A method for detecting potato late blight based on a multi-scale attention hybrid head

Abstract

The application discloses a kind of based on multi-scale attention mixed head's potato late blight detection method.It includes the following steps: obtaining potato late blight leaf image dataset, the diseased leaf in potato late blight leaf image is labeled, and standardization annotation dataset is obtained;Potato late blight detection model is constructed, and the potato late blight detection model includes multi-scale feature extraction network, multi-scale attention mixed head, decoding prediction network;Standardization annotation dataset is used to train potato late blight detection model;The potato leaf image to be measured is input into the trained potato late blight detection model, and the potato late blight detection model outputs the position of the diseased leaf on the potato leaf image to be measured, disease class.The application has good scale adaptability, sufficient feature interaction, strong anti-interference ability, and considers detection accuracy and deployment efficiency.

Description

Technical Field

[0001] This invention relates to the field of crop disease detection technology, and in particular to a method for detecting potato late blight based on a multi-scale attention hybrid head. Background Technology

[0002] Potato late blight is one of the major diseases of potatoes and has always been one of the most serious threats to the global potato industry, causing yield losses and control costs amounting to tens of billions of dollars every year.

[0003] Current detection technologies for potato late blight mainly include traditional manual inspection, non-destructive testing based on spectral analysis, and deep learning-based visual detection methods. Among these, deep learning-based visual detection has become the mainstream research direction due to its advantages of adapting to complex field environments and high detection efficiency. Currently, target detection frameworks such as the YOLO series, RTMDet, and RT-DETR-R18 are widely used to identify and locate disease targets. However, the complex farmland environment brings multiple challenges to detection technologies: on the one hand, there are significant scale differences in potato late blight diseased leaves in the field. Influenced by factors such as plant growth stage and leaf position, there are small-scale young diseased leaves, medium-scale mature diseased leaves, and large-scale old leaves, with significant differences in the feature expression of targets at different scales; on the other hand, the complex background interference in the field, such as soil, weeds, and crop residues, can easily be confused with the characteristics of lesions, further increasing the difficulty of detection.

[0004] Existing deep learning-based visual detection methods have significant technical shortcomings in addressing the aforementioned challenges: First, traditional detection models often employ non-interactive fully convolutional head structures, achieving scale adaptation only through simple multi-scale feature concatenation or weighted fusion. This lacks deep interaction and global semantic association between features at different scales, leading to the easy omission of small-scale diseased leaves due to insufficient detailed features. Second, the feature fusion process does not specifically address redundant background information, resulting in interference between background noise in shallow features and semantic information in deep features. This weakens the model's resistance to interference from complex backgrounds, making it prone to false detections. Third, existing multi-scale fusion methods often prioritize detection accuracy over model lightweighting, or sacrifice feature details to compress network width, making it difficult to balance detection efficiency and accuracy and adapt to the deployment requirements of portable equipment for field inspections in farmland. Fourth, some improved methods enhance performance by increasing the number of feature extraction layers or sensors, but this leads to a surge in model parameters and increased inference latency, which is not suitable for practical applications in field inspections.

[0005] Therefore, it is necessary to design a potato late blight detection method based on a multi-scale attention hybrid head to solve the technical problems of poor scale adaptability, insufficient feature interaction, weak resistance to background interference, and difficulty in balancing efficiency and accuracy in the existing technology, so as to meet the needs of precision and efficiency in actual field inspection. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for detecting potato late blight based on a multi-scale attention hybrid head, which has good scale adaptability, sufficient feature interaction, strong anti-interference ability, and balances detection accuracy and deployment efficiency.

[0007] To solve the above problems, the present invention adopts the following technical solution: The present invention discloses a method for detecting potato late blight based on a multi-scale attention hybrid head, comprising the following steps: S1: Obtain a dataset of potato late blight leaf images, label the diseased leaves in the potato late blight leaf images, and obtain a standardized labeled dataset; S2: Construct a potato late blight detection model, which includes a multi-scale feature extraction network, a multi-scale attention hybrid head, and a decoding prediction network; S3: The potato late blight detection model was trained using a standardized labeled dataset; S4: Input the potato leaf image to be tested into the trained potato late blight detection model. The potato late blight detection model outputs the location of the diseased leaves and the disease type on the potato leaf image to be tested.

[0008] Preferably, the multi-scale feature extraction network is an encoder of the DFINE model, used to extract features layer by layer from the input potato leaf image, obtain semantic and detail information at different scales, and output feature maps X1, X2, and X3 to the multi-scale attention hybrid head. The multi-scale attention hybrid head is used to perform cross-channel cross-attention modeling and cross-scale feature fusion processing on feature maps X1, X2, and X3, and outputs the fused feature maps P1 and P2 to the decoding prediction network. The decoding prediction network is a decoder of the DFINE model, used to perform channel splicing and aggregation on feature maps P1 and P2, progressive feature extraction through multi-level decoding layers, and introduce self-distillation between adjacent decoding layers to achieve feature knowledge transfer optimization, outputting the location and disease type of diseased leaves on potato leaf images.

[0009] Preferably, the multi-scale attention mixing head includes a preprocessing module, a multi-scale cross-attention module, and a cross-scale compression fusion module.

[0010] Preferably, the preprocessing module performs average pooling operations on the input feature maps X1, X2, and X3 respectively to generate feature maps f1, f2, and f3, and outputs them to the multi-scale cross-attention module; channel concatenation, channel shuffling, and channel segmentation operations are performed on the feature maps f1, f2, and f3 in sequence to obtain key vector k and value vector v; the key vector k is subjected to dimension reshaping and vector normalization processing in sequence to obtain standard vector K; the value vector v is subjected to dimension reshaping processing to obtain standard vector V; and the standard vectors K and V are output to the multi-scale cross-attention module.

[0011] Preferably, the multi-scale cross-attention module sequentially performs dimension reshaping and vector normalization on the input feature maps f1, f2, and f3 to obtain standard vectors F1, F2, and F3. Based on the standard vectors F... i The attention weight A is calculated based on the similarity to the standard vector K. i For i=1~3, the attention weight A i With standard vector V and feature map f i The corresponding feature map O is obtained by fusion. i , feature map O i Output to the cross-scale compression and fusion module.

[0012] Preferably, the cross-scale compression fusion module performs three bilinear interpolation upsampling operations on the input feature map O1, with the second bilinear interpolation upsampling outputting feature map C3 and the third bilinear interpolation upsampling outputting feature map C2; for feature map O1... j After performing one bilinear interpolation upsampling operation, the feature map X is compared with... j Perform element-wise addition, then sum to feature map C. j Perform channel splicing to output feature map T j For j=2~3, convolution is performed on feature maps T2 and T3 respectively, and the fused feature maps P1 and P2 are output to the decoding prediction network.

[0013] As a preferred option, based on the standard vector F i The attention weight A is calculated based on the similarity to the standard vector K. i The method is as follows: The standard vector F i The attention weights A are obtained by multiplying them element-wise with the standard vector K and then dividing by the scaling factor. i .

[0014] Preferably, the feature map O i The calculation formula is as follows: , in, For instance normalization function, Let the exponential normalization function take values ​​between (0,1). This indicates element-wise multiplication. This indicates element-wise addition. For dimension reshaping functions.

[0015] Preferably, the method for annotating diseased leaves in potato late blight leaf images includes the following steps: Using the Labelme open-source annotation software, we manually outlined the diseased leaves in the potato late blight leaf image with polygonal boxes to mark the location of the diseased leaves, labeled the disease category, and generated the corresponding JSON format annotation file.

[0016] Preferably, step S1 further includes the following steps: performing image enhancement processing on each potato late blight leaf image in the standardized annotation dataset to generate more potato late blight leaf images and expand the standardized annotation dataset.

[0017] The beneficial effects of this invention are: (1) A multi-scale attention hybrid head integrating a multi-scale cross-attention module and a cross-scale compression fusion module is designed and embedded into the encoder of the DFINE model, thereby simplifying the network structure while strengthening the deep interaction of diseased leaf features at different scales, and taking into account both detection accuracy and deployment efficiency. (2) In order to accurately capture the core features of small-scale young diseased leaves, medium-scale mature diseased leaves and large-scale diseased old leaves of potato late blight, and suppress the redundant interference of complex backgrounds such as farmland soil and weeds, a multi-scale cross-attention module is used to construct global semantic association to give higher weight to diseased leaf features and selectively ignore invalid background information, thereby significantly improving the detection accuracy of diseased leaves at different scales and reducing the risk of false detection in complex scenarios. (3) In order to further reduce the scale of network parameters and computational complexity and adapt to the real-time detection needs of resource-constrained scenarios such as farmland portable terminals and drones, the cross-scale compression fusion module compresses three input feature channels into two core output channels, greatly reducing feature redundancy. Attached Figure Description

[0018] Figure 1 This is a flowchart of an embodiment; Figure 2 This is a schematic diagram of the structure of a multi-scale attention hybrid head; Figure 3 These are multi-scale images of potato leaves infected with late blight; Figure 4 The method in this embodiment is for Figure 3 The detection result image obtained by detecting the image in the middle. Detailed Implementation

[0019] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0020] Example: This example describes a method for detecting potato late blight based on a multi-scale attention hybrid head, such as... Figure 1 As shown, it includes the following steps: S1: Obtain a dataset of potato late blight leaf images, label the diseased leaves in the potato late blight leaf images to obtain a standardized label dataset, perform image enhancement processing on each potato late blight leaf image in the standardized label dataset to generate more potato late blight leaf images and expand the standardized label dataset.

[0021] A potato late blight leaf image dataset was created by acquiring 536 potato late blight leaf images from the publicly available dataset, the Irish potato imagery dataset for detection of early and late blight diseases. Each image contains multiple potato leaves, some of which are infected with late blight. The 536 images cover late blight leaves at different scales: small, medium, and large. Generally, small-scale late blight leaves have a pixel area less than 32*32, medium-scale leaves have a pixel area between 32*32 and 96*96, and large-scale leaves have a pixel area greater than 96*96.

[0022] The method for enhancing potato late blight leaf images to generate more potato late blight leaf images includes the following steps: Based on the original potato late blight leaf images, geometric transformations, pixel noise perturbations, image brightness adjustments, and color attribute adjustments were performed to obtain the first, second, third, and fourth amplified images.

[0023] The method for labeling diseased leaves in potato late blight leaf images includes the following steps: Using the Labelme open-source annotation software, we manually outlined the diseased leaves in potato late blight leaf images using polygonal bounding boxes to mark the location of the diseased leaves and label them with disease category tags, generating a corresponding JSON format annotation file. The only disease category is late blight, and the label category is 0, which represents late blight.

[0024] S2: Construct a potato late blight detection model, which includes a multi-scale feature extraction network, a multi-scale attention hybrid head, and a decoding prediction network.

[0025] The multi-scale feature extraction network is the encoder of the DFINE model, which is used to extract features layer by layer from the input potato leaf image to obtain semantic and detail information at different scales, and output feature maps X1, X2, and X3 to the multi-scale attention hybrid head. The multi-scale attention hybrid head is used to perform cross-channel cross-attention modeling and cross-scale feature fusion processing on feature maps X1, X2, and X3, and outputs the fused feature maps P1 and P2 to the decoding and prediction network. The decoding prediction network is a decoder of the DFINE model, used to perform channel splicing and aggregation on feature maps P1 and P2, progressive feature extraction through multi-level decoding layers, and introduce self-distillation between adjacent decoding layers to achieve feature knowledge transfer optimization, outputting the location of diseased leaves and disease type on potato leaf images.

[0026] like Figure 2 As shown, the multi-scale attention hybrid head includes a preprocessing module, a multi-scale cross-attention module, and a cross-scale compression fusion module.

[0027] The preprocessing module performs average pooling on the input feature maps X1, X2, and X3 respectively to generate feature maps f1, f2, and f3, and outputs them to the multi-scale cross-attention module. The feature maps f1, f2, and f3 are sequentially processed by channel concatenation, channel shuffling, and channel segmentation to obtain the key vector k and the value vector v. The key vector k is sequentially processed by dimension reshaping and vector normalization to obtain the standard vector K. The value vector v is processed by dimension reshaping to obtain the standard vector V. The standard vectors K and V are output to the multi-scale cross-attention module.

[0028] The method for performing channel concatenation, channel shuffling, and channel segmentation operations sequentially on feature maps f1, f2, and f3 to obtain the key vector k and value vector v includes the following steps: Perform channel stitching on feature maps f1, f2, and f3 to obtain stitched image B. , in, This is a channel splicing function; Perform channel shuffling and channel splitting operations (1:1 splitting) on ​​the stitched image B to obtain the key vector k and value vector v. , in, This is the channel shuffling function. This is the channel segmentation function.

[0029] The formula for obtaining the standard vector K by successively reshaping the dimension and normalizing the vector k is: , in, It is a vector normalization function. This is a dimension reshaping function used to convert a two-dimensional matrix into a one-dimensional vector.

[0030] The formula for reshaping the dimensions of the value vector v to obtain the standard vector V is: .

[0031] The multi-scale cross-attention module sequentially reshapes and normalizes the input feature maps f1, f2, and f3 to obtain standard vectors F1, F2, and F3. The standard vector F... i The attention weights A are obtained by multiplying them element-wise with the standard vector K and then dividing by the scaling factor. i For i=1~3, the attention weight A i With standard vector V and feature map f i The corresponding feature map O is obtained by fusion. i , feature map O i Output to the cross-scale compression and fusion module.

[0032] Attention weight A i The calculation formula is as follows: , , in, It is a constant (in this embodiment, it is half the sum of the number of channels in feature maps X1, X2, and X3). This is the scaling factor (used to adjust the scaling ratio). For vector normalization function, For dimension reshaping functions, This indicates element-wise multiplication.

[0033] Pay attention weight A i With standard vector V and feature map f i The corresponding feature map O is obtained by fusion. i The method includes the following steps: First, the attention weight A i Instance normalization is performed to ensure numerical stability. Then, the attention weights are normalized to a probability distribution using the Softmax function. Finally, they are element-wise multiplied with the standard vector V and then multiplied with the feature map f. i Element-wise addition is performed, and finally, dimensionality reshaping is performed to obtain the feature map O. i ; Feature map O i The calculation formula is as follows: , in, For instance normalization function, Let the exponential normalization function take values ​​between (0,1). This indicates element-wise addition.

[0034] The cross-scale compression fusion module performs bilinear interpolation upsampling on the input feature map O1 to output feature map C4, performs bilinear interpolation upsampling on feature map C4 to output feature map C3, and performs bilinear interpolation upsampling on feature map C3 to output feature map C2; then, it performs bilinear interpolation upsampling on feature map O1 to output feature map C4. j After performing one bilinear interpolation upsampling operation, the feature map X is compared with... j Perform element-wise addition, then sum to feature map C. j Perform channel splicing to output feature map T j For j=2~3, convolution is performed on feature maps T2 and T3 respectively, and the fused feature maps P1 and P2 are output to the decoding prediction network.

[0035] Feature map T j The calculation formula is as follows: , in, This is a bilinear interpolation upsampling function.

[0036] Feature map The calculation formula is as follows: , in, It is a two-dimensional convolutional layer.

[0037] The multi-scale feature extraction network outputs feature maps X1, X2, and X3. For example, feature map X1 is 80*80 pixels, feature map X2 is 40*40 pixels, and feature map X3 is 20*20 pixels. Feature maps X1, X2, and X3 are then subjected to average pooling to uniformly scale their dimensions to the size of feature map X3. , For feature map The number of channels, The height of feature map X3, The width of feature map X3 is then used. Subsequently, these feature maps are input into a multi-scale cross-attention module to obtain global semantic features, ultimately generating three output feature maps O1, O2, and O3. After that, feature maps O1, O2, and O3 enter a cross-scale compression fusion module for cross-channel feature fusion. This fusion process effectively utilizes the shallowest channel features and compresses the network width, converting the three input channels into two output channels. After fusion, feature maps P1 and P2 are generated.

[0038] The multi-scale cross-attention module is used to fuse multi-scale features to build a global semantic representation and enhances feature expression capabilities through a multi-head interaction mechanism. Unlike traditional methods, it uses channel shuffling and channel splitting to generate keys (k) and values ​​(v), replacing the commonly used 1×1 convolution, thereby reducing redundant computation and effectively reducing the dimensionality of keys and values.

[0039] The cross-scale compression fusion module performs three bilinear interpolation upsampling operations on feature map O1 to generate three feature maps C4, C3, and C2. The remaining two feature maps O2 and O3 are restored to their original size through one bilinear interpolation upsampling operation, and then connected to the input feature maps X2 and X3 respectively via residual connections. These are then further connected to C2 and C3 to generate feature maps T2 and T3. Finally, they are fused through two parallel 1x1 Conv2d (two-dimensional convolutional layers) to obtain the final feature maps P1 and P2. The cross-scale compression fusion module effectively enhances the model's ability to capture detailed information by fusing the shallowest feature channels, thereby improving overall performance. As the network deepens and the number of feature channels gradually increases, the contribution of the shallowest channels decreases when fused with features from shallow to deep layers. This design strategy aligns with the actual feature representation rules and is beneficial for achieving network width compression. Finally, efficient information integration across feature layers is achieved through parallel 1×1 convolutional layers.

[0040] S3: The potato late blight detection model was trained using a standardized labeled dataset.

[0041] S4: Input the potato leaf image to be tested into the trained potato late blight detection model. The potato late blight detection model outputs the location of the diseased leaves and the disease type on the potato leaf image to be tested.

[0042] In this scheme, a multi-scale attention hybrid head integrating a multi-scale cross-attention module and a cross-scale compression fusion module is designed and embedded after the encoder of the DFINE model. This enhances the deep interaction of diseased leaf features at different scales while simplifying the network structure, balancing detection accuracy and deployment efficiency. To accurately capture the core features of small-scale young leaves, medium-scale mature leaves, and large-scale old leaves affected by potato late blight, while suppressing redundant interference from complex backgrounds such as farmland soil and weeds, a multi-scale cross-attention module is used to construct global semantic associations, giving higher weights to diseased leaf features and selectively ignoring invalid background information. This significantly improves the detection accuracy of diseased leaves at different scales and reduces the risk of false detections in complex scenarios. To further reduce the network parameter scale and computational complexity, and adapt to the real-time detection needs of resource-constrained scenarios such as portable farmland terminals and drones, the cross-scale compression fusion module compresses three input feature channels into two core output channels, greatly reducing feature redundancy.

[0043] Using the method of this embodiment to Figure 3 Nine multi-scale images of potato leaves infected with late blight were used for detection. Figure 3 The diseased leaves in the first three images are small-scale leaves. Figure 3 The diseased leaves in the second row of three images are mesoscale leaves. Figure 3 The diseased leaves in the third row of three images are large-scale leaves, and the detection results are as follows. Figure 4 As shown, Figure 4 The test results in the image show the location of the diseased leaves marked with blue rectangles, and the disease type is late blight.

[0044] Verification experiment: To verify the detection performance of the method in this embodiment, YOLO series models, RTMDet model, RT-DETR-R18 model, and the potato late blight detection model (MCCE-DFINE) trained by the method in this embodiment were used to detect the public potato late blight dataset. The detection statistics are shown in Table 1. Table 1

[0045] AP50 and AP50-95 are core metrics for evaluating model performance in object detection tasks. As shown in Table 1, the MCCE-DFINE model trained using the method in this embodiment outperforms the YOLOV8n, YOLO11n, YOLO12n, RTMDet, and RT-DETR-R18 models in both AP50 and AP50-95, achieving performance improvements. Regarding model complexity, the MCCE-DFINE model has only 4.39 M parameters and 9.91 G of floating-point operations, significantly lower than the RT-DETR-R18 model. It maintains moderate computational overhead while achieving high accuracy, making it more suitable for real-world application scenarios.

[0046] To verify the effectiveness of the multi-scale attention hybrid head, a systematic ablation experiment was conducted on the public potato late blight dataset. The potato late blight detection model (MCCE-DFINE) was compared with the baseline model DFINE (with the multi-scale attention hybrid head removed) under the same training conditions and hyperparameter settings. The results are shown in Table 2. Table 2

[0047] As shown in Table 2, the potato late blight detection model (MCCE-DFINE) improved the detection performance by 0.3% and 1.3% in AP50 and AP50-95, respectively, verifying the positive effect of the multi-scale attention hybrid head in enhancing feature representation and semantic fusion. Although the number of parameters and floating-point operations increased slightly, the resulting improvement in detection performance was significant, indicating that the multi-scale attention hybrid head can effectively improve the potato late blight detection model (MCCE-DFINE) in terms of multi-scale feature fusion and detail perception.

Claims

1. A method for detecting potato late blight based on a multi-scale attention hybrid head, characterized in that, Includes the following steps: S1: Obtain a dataset of potato late blight leaf images, label the diseased leaves in the potato late blight leaf images, and obtain a standardized labeled dataset; S2: Construct a potato late blight detection model, which includes a multi-scale feature extraction network, a multi-scale attention hybrid head, and a decoding prediction network; S3: The potato late blight detection model was trained using a standardized labeled dataset; S4: Input the potato leaf image to be tested into the trained potato late blight detection model. The potato late blight detection model outputs the location of the diseased leaves and the disease type on the potato leaf image to be tested. The multi-scale attention mixing head includes a preprocessing module, a multi-scale cross-attention module, and a cross-scale compression fusion module; The preprocessing module performs average pooling on the input feature maps X1, X2, and X3 respectively to generate feature maps f1, f2, and f3, and outputs them to the multi-scale cross-attention module. Channel concatenation, channel shuffling, and channel segmentation are then performed on the feature maps f1, f2, and f3 sequentially to obtain key vector k and value vector v. The key vector k is then subjected to dimensionality reshaping and vector normalization to obtain standard vector K. The value vector v is also subjected to dimensionality reshaping to obtain standard vector V. Standard vectors K and V are then output to the multi-scale cross-attention module. The multi-scale cross-attention module sequentially reshapes and normalizes the input feature maps f1, f2, and f3 to obtain standard vectors F1, F2, and F3. Based on the standard vectors F... i The attention weight A is calculated based on the similarity to the standard vector K. i For i=1~3, the attention weight A i With standard vector V and feature map f i The corresponding feature map O is obtained by fusion. i , feature map O i Output to the cross-scale compression and fusion module; The cross-scale compression fusion module performs three bilinear interpolation upsampling operations on the input feature map O1. The second bilinear interpolation upsampling outputs feature map C3, and the third bilinear interpolation upsampling outputs feature map C2. For feature map O1... j After performing one bilinear interpolation upsampling operation, the feature map X is compared with... j Perform element-wise addition, then sum to feature map C. j Perform channel splicing to output feature map T j For j=2~3, convolution is performed on feature maps T2 and T3 respectively, and the fused feature maps P1 and P2 are output to the decoding prediction network.

2. The method for detecting potato late blight based on a multi-scale attention hybrid head according to claim 1, characterized in that, The multi-scale feature extraction network is used to extract features layer by layer from the input potato leaf image to obtain semantic and detail information at different scales, and output feature maps X1, X2, and X3 to the multi-scale attention hybrid head. The multi-scale attention hybrid head is used to perform cross-channel cross-attention modeling and cross-scale feature fusion processing on feature maps X1, X2, and X3, and outputs the fused feature maps P1 and P2 to the decoding prediction network. The decoding prediction network is used to perform channel splicing and aggregation on feature maps P1 and P2, progressive feature extraction through multi-level decoding layers, and introduces self-distillation between adjacent decoding layers to achieve feature knowledge transfer optimization, outputting the location of diseased leaves and disease type on potato leaf images.

3. The method for detecting potato late blight based on a multi-scale attention hybrid head according to claim 1, characterized in that, According to the standard vector F i The attention weight A is calculated based on the similarity to the standard vector K. i The method is as follows: The standard vector F i The attention weights A are obtained by multiplying them element-wise by the standard vector K and then dividing by the scaling factor. i .

4. The method for detecting potato late blight based on a multi-scale attention hybrid head according to claim 1, characterized in that, The feature map O i The calculation formula is as follows: ， in, For instance normalization function, For exponential normalization function, This indicates element-wise multiplication. This indicates element-wise addition. For dimension reshaping functions.

5. The method for detecting potato late blight based on a multi-scale attention hybrid head according to claim 1, characterized in that, The method for annotating diseased leaves in potato late blight leaf images is as follows: manually outline the diseased leaves in potato late blight leaf images with polygonal boxes to mark the location of the diseased leaves and label the disease category.

6. The method for detecting potato late blight based on a multi-scale attention hybrid head according to claim 1, characterized in that, Step S1 also includes the following steps: performing image enhancement processing on each potato late blight leaf image in the standardized annotation dataset to generate more potato late blight leaf images and expand the standardized annotation dataset.

Images