural Network (CNN), achieving over 95% accuracy by fine-tuning the CNN model's structure and parameters. However, these methods can only identify the presence and type of disease in an image, not its location. Especially when multiple diseases exist in a single image, relying solely on classification may lead to misdiagnosis or missed detection of certain diseases. Object detection is an image processing technique used for classification and localization. Zhou et al. proposed an object detection technique for rice disease detection based on the fusion of FCM-KM and Faster R-CNN. This method uses FCM-KM analysis to determine different sizes of candidate boxes for Faster R-CNN. The final trained model has good performance in terms of detection accuracy and inference speed for various rice diseases. Among them, the detection accuracy and inference speed for rice blast are 96.71% and 1.5 FPS, respectively.
[0005] However, the above literature still has certain limitations in rice blast detection: (1) It does not fully consider the shape and size of rice blast lesions, especially lacks detection of small target lesions in the early stage of rice blast. (2) It lacks fine-grained classification of lesions in real farmland background, usually treating them as the same category. Such a rough classification cannot well assess the infection status of rice blast in real farmland environment. (3) The inference speed of the model is relatively slow, and real-time detection tasks often have higher requirements for inference speed. Summary of the Invention
[0006] In the field of object detection, small targets are typically defined by area or length thresholds. A general definition of a small target is one whose bounding box area is less than 32×32 pixels. For an image with a resolution of 512×512 pixels, the corresponding relative area (the ratio of the bounding box area to the image area) should be less than 0.39%. Therefore, this invention defines objects with a relative area less than 0.39% as small targets. In the rice leaf blast lesion image sample dataset, all brown spot and acute lesions, as well as 82% of chronic lesions, fall into the category of small targets.
[0007] The purpose of this invention is to address the shortcomings of existing rice leaf blast research, such as the lack of fine-grained classification of lesions in real farmland settings, early detection of small targets in lesions, and slow model computation speed. This invention proposes a small target detection method for rice leaf blast image recognition. This method achieves a mAP of 95.4% for detecting different types of rice leaf blast and an inference speed of 122.1 FPS, demonstrating both high accuracy and high inference speed.
[0008] To solve the above problems, the technical solution of the present invention is as follows:
[0009] A method for small target detection in rice leaf blast image recognition, characterized by the following steps:
[0010] (1) Collect rice leaf blast lesion image sample dataset and perform image preprocessing to obtain rice leaf blast image sample data;
[0011] (2) Screen rice leaf blast image samples to obtain initial sample data of rice leaf blast images for disease detection;
[0012] (3) Use image annotation software to annotate the lesions in the initial sample data of rice leaf blast images with rectangular boxes to obtain a dataset of rice leaf blast lesion images.
[0013] (4) Data augmentation was performed on the rice leaf blast lesion image sample dataset to increase the number of samples and highlight the target features of the samples. The augmentation measures included: horizontal flipping, random cropping, brightness variation, HSV color conversion and Mosaic enhancement. Each measure had a 50% probability of being selected for enhancement for each sample.
[0014] (5) Target detection was performed on the enhanced rice leaf blast lesion image sample dataset. The anchorless target detection algorithm CenterNet was used as the baseline network, and an attention mechanism and feature fusion module were introduced. A loss function suitable for small target detection was designed to obtain the final GCPDFFNet detection model.
[0015] (6) Use the final GCPDFFNet detection model obtained in step (5) to detect rice leaf images and obtain the detection results of rice leaves.
[0016] Furthermore, in the initial sample data of rice leaf blast images in step (2): rice leaf blast lesions are divided into three types according to morphological characteristics: brown spot type, acute type and chronic type. All rice leaf blast samples are screened based on similarity to remove duplicates and samples with blurred target display in the image. The remaining samples are used as images to be labeled. Then, the Labelme image labeling software is used to label the lesions in the image with rectangular boxes.
[0017] Furthermore, in the specific method of sample data augmentation, the horizontal flipping is to mirror the image along the horizontal direction;
[0018] The random cropping process involves randomly selecting a portion of the original image and enlarging that portion to the same size as the original image to obtain a new image.
[0019] The brightness change is calculated by multiplying the image's pixel count by a constant, as shown in the following formula:
[0020] g(i,j)=αf(i,j)
[0021] Here, parameter α is called gain, f(i,j) is the pixel value of the input image, and g(i,j) is the pixel value of the output image; i and j represent the i-th row and j-th column of the pixel array, respectively.
[0022] The HSV color conversion involves converting an image from the RGB color space to the HSV color space, then applying random adjustments to increase or decrease the values of the H, S, and V channels, and finally converting it back from the HSV color space to the RGB color space to obtain the enhanced image. HSV color conversion can effectively simulate different lighting conditions and expand the sample space of the dataset.
[0023] The Mosaic enhancement involves randomly scaling, cropping, and arranging four original images, then stitching these four processed images together into one image, and finally scaling the stitched image to a specified size to obtain the enhanced image.
[0024] Each of the aforementioned data augmentation methods has a 50% probability of choosing whether to augment each sample. The purpose is to suppress the sample imbalance caused by the augmentation method for each disease type, so that the number of samples augmented by each method reaches a dynamic balance.
[0025] Furthermore, the specific steps for obtaining the GCPDFFNet detection model are as follows:
[0026] 1) Select the detection network architecture, and choose CenterNet, a deep learning-based anchorless detection algorithm, as the baseline network;
[0027] 2) Select ResNet50 as the backbone feature extraction network, and apply GC attention mechanism after the C2, C3 and C5 feature maps of ResNet50 respectively to extract lesion features;
[0028] 3) A PDFF feature fusion module is introduced after the P2 and P3 feature maps of the detection module; the PDFF module fuses the P2 and P3 feature maps and the feature maps extracted by the GC attention mechanism in C2 and C3 respectively;
[0029] 4) The proposed SNWD loss function, suitable for small object detection, is used for training to obtain the GCPDFFNet detection model.
[0030] Furthermore, the implementation principle of the PDFF module fusing high-level feature maps and low-level feature maps is as follows:
[0031] Given a high-level feature map X C×H×W and a low-level feature map Y C′×H×W Where C and C' represent the number of channels in the feature map, and H and W represent the length and width of the feature map, respectively; the low-level feature map Y C′×H×W The algorithm splits into two paths, one of which extracts the feature map P(Y). C×H×W The other path retains the original characteristics. P(Y) C×H×W It can be written as:
[0032] P(T) C×H×W =Conv(CBL(Y) C＇×H×W ))
[0033] Where Conv represents a convolution with kernel=1, stride=1, and padding=1, CBL consists of Conv, BN, and ReLU; CBL converts the low-level feature map Y C′×H×W The number of channels is changed from C' to C to ensure compatibility with the high-level feature map X. C×H×W The number of channels is consistent, while Conv further extracts features; the intermediate feature map W (C+C′)×H×W It is obtained by addition and concatenation, and its expression is:
[0034]
[0035] Output feature map Z of PDFF module C×H×W Through W (C+C′)×H×W The following two consecutive CBLs extract deeper information to generate:
[0036] Z C×H×W =CBL(CBL(W) (C+C′)×H×W ))
[0037] The PDFF module uses element-wise summation, concatenation, and convolution operations to process the input high-level feature map X. C×H×W and a low-level feature map Y C′×H×W The fusion process yields the output feature map Z. C×H×W .
[0038] Furthermore, the processed rice leaf blast lesion image sample dataset and the rice leaf images to be detected are used for target disease detection and identification. The quality of the detection result is jointly evaluated by the mean average precision (mAP) and the frame rate (FPS). mAP is obtained by averaging the AP values of the detected rice leaf blast lesions of various types. The AP value is an approximation of the area under the precision-recall curve, and the calculation formula is as follows: Where N represents the number of types of rice leaf blast lesions; FPS is usually the rate at which continuous images are captured or displayed, and the higher the FPS, the faster the model's inference speed; the above model for detecting and identifying rice diseases is the final GCPDFFNet detection model.
[0039] This invention provides a small target detection method for rice leaf blast image recognition. Rice leaf blast lesions are classified into three types based on morphological characteristics: brown spot type, acute type, and chronic type. Data augmentation methods are employed to expand the dataset sample space, enhancing the model's robustness and generalization ability. The proposed GCPDFFNet detection model, by introducing an attention mechanism and feature fusion module and redesigning the loss function, effectively addresses the difficulty of small target detection. Experimental validation was conducted on a rice leaf blast lesion image sample dataset, achieving an mAP of 95.4% and an inference speed of 122.1 FPS for the detection of the three types of rice leaf blast lesions, demonstrating high accuracy and high inference speed. Attached Figure Description
[0040] Figure 1 This is an overall flowchart of the method for detecting rice leaf blast lesions according to the present invention.
[0041] Figure 2-1 These are illustrations of three types of rice leaf blast lesions according to the present invention. Figure 2-2 The three types of rice leaf blast lesions of this invention are illustrated using Labelme annotations.
[0042] Figure 3-1 This invention provides data augmentation illustrations of a dataset of rice leaf blast lesion images, arranged from left to right and top to bottom as follows: horizontal flipping, random cropping, brightness variation, and HSV color conversion. Figure 3-2 Enhanced Mosaic graphics.
[0043] Figure 4This is a structural diagram of the GCPDFFNet detection model of the present invention, which consists of four main modules: input module, network backbone, detection module, and output module. A GC attention mechanism is applied after the C2, C3, and C5 feature maps of the ResNet50 network backbone to extract lesion features. A PDFF feature fusion module is introduced after the P2 and P3 feature maps of the detection module. The PDFF module fuses the P2 and P3 feature maps with the feature maps extracted by the GC attention mechanism from C2 and C3, respectively. During model training, the SNWD loss function is used for model regression.
[0044] Figure 5 This is a structural diagram of the PDFF feature fusion module of the present invention. The PDFF module fuses the input high-level feature map X through element-wise summation, concatenation, and convolution operations. C×H×W and a low-level feature map Y C′×H×W The fusion process yields the output feature map Z. C×H×W .
[0045] Figure 6 This is a visualization of the class activation graph of the GCPDFFNet detection model of the present invention at the last layer of the ResNet50 backbone network.
[0046] Figure 7 This is a diagram showing the detection effect of the PDFF module of the GCPDFFNet detection model of the present invention on small-sized rice leaf blast lesions.
[0047] Figure 8 This refers to the change in detection accuracy of the GCPDFFNet detection model of the present invention on the rice leaf blast lesion image sample dataset during the improvement process. Detailed implementation method:
[0048] To make the description of the objectives, technical solutions, and advantages of this invention clearer and more accurate, the invention will be further described below in conjunction with embodiments and accompanying drawings. It should be specifically noted that the illustrative embodiments and descriptions of this invention are only for explaining the invention, but the embodiments of this invention are not limited thereto:
[0049] The small target detection method for rice leaf blast image recognition provided by this invention trains the model on a dataset of rice leaf blast lesion images taken in real farmland scenes, and continuously optimizes the algorithm to address the difficulties in small target detection, resulting in the GCPDFFNet detection model.
[0050] In this embodiment, refer to the appendix Figure 1This paper describes a method for detecting three types of rice blast disease lesions in a dataset of rice leaf blast lesion images. The specific implementation process is as follows: First, rice leaf blast lesion image samples are collected from real farmland environments and preprocessed to obtain rice leaf blast image sample data. Then, the rice leaf blast image samples are screened, labeled, and augmented to obtain a dataset of rice leaf blast lesion images. For the augmented rice leaf blast lesion images, target detection is performed using the anchor-free target detection algorithm CenterNet as the baseline network. A GC attention mechanism and a PDFF feature fusion module are introduced, and a SNWD loss function suitable for small target detection is designed to obtain the GCPDFFNet detection model. Finally, the GCPDFFNet detection model is used to detect the three types of rice leaf blast lesions. The detection results are evaluated by the mean average precision (mAP) and frame rate (FPS). mAP is obtained by averaging the AP values of the detected rice leaf blast lesions of each type. The AP value is an approximation of the area under the precision-recall curve, and the calculation formula is... Where N represents the number of types of rice blast lesions; FPS is typically the rate at which consecutive images are captured or displayed; a higher FPS indicates a faster inference speed. Dataset preprocessing involved using OpenCV to resize images to 512×512 pixels, cropping any unnecessary parts of the original images. Then, all rice blast image samples were filtered based on similarity, removing identical or extremely low-quality samples, retaining the remaining images for annotation. Labelme image annotation software was then used to annotate the lesions with bounding boxes. The final result was a dataset of 500 images of rice blast lesions, including brown spot, acute, and chronic types. Images of brown spot, acute, and chronic types are referenced... Figure 2-1 ;Labelme annotation example reference Figure 2-2 .
[0051] In this embodiment, data augmentation of the rice leaf blast lesion image sample dataset includes: horizontal flipping, random cropping, brightness variation, HSV color conversion, and Mosaic enhancement. For example... Figure 3-1 The data augmentation methods, arranged from left to right and top to bottom, are horizontal flipping, random cropping, brightness variation, and HSV color conversion. These methods are performed before training, expanding the image sample to 1200 images after augmentation. Horizontal flipping involves mirroring the image along the horizontal direction. Random cropping involves randomly selecting a region from the original image and enlarging that region to the same size as the original image to obtain a new image. Both horizontal flipping and random cropping perform data augmentation from the perspective of image spatial location, preventing the detection algorithm from treating the location of lesions in the image as an important feature and affecting the model's detection performance.
[0052] The brightness change is calculated by multiplying the image's pixel count by a constant, as shown in the following formula:
[0053] g(i,j)=αf(i,j)
[0054] Here, parameter α is called gain, f(i,j) is the pixel value of the input image, and g(i,j) is the pixel value of the output image; i and j represent the i-th row and j-th column of the pixel array, respectively.
[0055] The HSV color conversion involves converting an image from the RGB color space to the HSV color space, then applying random adjustments to increase or decrease the values of the H, S, and V channels, and finally converting it back from the HSV color space to the RGB color space to obtain the enhanced image. HSV color conversion can effectively simulate different lighting conditions and expand the sample space of the dataset.
[0056] Both the brightness variation and the HSV color conversion are data augmentations performed at the color level to enrich the lighting variations in the dataset, thereby improving the robustness and generalization of the model.
[0057] like Figure 3-2 The Mosaic enhancement process involves randomly scaling, cropping, and arranging four original images, then stitching these four processed images together into a single image. Finally, the stitched image is scaled to a specified size to obtain the enhanced image. This Mosaic enhancement is used during model training; it enriches the background of the training images to improve the model's generalization ability, better utilizes GPU resources, and enhances training efficiency.
[0058] In this embodiment, all rice blast disease samples are screened based on similarity to avoid repeated learning of the same samples by the model in later stages, thus avoiding overfitting to some extent. Deduplication refers to selecting only one image for each completely identical sample to achieve the purpose of removing duplicates. For removing blurry samples from the images, the Laplacian gradient function is selected for blur evaluation, and the blur threshold is set to 400. Image samples with a blur value greater than or equal to 400 are deleted.
[0059] In this embodiment, refer to the appendix. Figure 4The GCPDFFNet detection model consists of four main parts: an input module, a network backbone, a detection module, and an output module. A GC attention mechanism is applied after the C2, C3, and C5 feature maps of the ResNet50 backbone to extract lesion features. A PDFF feature fusion module is introduced after the P2 and P3 feature maps of the detection module. The PDFF module fuses the P2 and P3 feature maps with the feature maps extracted by the GC attention mechanism from C2 and C3, respectively. During model training, the SNWD loss function is used for model regression. Addressing the challenges of complex backgrounds and difficulty in detecting small targets in rice blast lesion detection, the GCPDFFNet detection model uses CenterNet as the baseline network and introduces a GC attention mechanism to focus more on lesion regions, improving feature extraction capabilities. The proposed PDFF module fuses low-level feature maps with rich detail information with high-level feature maps with high-level semantic information, enhancing the recognition effect of small lesions. Simultaneously, the SNWD loss function is designed to accelerate model convergence speed and efficiency, improving the detection performance of rice blast lesions.
[0060] In this embodiment, refer to the appendix. Figure 5 The implementation principle of the PDFF feature fusion module is as follows: Given a high-level feature map X C×H×W and a low-level feature map Y C′×H×W Where C(C'), H, and W represent the number of channels, length, and width of the feature map, respectively; and the low-level feature map Y... C′×H×W The algorithm splits into two paths, one of which extracts the feature map P(Y). C×H×W The other path retains the original characteristics. P(Y) C×H×W It can be written as:
[0061] P(Y) C×H×W =Conv(CBL(Y) C＇×H×W ))
[0062] Where Conv represents a convolution with kernel=1, stride=1, and padding=1, CBL consists of Conv, BN, and ReLU; CBL converts the low-level feature map Y C′×H×W The number of channels is changed from C' to C to ensure compatibility with the high-level feature map X. C×H×W The number of channels is consistent, while Conv further extracts features; the intermediate feature map W (C+C′)×H×W It is obtained by addition and concatenation, and its expression is:
[0063]
[0064] Output feature map Z of PDFF module C×H×W Through W (C+C′)×H×WThe following two consecutive CBLs extract deeper information to generate:
[0065] Z C×H×W =CBL(CBL(W) (C+C′)×H×W ))
[0066] The PDFF module uses element-wise summation, concatenation, and convolution operations to process the input high-level feature map X. C×H×W and a low-level feature map Y C′×H×W The fusion process yields the output feature map Z. C×H×W .
[0067] In this embodiment, the appendix Figure 6 The visualization of class activation maps of the GCPDFFNet detection model in the last layer of the ResNet50 backbone network is presented. In the figure, (a) is the image of the detected rice leaf, and (b) and (c) are the visualizations of class activation maps of the last layer of the backbone network of the GCPDFFNet detection model before and after the addition of the GC module, respectively. The bright spots in the figure represent the regions of interest of the model. It can be seen that the addition of the GC module helps the model focus on the rice leaf blast lesion region, which is helpful for target localization and regression, thereby improving the model's detection accuracy.
[0068] In this embodiment, the appendix Figure 7 The results of the CPDFFNet detection model's PDFF module in detecting small rice leaf blast lesions are presented. Figures (a) and (b) show the detection results of the CPDFFNet model before and after adding the PDFF module for small brown spot lesions, respectively. The figures demonstrate that the PDFF module helps the CPDFFNet detection model improve the classification results of similar lesions and avoid misidentifying some background as lesions. The PDFF module, through parallel processing, fully integrates with high-level features without losing low-level features, ensuring that the fused feature map contains rich spatial and semantic information, thus greatly improving the detection accuracy of small brown spot lesions.
[0069] In this embodiment, the appendix Figure 8The changes in detection accuracy of the GCPDFFNet detection model on a dataset of rice leaf blast lesion images during the improvement process are presented. Different models (CenterNet, CenterNet+GC, CenterNet+GC+PDFF, GCPDFFNet) show different mean accuracy (AP) for the three types of rice leaf blast lesions. It can be seen that the improved GCPDFFNet detection model significantly improves the detection accuracy of rice leaf blast lesions, especially for the smallest brown dot lesions. Experimental results show that these improvements are effective in enhancing model performance. Table 1 also presents the performance comparison results of different models on the dataset during the improvement process. The final GCPDFFNet detection model achieved an mAP of 95.4% for different categories of rice leaf blast detection and an inference speed of 122.1 FPS. In Table 1, the "Model" column lists the names of the different models improved during the experiment. The overall experimental procedure for model improvement includes:
[0070] Starting from the baseline CenterNet detection model, this paper introduces a GC attention mechanism to improve the model's feature extraction capability, resulting in a 3.1% improvement in detection mAP and a 16.7% decrease in FPS. Then, to address the poor performance in small object detection, a PDFF module is proposed to fuse low-level and high-level features, resulting in a 6.0% improvement in detection mAP and a 6.5% decrease in FPS. Finally, a SNWD loss function suitable for small objects is designed to accelerate model convergence and improve accuracy, ultimately improving the model's mAP by 2.7% while only decreasing FPS by 2.6. In conclusion, compared to CenterNet, the proposed GCPDFFNet detection model achieves an 11.8% improvement in mAP, reaching 95.4%, while the FPS only decreases from 147.9 to 122.1, maintaining a relatively fast inference speed.
[0071] Table 1 shows the performance comparison results of different models on the dataset during the improvement process.
[0072]
[0073] The above description is merely a further embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitution or partial modification made by those skilled in the art within the scope of the technology disclosed in the present invention, according to the technical process and method steps of the present invention, and the resulting functions and effects do not exceed the information covered by the specification and drawings, shall fall within the scope of protection of the present invention.
Claims
1. A method for small target detection in rice leaf blast image recognition, characterized in that, Includes the following steps: (1) Collect rice leaf blast lesion image sample dataset and perform image preprocessing to obtain rice leaf blast image sample data; (2) Screen rice leaf blast image samples to obtain initial sample data of rice leaf blast images for disease detection; (3) Use image annotation software to annotate the lesions in the initial sample data of rice leaf blast images with rectangular boxes to obtain a dataset of rice leaf blast lesion images; (4) Data augmentation was performed on the rice leaf blast lesion image sample dataset to increase the number of samples and highlight the target features of the samples; The enhancement measures include: horizontal flipping, random cropping, brightness variation, HSV color conversion, and Mosaic enhancement. Each measure has a 50% probability of being selected for enhancement for each sample. (5) Target detection was performed on the enhanced rice leaf blast lesion image sample dataset. The anchorless target detection algorithm CenterNet was used as the baseline network, and an attention mechanism and feature fusion module were introduced. A loss function suitable for small target detection was designed to obtain the final GCPDFFNet detection model. (6) Use the final GCPDFFNet detection model obtained in step (5) to detect rice leaf images and obtain the detection results of rice leaves; The specific steps for obtaining the GCPDFFNet detection model in step (5) are as follows: 1) Select the detection network architecture, and choose CenterNet, a deep learning-based anchorless detection algorithm, as the baseline network; 2) ResNet50 was selected as the backbone feature extraction network, and GC attention mechanism was applied to the C2, C3 and C5 feature maps of ResNet50 to extract lesion features. 3) A PDFF feature fusion module is introduced after the P2 and P3 feature maps of the detection module; the PDFF module fuses the P2 and P3 feature maps and the feature maps extracted by the GC attention mechanism in C2 and C3 respectively; 4) The proposed SNWD loss function, suitable for small object detection, is used for training to obtain the GCPDFFNet detection model.
2. The method for small target detection in rice leaf blast image recognition according to claim 1, characterized in that, In the initial sample data of rice leaf blast images in step (2): Rice leaf blast lesions are classified into three types based on morphological characteristics: brown spot type, acute type, and chronic type. All rice leaf blast samples were screened based on similarity to remove duplicates and samples with blurred target displays in the images. The remaining samples were used as images to be labeled. Then, the Labelme image annotation software was used to annotate the lesions in the images with rectangular boxes.
3. The method for small target detection in rice leaf blast image recognition according to claim 1, characterized in that, In the specific method of sample data augmentation in step (4): The horizontal flipping refers to mirroring the image along the horizontal direction. The random cropping process involves randomly selecting a portion of the original image and enlarging that portion to the same size as the original image to obtain a new image. The brightness change is calculated by multiplying the image's pixel count by a constant, as shown in the following formula: Here, parameter α is called gain, f(i,j) is the pixel value of the input image, and g(i,j) is the pixel value of the output image; i and j represent the i-th row and j-th column of the pixel array, respectively. The HSV color conversion involves converting an image from the RGB color space to the HSV color space, then applying random adjustments to increase or decrease the values of the H, S, and V channels, and finally converting it back from the HSV color space to the RGB color space to obtain the enhanced image. HSV color conversion can effectively simulate different lighting conditions and expand the sample space of the dataset. The Mosaic enhancement involves randomly scaling, cropping, and arranging four original images, then stitching these four processed images together into one image, and finally scaling the stitched image to a specified size to obtain the enhanced image. Each of the aforementioned data augmentation methods has a 50% probability of choosing whether to augment each sample. The purpose is to suppress the sample imbalance caused by the augmentation method for each disease type, so that the number of samples augmented by each method reaches a dynamic balance.
4. The method for small target detection in rice leaf blast image recognition according to claim 1, characterized in that, The implementation principle of the PDFF module fusing high-level feature maps and low-level feature maps is as follows: Given a high-level feature map and a low-level feature map ,in The number of channels in the feature map. , These represent the length and width of the feature map, respectively. Low-level feature maps The process is divided into two paths, one of which extracts the feature map. The other route retains its original characteristics; It can be written as: in This represents a convolution with kernel=1, stride=1, and padding=1. Depend on Composed of BN and ReLU; Low-level feature maps The number of channels is determined by Change into Ensure compatibility with high-level feature maps The number of channels is the same, while Further feature extraction; intermediate feature map It is obtained by addition and concatenation, and its expression is: Output feature map of PDFF module pass The next two consecutive Generate by extracting deeper information: The PDFF module uses element-wise summation, concatenation, and convolution operations to process the input high-level feature map. and a low-level feature map The output feature map is obtained by fusion. .
5. The method for small target detection in rice leaf blast image recognition according to claim 1, characterized in that, Using the above step (5), the target disease detection and identification are performed on the rice leaf blast lesion image sample dataset processed in step (3) and the rice leaf image to be detected. The quality of the detection result is jointly evaluated by the mean of the average detection accuracy mAP and the frame rate FPS. mAP is obtained by averaging the AP values of various types of rice leaf blast lesions detected. The AP value is an approximation of the area under the precision-recall curve. The calculation formula is as follows: , where N represents the number of types of rice leaf blast lesions; FPS is usually the rate at which continuous images are captured or displayed, and the higher the FPS, the faster the model's inference speed; the above model for detecting and identifying rice diseases is the final GCPDFFNet detection model.