Grape downy mildew unmanned aerial vehicle remote sensing detection method based on improved YOLOv11

Abstract

The invention discloses an improved YOLOv11-based grape downy mildew unmanned aerial vehicle remote sensing detection method, which comprises the following steps of (1) shooting a grape downy mildew image by using an unmanned aerial vehicle, establishing a data set and performing preprocessing; (2) designing a CUG joint method to optimize the image; marking the data, and dividing the data into a training set and a verification set; (3) constructing an improved YOLOv11 network model, and training the improved YOLOv11 network model; (4) carrying out unmanned aerial vehicle grape downy mildew real-time detection by utilizing the improved YOLOv11 network model; according to the method, the accurate attention and recognition capability on the disease area is enhanced, and the overall detection performance is greatly improved.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision and agricultural technology, and in particular to an unmanned aerial vehicle remote sensing detection method for grape downy mildew based on improved YOLOv11. Background Art

[0002] Grape downy mildew is a plant disease that poses a serious threat to the grape cultivation industry. It is mainly caused by downy mildew and often causes rot in grape leaves, fruits and branches, affecting yield and quality. Traditional disease detection methods mostly rely on manual observation and empirical judgment, which is not only inefficient but also easily affected by subjective factors. Therefore, the development of efficient and accurate automated detection technology has become an urgent need in grape cultivation management.

[0003] In recent years, with the development of drone technology, remote sensing detection based on drones has gradually become an important means of agricultural disease detection. Drones can cover a large area of ​​planting area in a short time and obtain high-resolution image data, which provides convenience for subsequent disease analysis. However, relying solely on image acquisition cannot effectively solve the problem of disease detection. How to extract useful features from the acquired images and accurately identify diseases has become the focus of research.

[0004] The rapid development of deep learning technology has brought new opportunities to the fields of image processing and object detection. In particular, the successful application of CNN and Transformer in image classification and object detection has provided the possibility for automatic detection of grape downy mildew. However, existing deep learning models still face challenges of insufficient detection accuracy and poor robustness when dealing with complex backgrounds, low contrast and subtle disease features. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a UAV remote sensing detection method for grape downy mildew based on improved YOLOv11, which improves the processing effect in the UAV scene by improving and optimizing image feature enhancement technology and deep learning models, thereby ensuring that the technology can accurately detect diseases in complex grape fields and provide technical support for the prevention and control of grape downy mildew.

[0006] Technical solution: The invention discloses a method for detecting grape downy mildew by drone remote sensing based on improved YOLOv11, comprising the following steps:

[0007] (1) Use drones to take images of grape downy mildew, establish a dataset, and perform preprocessing;

[0008] (2) Design a CUG joint method to optimize image processing; annotate the data and divide it into a training set and a validation set;

[0009] (3) An improved YOLOv11 network model was constructed and trained;

[0010] (4) Use the improved YOLOv11 network model to perform real-time detection of grape downy mildew using drones.

[0011] Furthermore, step (1) is specifically as follows: the 3840×2160 grape downy mildew image data is cut into two 2160×2160 images on the left and right.

[0012] Furthermore, step (2) is specifically as follows: using the CLAHE method to obtain the local dynamic range and improve the illumination; using the Unsharp Masking method to obtain the edge clarity; and using the Gamma method to correct the contrast.

[0013] Furthermore, the CLAHE algorithm converts the image from BGR to LAB space, processes only the L channel, and sets the parameter cliplimit = 2.0; the enhanced L channel is merged with the original A and B channels and converted back to BGR space to maintain the color information.

[0014] Furthermore, Unsharp Masking first performs Gaussian blur on the image, subtracts the blurred part from the original image, and obtains the high-frequency information of the edge. The formula is as follows:

[0015] Unsharp Masked Image=α×Original Image-β×Blurred Image+Offset

[0016] Among them, α and β control the ratio of the original image to the blurred image.

[0017] Furthermore, in gamma correction, the image contrast is further adjusted by creating a gamma lookup table. The formula is as follows:

[0018]

[0019] Among them, I in is the pixel value of the input image, I out It is the output pixel value after Gamma correction, γ is the Gamma value, which determines the contrast characteristics of the image.

[0020] Furthermore, in step (2), Labelimg is used to mark the affected areas, including the "early" and "late" states on the front side of the leaves and the "disease" state on the back side of the leaves, which are marked as 0, 1, and 2, respectively.

[0021] Furthermore, step (3) is specifically as follows: first, the AIFI module is introduced at the tail of Backbone. The AIFI module obtains the relationship between local disease features and global context by adaptively fusing multi-scale features; second, the custom CBSC module is introduced in the Neck part of the feature pyramid; the CBSC module includes: two parallel branches: the upper branch consists of a channel attention module and a spatial attention module, and the lower branch consists of spatial and channel reconstruction units.

[0022] Furthermore, the spatial reconstruction unit is as follows: First, the input size D is 40×40×384 and is first group normalized, and the calculation formula is:

[0023]

[0024] Among them, γ and β are trainable scaling and translation parameters; μ and σ are the mean and standard deviation respectively; ε is a constant to prevent division by zero;

[0025] Weight W γ Calculated by the weight distribution formula, expressed as:

[0026]

[0027] Where C is the number of channels;

[0028] Then, the spatial attention weight W = Gate(Sigmoid(W γ ·X out )) and applied to the reconstructed input:

[0029]

[0030] Among them, the threshold is set to 0.5;

[0031] Finally, the feature maps of different parts are combined by addition fusion and splicing to obtain the spatial reconstruction unit output X new .

[0032] Furthermore, the channel reconstruction unit is specifically as follows: first, X new It is divided into two parts. One part is processed by global convolution and point convolution, and then added to get the output Y1; the other part is processed by point convolution and the original input to get the union Y2. The convolution kernel size of the point convolution is 1×1, and the convolution kernel size of the global convolution is 3×3. Then the two outputs are pooled and normalized to get S1 and S2; finally, the output Y=S1Y1+S2Y2 is obtained.

[0033] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: a CUG joint method is designed to enable the grape leaf images collected by drones to better retain and enhance key structural information such as texture and edge in the preprocessing stage, thereby significantly improving the feature clarity and distinguishability of the diseased area, solving the problem that the features are not obvious and the diseased area is difficult to extract under the traditional method; the AIFI module is introduced at the tail of Backbone, and the self-attention mechanism is used to establish an efficient association between local disease features and global context in complex backgrounds. This mechanism can highlight key areas and suppress background interference in complex environments, especially in scenes with densely distributed leaves, significantly improving the integrity and expression ability of feature extraction and improving the robustness of the model; the feature pyramid (Neck) part designs and applies a two-dimensional hybrid module CBSC, which effectively balances the expression ability of details and semantic information through the combination of channel and spatial attention, and the upper and lower branch structures reconstructed by spatial and channel information, avoiding the interference of useless background on disease features. In scenes with dense leaf distribution, the CBSC module significantly reduces the missed detection rate and false detection rate, while strengthening the precise attention and recognition ability of the diseased area, greatly improving the overall detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of the present invention;

[0035] Figure 2 It is a structural diagram of the CUG joint optimization method of the present invention;

[0036] Figure 3 It is a structural diagram of the improved YOLOv11 of the present invention;

[0037] Figure 4 It is the internal structure diagram of AIFI of the present invention;

[0038] Figure 5 is a structural diagram of the CBSC of the present invention;

[0039] Figure 6 It is a trend graph of the performance index of the improved YOLOv11 of the present invention versus the number of iterations;

[0040] Figure 7 It is the YOLOv11 model detection diagram of the present invention;

[0041] Figure 8 Correct detection graph for the improved YOLOv11 implementation of the present invention. DETAILED DESCRIPTION

[0042] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0043] like Figure 1As shown, the embodiment of the present invention provides a UAV remote sensing detection method for grape downy mildew based on improved YOLOv11, comprising the following steps:

[0044] S1. Data acquisition: Use a drone equipped with a high-definition camera to shoot grape planting areas to obtain grape downy mildew image data. The drone's flight height is about 2-3 meters, and the shooting method is vertical. During the collection process, keep the drone in stable flight to avoid image blur due to shaking.

[0045] S2. Data preprocessing: Image slicing is performed based on the grape downy mildew image data of a fixed size. Specifically, the 3840×2160 grape downy mildew image data is cut into two 2160×2160 blocks on the left and right. This slicing method has many advantages: First, it is easier to label the image after segmentation, making the details and features of each area easier to identify and process. Secondly, smaller image blocks can improve the graphics card's cache efficiency for images, reduce memory usage, and thus speed up the subsequent training and inference of deep learning models. In addition, this slicing method can effectively deal with the performance bottleneck of large images in the processing process, ensuring that the model has a higher response speed and better results during training and application. The CUG joint method is proposed to optimize the image. As Figure 2 As shown in the figure, it is composed of CLAHE, Unsharp Masking and Gamma Correction in series. The CLAHE algorithm improves the image performance under complex lighting conditions by enhancing the local dynamic range; Unsharp Masking improves edge clarity and strengthens the contrast between leaves and diseased areas; Gamma Correction further highlights the detailed features of the disease by adjusting the brightness and contrast of the image.

[0046] S21, CLAHE algorithm converts the image from BGR to LAB color space, where the L channel represents brightness and the A and B channels represent color information. Only the L channel is processed, and the local contrast is improved by equalizing the brightness histogram in a small area, and the parameter cliplimit = 2.0 is set to avoid excessive enhancement of noise. The enhanced L channel is then merged with the original A and B channels to maintain the original color information. Finally, it is converted to BGR space.

[0047] S22, Unsharp Masking First, Gaussian blur the image and subtract the blurred part from the original image to highlight high-frequency information such as edges. Set α to 1.5 and β to 0.5.

[0048] Unsharp Masked Image=α×Original Image-β×Blurred Image+Offset#(1)

[0049] S23. In gamma correction, the image contrast is further adjusted by creating a gamma lookup table.

[0050]

[0051] I in is the pixel value of the input image, I out It is the output pixel value after Gamma correction, γ is the Gamma value, which determines the contrast characteristics of the image.

[0052] Labelimg was used to annotate the affected areas, and the annotation content was divided into three categories: two different degrees of "early" and "late" observed from the front of the leaf, and the "disease" state observed from the back of the leaf. And marked as 0, 1, and 2 in order. The 860 data sets were further divided into training sets, validation sets, and test sets. 700 image blocks containing grape downy mildew were selected as training sets, 80 as validation sets, and another 80 as test sets. The number of annotation boxes was 8060, 952, and 854 respectively to ensure the scientific nature of model training and evaluation. The offline data enhancement used was to translate, rotate, and flip the original training images horizontally or vertically. Only by training the detection network can rich angle information of the target be obtained, which enhances the generalization ability of the deep learning model. After data enhancement, 4206 training images were obtained.

[0053] S3. Improvement of YOLOv11 model: Figure 3 As shown, the AIFI module is introduced at the end of Backbone. Figure 4 It can be seen that this module improves the model's ability to capture the relationship between local disease features and global context in complex backgrounds through adaptive fusion of multi-scale features, and improves the feature extraction effect under densely distributed leaves. The module focuses on the internal embedded multi-head self-attention mechanism, determines num_heads = 8, and uses the convolution kernel size of 3×3 to transform the input feature 20×20×256 linearly to generate multiple query (Q), key (K) and value (V) vectors to capture the relationship between different features. For each attention head, the attention weight is calculated, and the formula is:

[0054]

[0055] Among them, d k is the dimension of the key. Finally, the outputs of each head are concatenated, and the final output is obtained through linear transformation with the same size of 20◇20◇256 semantic information.

[0056] like Figure 5As shown in the figure, the CBSC module is designed and applied in the feature pyramid (Neck) part, which helps to balance the details and semantic information and avoid the interference of useless background information, making the model more accurate in the fusion process, paying more attention to densely distributed leaves, and reducing the missed detection rate and false detection rate. The CBSC module is divided into two parallel branches: the upper branch consists of the channel attention module and the spatial attention module, and the lower branch consists of the spatial and channel reconstruction units.

[0057] Steps for branching:

[0058] a: Channel attention module

[0059] The input feature F of size 40×40×384 is subjected to GAP and GMP operations to obtain an output of 1×1×384, which is then passed through two layers of MLP and added together and activated by Sigmoid to generate channel attention weights, which are finally multiplied element-wise with the original F.

[0060] F c =Sigmoid(MLP(GAP(F))+MLP(GMP(F)))⊙F#(4)

[0061] b: Spatial attention module

[0062] F c Perform GAP and GMP operations to obtain a 40×40×1 feature map, which is then concatenated and passed through a 3×3 convolution to generate spatial attention weights. After Sigmoid activation, it is combined with F c Multiply element by element to get the spatially enhanced feature map F cs .

[0063] F cs =Sigmoid(Conv2D(Concat(GAP(F c ),GMP(F c ))))⊙F c #(5)

[0064] Steps for branching:

[0065] c: Space reconstruction unit

[0066] The input F with size 40×40×384 is first group normalized, and the calculation formula is:

[0067]

[0068] where γ and β are trainable scaling and translation parameters; μ and σ are the mean and standard deviation, respectively; and ε is a constant to prevent division by zero.

[0069] Next, the weight W γCalculated by the weight distribution formula, expressed as:

[0070]

[0071] Where C is the number of channels.

[0072] Then, the spatial attention weight W = Gate(Sigmoid(W γ ·X out )) and applied to the reconstructed input.

[0073]

[0074] Among them, the threshold is set to 0.5.

[0075] Next, we will combine the feature maps of different parts by adding fusion and splicing to obtain the spatial reconstruction unit output X new .

[0076] d: Channel reconstruction unit

[0077] First, X new It is divided into two parts. One part is processed by global convolution and point convolution, and then added to get the output Y1; the other part is processed by point convolution and the original input to get the union Y2. The convolution kernel size of the point convolution is 1×1, and the convolution kernel size of the global convolution is 3×3. Then the two outputs are pooled and normalized to get S1 and S2. Finally, the output Y=S1Y1+S2Y2 is obtained.

[0078] Finally, the results of the upper and lower branches, F cs The final output of the CBSC module is obtained by adding Y element by element and passing through the activation function.

[0079] F CBSC =Sigmoid(α1F cs +α2Y)#(9)

[0080] S4. Model training: Confirm that the training environment is configured as operating system: Windows 11professional 64-bits, CPU: IntelCore i9-14900K, GPU: NVIDIA GeForce RTX 4090, CUDA 12.6, PyTorch2.5.0, Python 3.10.

[0081] The model hyperparameters are adjusted to set the input image size to 640×640, the number of iterations to 300, the batch size to 32, the category of the detected object to 3, the optimizer to stochastic gradient descent, the learning rate to 0.01, the momentum term to 0.937, and the weight decay to 0.0005.

[0082] Determination of performance indicators: Precision refers to the proportion of positive samples predicted by the model to those that are actually positive, which measures the accuracy of the prediction. Recall refers to the proportion of positive samples correctly predicted to be positive among actual positive samples, reflecting the recall ability of the model. mAP50 is the average accuracy calculated by comprehensively considering Precision and Recall when the intersection-over-union ratio threshold is 0.5, and is one of the key indicators for evaluating the performance of target detection models. Parameters refers to the total number of trainable parameters in the model, reflecting the complexity and capacity of the model.

[0083] S5. Experimental Results and Analysis

[0084] according to Figure 6 It shows that the model is set for 300 iterations and converges at iteration 223. The loss curve tends to be stable. The precision curve, recall curve and average precision curve of the model show a stable trend, indicating that the model has basically converged.

[0085] In order to test the effect of the CUG joint optimization method on the initial model, the present invention made the following comparison results, as shown in Table 1:

[0086] Table 1

[0087]

[0088]

[0089] The experimental results show that the drone grape leaf image reconstructed using the CUG joint optimization algorithm can make the original drone grape leaf image have more texture, edge and other structural information, which greatly improves the detection results.

[0090] In order to test the effect of each module on the model improvement, the present invention conducted the following ablation experiments, and the experimental results are shown in Table 2:

[0091] Table 2

[0092]

[0093] Note: √ indicates that the module is used

[0094] 1. Introducing the AIFI module: Precision dropped from 0.831 to 0.778, but Recall increased to 0.832 and mAP50 increased to 0.848, indicating that AIFI improved the detection coverage but may have increased false positives.

[0095] 2. Introducing the CBSC module: Precision slightly dropped to 0.818, Recall was 0.778, and Recall was 0.832. The effect was not as good as AIFI, and it mainly optimized spatial features.

[0096] 3. Joint introduction of AIFI and CBSC: Precision is 0.806, Recall is 0.809, and mAP50 reaches the highest 0.857, showing a balance between precision and recall, significantly improving the overall detection performance, thereby improving missed detection and false detection problems.

[0097] In order to demonstrate the effect of the invention, Figure 7 and Figure 8 Description: Compare the detection image of the improved model with the detection image of the original model. Figure 7 Original model detection diagram, Figure 8 Improved model detection graph. The red marks in the original model detection graph represent the false detection and missed detection compared with the improved model detection graph. When AIFI and CBSC are added at the same time, the detection results are the best.

Claims

1. A UAV remote sensing detection method for grape downy mildew based on improved YOLOv11, characterized in that: Includes the following steps: (1) Use drones to take images of grape downy mildew, establish a dataset, and perform preprocessing; (2) Design a CUG joint method to optimize image processing; annotate the data and divide it into a training set and a validation set; (3) An improved YOLOv11 network model was constructed and trained; (4) Use the improved YOLOv11 network model to perform real-time detection of grape downy mildew using drones.

2. The method for detecting grape downy mildew by drone remote sensing based on improved YOLOv11 according to claim 1, characterized in that: Step (1) is as follows: the 3840×2160 grape downy mildew image data is cut into two 2160×2160 images on the left and right.

3. The method for detecting grape downy mildew by drone remote sensing based on improved YOLOv11 according to claim 1, characterized in that: Step (2) is as follows: using the CLAHE method to obtain the local dynamic range and improve the illumination; using the UnsharpMasking method to obtain edge clarity; and using the Gamma method to correct the contrast.

4. The method for detecting grape downy mildew by drone remote sensing based on improved YOLOv11 according to claim 3, characterized in that: The CLAHE algorithm converts the image from BGR to LAB space, processes only the L channel, and sets the parameter cliplimit = 2.0; the enhanced L channel is merged with the original A and B channels and converted back to BGR space to maintain the color information.

5. The method for detecting grape downy mildew by drone remote sensing based on improved YOLOv11 according to claim 3, characterized in that: Unsharp Masking first performs Gaussian blur on the image, subtracts the blurred part from the original image, and obtains the high-frequency edge information. The formula is as follows: Unsharp MaskedImage=α×Original Image-β×BlurredImage+Offset Among them, α and β control the ratio of the original image to the blurred image.

6. The method for detecting grape downy mildew by drone remote sensing based on improved YOLOv11 according to claim 3, characterized in that: In gamma correction, the image contrast is further adjusted by creating a gamma lookup table. The formula is as follows: Among them, I in is the pixel value of the input image, I out It is the output pixel value after Gamma correction, γ is the Gamma value, which determines the contrast characteristics of the image.

7. The method for detecting grape downy mildew by drone remote sensing based on improved YOLOv11 according to claim 1, characterized in that: In step (2), Labelimg is used to mark the affected areas, including the "early" and "late" states on the front of the leaves and the "disease" state on the back of the leaves, which are marked as 0, 1, and 2 respectively.

8. The method for detecting grape downy mildew by drone remote sensing based on improved YOLOv11 according to claim 1, characterized in that: Step (3) is as follows: First, the AIFI module is introduced at the end of Backbone. The AIFI module obtains the relationship between local disease features and global context by adaptively fusing multi-scale features; Secondly, a custom CBSC module is introduced in the Neck part of the feature pyramid; The CBSC module consists of two parallel branches: the upper branch consists of a channel attention module and a spatial attention module, and the lower branch consists of spatial and channel reconstruction units.

9. The method for detecting grape downy mildew by drone remote sensing based on improved YOLOv11 according to claim 8, characterized in that: The spatial reconstruction unit is as follows: First, the input size of F is 40×40×384, and it is first group normalized, and the calculation formula is: Among them, γ and β are trainable scaling and translation parameters; μ and σ are the mean and standard deviation respectively; ε is a constant to prevent division by zero; Weight W γ Calculated by the weight distribution formula, expressed as: Where C is the number of channels; Then, the spatial attention weight W = Gate(Sigmoid(W γ ·X out )) and applied to the reconstructed input: Among them, the threshold is set to 0.5; Finally, the feature maps of different parts are combined by addition fusion and splicing to obtain the spatial reconstruction unit output X new .

10. The method for detecting grape downy mildew by drone remote sensing based on improved YOLOv11 according to claim 8, characterized in that: The channel reconstruction unit is as follows: First, X new It is divided into two parts. One part is processed by global convolution and point convolution, and then added to get the output Y1; the other part is processed by point convolution and the original input to get the union Y2. The convolution kernel size of the point convolution is 1×1, and the convolution kernel size of the global convolution is 3×3. Then the two outputs are pooled and normalized to get S1 and S2; finally, the output Y=S1Y1+S2Y2 is obtained.

Images