Method for detecting diseases and insect pests of tea leaves in complex environment based on improved YOLOv11 model

Abstract

The invention provides an improved YOLOv11 model-based tea plant disease and insect pest detection method in a complex environment, which comprises the following steps of: acquiring an original image through field shooting, performing data amplification by using a data enhancement technology, and realizing bounding box labeling by using a labeling tool label Img to form a training data set; the method comprises the following steps: constructing a network model of tea plant diseases and insect pests, improving an original model structure of YOLOv11n, and improving a Head module by using an MSDAF module; improving a loss function; performing end-to-end training on the network architecture by using the training set and the test set, and adjusting hyper-parameters; and performing field detection on the tea leaf scab area by using the trained network model. The invention provides a novel tea disease and insect pest detection algorithm based on a YOLOv11 target detection algorithm, and the average precision of tea disease and insect pest detection in a complex environment can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the intersection of computer vision and agricultural intelligence, and in particular relates to a method for detecting tea plant diseases and insect pests in complex environments based on an improved YOLOv11 model. Background Art

[0002] Tea is an important economic crop, and pest and disease control has always been a challenge in tea plantation ecological management, seriously affecting tea quality and yield. Existing manual pest and disease control methods rely on visually observing pest and disease characteristics on the surface of tea leaves to identify pest species. This presents challenges and is inefficient, hindering efficient ecological management of modern tea plantations. Therefore, tea plantation pest and disease detection is essential.

[0003] In recent years, research on target detection methods for tea diseases and pests has also been innovative.

[0004] For example, a Chinese patent with the existing publication number (CN119206521A) proposes a deep learning-based tea pest and disease detection method. This method uses RGB and infrared dual sensors to collaboratively collect multispectral image data of tea leaves, uses a planned perceptual distortion template to perform adaptive pixel offset processing on low-bit images, generates multiple descriptive features through a quantizer to enhance image representation, and inputs a selective state space model for training and optimization. This method then detects tea pests and diseases and determines their categories. The drawback of this existing technology is that the traditional CNN architecture has difficulty capturing long-range spatial correlations in complex backgrounds (such as occlusion by branches and leaves, and mottled soil), and multispectral data fusion suffers from alignment errors and noise interference.

[0005] Another example is a Chinese patent application with publication number (CN117576564A), which describes a method for identifying and warning tea plant pests and diseases. This method uses a multispectral sensor network to integrate visible light and infrared images to obtain a set of tea growth image information. This information is then preprocessed and subjected to multi-level feature recognition to improve pest and disease identification accuracy and early warning efficiency. However, this prior art approach is limited by bottlenecks such as poor adaptability to field environments, high hardware costs, and insufficient model generalization capabilities. Summary of the Invention

[0006] This paper addresses the issues of existing tea pest and disease detection algorithms, which suffer from low detection accuracy and high computational complexity in complex environments. Based on the YOLOv11 object detection algorithm, this new tea pest and disease detection algorithm is proposed. This algorithm effectively improves the mean average precision (mAP) of tea pest and disease detection in complex environments. Compared with the existing YOLO algorithm, the improved YOLO algorithm not only significantly improves accuracy and detection performance, but also reduces the number of model parameters and computational complexity.

[0007] The present invention provides the following technical solutions:

[0008] A method for detecting tea plant diseases and insect pests in a complex environment based on an improved YOLOv11 model includes the following steps:

[0009] S1. Obtain original images through field photography, use data augmentation technology to amplify the data, and use the labelImg annotation tool to implement bounding box annotation to form a training dataset.

[0010] S2. Build a network model for tea pests and diseases, improve the original YOLOv11n model structure, which includes the Backbone module and the Head module. Use the MSDAF module to improve the Head module; improve the loss function.

[0011] (1) This paper proposes a tea pest and disease detection method in a complex environment based on an improved YOLOv11 architecture, and designs a multi-scale dual-path attention fusion module (MSDAF). Figure 2 As shown in the figure. To solve the problem that the features of small target spots are easily submerged by the background, multi-level feature extraction is performed to obtain feature tensors F at different levels from low to high, namely F2, F3, F4 and F5. Then, the feature tensor F is obtained by combining the output of the channel-space dual-path attention mechanism and the step-by-step multiplication operation. M ; Then use step-by-step addition to get the feature tensor F of the dual-path attention integration A ;Finally, the feature tensor F and the feature tensor F M The concatenation is performed to combine the attention features with the convolution features.

[0012] CA(F)=Sigmoid(Layer(AO(F))+Layer(MP(F)))

[0013] CA(F)=Sigmoid(Layer(AP(F))+Layer(MP(F)))

[0014]

[0015] Fout =Concat(F, F M )

[0016] Among them, CA represents the channel attention mechanism; F c Represents the feature tensor output by channel attention; SA represents the spatial attention mechanism; F s Represents the feature tensor output by spatial attention; MP represents maximum pooling; F M Represents the feature tensor after maximum pooling; AP represents average pooling; F A Represents the feature tensor after average pooling; Indicates step-by-step addition; Represents step-by-step multiplication; Layer represents a single hidden layer multilayer perceptron; Concat represents feature tensor combination; Conv represents two-dimensional convolution; Sigmoid represents activation function; F out The feature tensor representing the final output.

[0017] (2) Spatial attention path: Improved atrous convolutional pyramid (ASPP), using atrous convolution to capture the edge and texture features of lesions at different scales. The output feature fusion formula is as follows:

[0018] F spatial =Conv 1×1 ([DConv r＝1 (F), DConv r＝3 (F), DConv r＝5 (F)])

[0019] Where DConvr=k represents the dilation rate of the dilation convolution k, and [·] represents the concatenation of feature maps along the channel dimension.

[0020] (3) Channel attention path: Combine global average pooling (GAP) and global maximum pooling (GMP) to generate channel weights, which are expressed as follows:

[0021] W channel =Sigmoid(MLP(GAP(F))+MLP(GMP(F)))

[0022] Among them, W channel represents channel weight; GAP represents global average pooling, GMP represents global maximum pooling; MLP represents a two-layer fully connected network. Adding the gated fusion mechanism, the expression for dynamically balancing space and channel paths is:

[0023] F out =Sigmoid(W g )⊙F spatial +(1-Sigmoid(W g ))⊙Fchannel

[0024] Among them, F channel represents the feature tensor passed through the channel attention mechanism; Wg represents the gating weight, which is learned from the concatenated features [Fspatial, Fchannel] through a fully connected layer and mapped to the range [0, 1]. Compared with traditional attention mechanisms such as SE and CBAM, the computational overhead is reduced, but the mean average performance (MAP) is improved.

[0025] (4) Improve the model loss function. The traditional IoU loss is not good at distinguishing overlapping lesions, and the weights of difficult and easy samples are fixed. The dynamic focus WIoU loss function (DFWIoU) can improve the optimization of the detection head and adaptively adjust the sample weights according to the overlap (IoU) between the predicted box and the real box, such as Figure 3 As shown, the formula is:

[0026]

[0027] Where w represents the sample weight; IoU represents the intersection over union ratio. The weight of high IoU samples (easy samples) is reduced, and the weight of low IoU samples (difficult samples) is increased. ∈ represents the minimum value (1e-5) to prevent the denominator from being zero. The gradient alignment constraint of the lesion edge is introduced to suppress background false detection.

[0028] Add a boundary-sensitive penalty term, introduce the gradient information of the lesion edge, and suppress background false detection. The formula is:

[0029]

[0030] in, represents the Sobel gradient of the real lesion, Represents the gradient of the prediction area.

[0031] Finally, the total loss function formula is:

[0032] L total =L DFWIoU +0.5·L edge +L prune

[0033] Among them, L total Represents the total loss function; L DFWIoU Represents the dynamic focus WIoU loss function, based on weighted IoU calculation, L edge represents the boundary-sensitive loss function, with a weight coefficient of 0.5 determined through grid search, and Lprune represents the pruning regularization loss function. In dense lesion scenarios, DFWIoU improves the discrimination of overlapping lesions.

[0034] S3. Use the training set and test set obtained in step 1 to perform end-to-end training through the network architecture designed in step 2, and adjust the hyperparameters.

[0035] S3.1 Input the training set data, output the fused feature map through the network model, and save the weight file;

[0036] The S3.2 network model uses pre-trained weights to infer the validation set and obtain preliminary performance indicators;

[0037] S3.3 Based on the performance indicators, the model hyperparameters are adjusted and the optimal weights are obtained after multiple rounds of iterations;

[0038] S3.4 uses the optimal weights to test the model generalization ability on the test set to verify the model effectiveness and deployment feasibility.

[0039] S4. Use the trained network model to conduct field detection of tea leaf diseased areas.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1.) The Multi-Scale Dual-Path Attention Fusion (MSDAF) module, combined with a channel-spatial attention mechanism, effectively captures local details and long-range spatial correlations of small target lesions. This addresses the feature overwhelm of traditional CNNs in complex backgrounds such as foliage occlusion and soil mottle. This improves the recall rate of small targets and significantly increases mean average precision (MAP). Furthermore, the MSDAF module dynamically balances spatial and channel features through a gated fusion mechanism. Compared to traditional attention mechanisms (such as SE and CBAM), this module reduces computational overhead and reduces both model parameter count and computational complexity.

[0042] 2.) The improved atrous convolutional pyramid (ASPP) enhances the ability to extract edge and texture features of lesions at different scales and reduces alignment errors and noise interference in multispectral data fusion.

[0043] 3.) A dynamic focus WIoU loss function (DFWIoU) is proposed. It adaptively adjusts sample weights based on the intersection-over-union (IoU) ratio, reducing the weight of easy samples (high IoU) and increasing the weight of difficult samples (low IoU), thereby improving the discrimination of overlapping lesions. A boundary-sensitive penalty term (Ledge) is introduced to align the edges of the real and predicted lesions using the Sobel gradient, reducing false detections in complex backgrounds. The network architecture incorporates a pruning regularization loss (Lprune) to further reduce the model size and improve deployment feasibility.

[0044] 4. Image stitching and resizing simulate real-world occlusion scenarios, alleviating the generalization issues inherent in existing technologies due to limited data. Single-modal RGB image input eliminates the need for multispectral sensors, reducing hardware deployment costs and addressing the high cost bottleneck of existing multimodal fusion solutions. The end-to-end training framework supports fast inference of lightweight models on edge devices, meeting the real-time detection requirements of tea plantations. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flowchart provided for an embodiment of the present invention.

[0046] Figure 2 Schematic diagram of the multi-scale dual-path attention fusion module (MSDAF) provided in an embodiment of the present invention.

[0047] Figure 3 Schematic diagram of the dynamic focus WIoU loss function (DFWIoU) provided in an embodiment of the present invention.

[0048] Note: Wt, Ht represent the width and height of the overlapping part of the target frame and the prediction frame; w, h represent the width and height of the prediction frame; Wgt, hgt represent the width and height of the target frame; (x, y) represents the center coordinates of the prediction frame; (x gt ,y gt ) represents the center coordinates of the target box; Wg, Hg represent the width and height of the minimum outer frame that contains both the prediction box and the target box.

[0049] Figure 4 This is a diagram of the overall architecture of the improved YOLOv11n provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention is described in detail below with reference to the embodiments and accompanying drawings. However, it should be understood that the embodiments and accompanying drawings are merely exemplary descriptions of the present invention and do not constitute any limitation on the scope of protection of the present invention. All reasonable variations and combinations within the scope of the inventive concept of the present invention fall within the scope of protection of the present invention.

[0051] The present invention will be further described below with reference to the accompanying drawings.

[0052] Example 1

[0053] Flowchart as Figure 1 As shown:

[0054] S1: Collect images of tea garden pests and diseases, perform denoising and standardization preprocessing, annotate the images and perform data enhancement to obtain a dataset (training set / validation set / test set = 7:2:1), as follows:

[0055] S1.1 collects images of five common tea plantation pests and diseases (tea aphids, tea brown spot, tea black rot, tea rust, and tea white spot), covering scenes with different lighting conditions and growth stages.

[0056] S1.2 screens the image quality, performs blur detection, frequency domain ghosting elimination, and similarity filtering to optimize data;

[0057] S1.3 Use the LabelImg tool to label the lesion area. The annotation file format is YOLO standard. Ensure that the bounding box accurately covers the edge of the lesion.

[0058] S1.4 randomly selects images for stitching, adjusts the size, and simulates a complex field background;

[0059] S1.5 illumination normalization, local contrast enhancement of the image, random noise injection, Gaussian noise and salt and pepper noise are added to simulate rainy days or equipment noise interference to generate diversity enhanced samples.

[0060] S2: Build a network model for tea pests and diseases, improving the original YOLOv11n model structure, which includes the Backbone module and the Head module. The Head module is improved using the MSDAF module; and the loss function is improved. Specifically, the following steps are included:

[0061] S2.1 constructs a multi-scale dual-path attention fusion module (MSDAF), such as Figure 2 As shown. By performing multi-level feature extraction on the feature information, we can obtain the feature tensors F of different levels from low to high levels, namely F2, F3, F4 and F5. Then, we can obtain the feature tensor F by combining the output of the channel-space dual-path attention mechanism and the step-by-step multiplication operation. M ; Then use step-by-step addition to get the feature tensor F of the dual-path attention integration A ;Finally, the feature tensor F and the feature tensor F M The concatenation is performed to combine the attention features with the convolution features.

[0062] CA(F)=Sigmoid(Layer(AP(F))+Layer(MP(F)))

[0063] SA(F)=Sigmoid(Conv(Concat(AP(F),MP(F))))

[0064]

[0065] F out =Concat(F, F M )

[0066] Among them, CA represents the channel attention mechanism; F c Represents the feature tensor output by channel attention; SA represents the spatial attention mechanism; F s Represents the feature tensor output by spatial attention; MP represents maximum pooling; F M Represents the feature tensor after maximum pooling; AP represents average pooling; F A Represents the feature tensor after average pooling; Indicates step-by-step addition; Represents step-by-step multiplication; Layer represents a single hidden layer multilayer perceptron; Concat represents feature tensor combination; Conv represents two-dimensional convolution; Sigmoid represents activation function; F out The feature tensor representing the final output.

[0067] This improved method effectively captures the local details and long-distance spatial correlations of small target lesions, solves the problem of feature submersion of traditional CNN in complex backgrounds such as branch and leaf occlusion and soil mottledness, improves the recall rate of small targets, and significantly improves the average precision (mAP), while reducing the number of model parameters and computational complexity.

[0068] S2.2 improves the atrous convolutional pyramid (ASPP) and uses atrous convolution to capture the edge and texture features of lesions at different scales. The output feature fusion formula is as follows:

[0069] F spatial =Conv 1×1 ([DConv r＝1 (F), DConv r＝3 (F), DConv r＝5 (F)])

[0070] Among them, F spatial represents the feature tensor passed through the spatial attention mechanism; DConvr=k represents the dilated convolution with a dilation rate of k, and [·] represents the concatenation of feature maps along the channel dimension.

[0071] This improvement can enhance the ability to extract edge and texture features of lesions at different scales and reduce alignment errors and noise interference in multispectral data fusion.

[0072] S2.3 combines global average pooling (GAP) with global maximum pooling (GMP) to generate channel weights, which are expressed as follows:

[0073] W channel =Sigmoid(MLP(GAP(F))+MLP(GMP(F)))

[0074] Among them, W channelrepresents channel weight; GAP represents global average pooling, GMP represents global maximum pooling; MLP represents a two-layer fully connected network. Adding the gated fusion mechanism, the expression for dynamically balancing space and channel paths is:

[0075] F out =Sigmoid(W g )⊙F patial +(1-Sigmoid(W g ))⊙F channel

[0076] Among them, F channel represents the feature tensor passed through the channel attention mechanism; Wg represents the gating weight, which is learned from the concatenated features [Fspatial, Fchannel] through a fully connected layer and mapped to the range [0, 1]. Compared with traditional attention mechanisms such as SE and CBAM, the computational overhead is reduced, but the mean average performance (MAP) is improved.

[0077] S2.4 Improve the model loss function. Using the dynamic focus WIoU loss function (DFWIoU) can improve the optimization of the detection head and adaptively adjust the sample weight according to the overlap (IoU) between the predicted box and the real box, such as Figure 3 As shown, the expression is:

[0078]

[0079] Where w represents the sample weight; IoU represents the intersection over union ratio. The weight of high IoU samples (easy samples) is reduced, and the weight of low IoU samples (difficult samples) is increased. ∈ represents the minimum value (1e-5) to prevent the denominator from being zero. The gradient alignment constraint of the lesion edge is introduced to suppress background false detection.

[0080] Add a boundary-sensitive penalty term, introduce the gradient information of the lesion edge, and suppress background false detection. The formula is:

[0081]

[0082] in, represents the Sobel gradient of the real lesion, Represents the gradient of the prediction area.

[0083] Finally, the total loss function formula is:

[0084] L total =L DFWIoU +0.5·L edge +L prune

[0085] Among them, L total Represents the total loss function; L DFWIoURepresents the dynamic focus WIoU loss function, based on weighted IoU calculation, L edge represents the boundary sensitive loss function, the weight coefficient 0.5 is determined by grid search, L prune represents the pruning regularization loss function.

[0086] The loss function improvement provided in this step makes up for the problem that the traditional IoU loss has insufficient discrimination of overlapping lesions and the weight of difficult and easy samples is fixed. It can improve the discrimination of overlapping lesions in dense lesion scenes, making the subsequent target frame and prediction frame more accurate.

[0087] In step 2, the Backbone module converts the raw data of the input training set into feature maps of different semantic levels through downsampling. The Backbone module includes the Conv module, C3K2 module, SPPF module and C2PSA module. The specific structure is as follows: Figure 4 As shown;

[0088] The Head module adds the MSDAF module, and comprehensively utilizes the feature information of different scales output by the Backbone module through multi-level feature extraction and channel-space dual-path attention mechanism, as well as feature fusion of each layer, and outputs feature information. Then, the detection head with a decoupled head structure is used to perform target detection on the output feature information, predict the bounding box and its category and confidence.

[0089] S3: Use the training set and test set obtained in step 1 to perform end-to-end training through the network architecture designed in step 2, and adjust the hyperparameters.

[0090] S3.1 Input the training set data, output the fused feature map through the network model, and save the weight file;

[0091] The S3.2 network model uses pre-trained weights to infer the validation set and obtain preliminary performance indicators;

[0092] S3.3 Based on the performance indicators, the model hyperparameters are adjusted and the optimal weights are obtained after multiple rounds of iterations;

[0093] S3.4 uses the optimal weights to test the model generalization ability on the test set to verify the model effectiveness and deployment feasibility.

[0094] S4: Use the trained network model to conduct field detection of tea leaf diseased areas.

[0095] The above embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions that fall within the scope of protection of the present invention are within the scope of protection of the present invention. It should be noted that improvements and modifications that can be made by a person skilled in the art without departing from the principles of the present invention are also considered to be within the scope of protection of the present invention.

Claims

1. A method for detecting tea pests and diseases in complex environments based on an improved YOLOv11 model, characterized in that: The following steps are involved: S1. Obtain original images through field photography, amplify the data using data augmentation techniques, and annotate bounding boxes using the labelImg annotation tool to form a training dataset. S2. Build a network model for tea pests and diseases, improve the original YOLOv11n model structure, use the MSDAF module to improve the Head module, and improve the loss function; S3. Use the training and test sets obtained in step 1 to perform end-to-end training on the network architecture designed in step 2, and adjust the hyperparameters. S4. Use the trained network model to conduct field detection of tea leaf diseased areas.

2. The method for detecting tea plant diseases and insect pests in a complex environment based on the improved YOLOv11 model according to claim 1, wherein S2 include: A multi-scale dual-path attention fusion module (MSDAF) is designed to address the problem that small target lesion features are easily submerged by the background. Multi-level feature extraction is performed to obtain feature tensors F at different levels from low to high, namely F2, F3, F4, and F5. Then, the feature tensor F is obtained by combining the output of the channel-space dual-path attention mechanism and the stage-by-stage multiplication operation. M ; Then use step-by-step addition to get the feature tensor F of the dual-path attention integration A ;Finally, the feature tensor F and the feature tensor F M Perform concatenation to combine attention features with convolution features; CA(F)=Sigmoid(Layer(AP(F))+Layer(MP(F))) SA(F)=Sigmoid(Conv(Concat(AP(F),MP(F)))) F out =Concat(F,F M ) Among them, CA represents the channel attention mechanism; F c Represents the feature tensor output by channel attention; SA represents the spatial attention mechanism; F s Represents the feature tensor output by spatial attention; MP represents maximum pooling; F M Represents the feature tensor after maximum pooling; AP represents average pooling; F A Represents the feature tensor after average pooling; Indicates step-by-step addition; Represents step-by-step multiplication; Layer represents a single hidden layer multilayer perceptron; Concat represents feature tensor combination; Conv represents two-dimensional convolution; Sigmoid represents activation function; F out The feature tensor representing the final output.

3. The method for detecting tea plant diseases and insect pests in a complex environment based on the improved YOLOv11 model according to claim 2, characterized in that: Spatial attention path in S2: Improve the dilated convolutional pyramid ASPP and use dilated convolution to capture the edge and texture features of lesions at different scales. The output feature fusion formula is as follows: F spatial =Conv 1×1 ([DConv r＝1 (F),DConv r＝3 (F),DConv r＝5 (F)]D) Among them, F saptial Represents the feature tensor through the spatial attention mechanism; DConvr=k represents the hollow convolution with a dilation rate of k, and [] represents the concatenation of feature maps along the channel dimension.

4. The method for detecting tea plant diseases and insect pests in a complex environment based on the improved YOLOv11 model according to claim 3, characterized in that: Channel attention path in S2: Combine global average pooling GAP and global maximum pooling GMP to generate channel weights, expressed as follows: W channel =Sigmoid(MLP(GAP(F))+MLP(GMP(F))) Among them, W channel Represents channel weight; GAP represents global average pooling, GMP represents global maximum pooling; MLP represents a two-layer fully connected network; adding a gated fusion mechanism, the expression for dynamically balancing space and channel paths is: F out =Sigmoid(W g )☉F spatial +(1-Sigmoid(W g ))☉F channel Among them, F channel Represents the feature tensor through the channel attention mechanism; Wg represents the gating weight, which is learned from the splicing features [Fspatial, Fchannel] through the fully connected layer and maps the weight to the [0, 1] interval.

5. The method for detecting tea plant diseases and insect pests in a complex environment based on the improved YOLOv11 model according to claim 2, characterized in that: The improved model loss function in S2 is specifically the dynamic focus WIoU loss function DFWIoU to improve the optimized detection head, and adaptively adjust the sample weight according to the overlap IoU between the predicted box and the real box. The formula is: Where w represents the sample weight; IoU represents the intersection over union ratio. The weight of high IoU samples (easy samples) is reduced, and the weight of low IoU samples (difficult samples) is increased. ∈ represents the minimum value (1e-5). To prevent the denominator from being zero, the gradient alignment constraint of the lesion edge is introduced to suppress background false detection. Add a boundary-sensitive penalty term, introduce the gradient information of the lesion edge, and suppress background false detection. The formula is: in, represents the Sobel gradient of the real lesion, Represents the gradient of the prediction area; Finally, the total loss function formula is: L total =L DFWIoU +0.5 L edge +L prune Among them, L total Represents the total loss function; L DFWIoU Represents the dynamic focus WIoU loss function; based on weighted IoU calculation, L edge represents the boundary sensitive loss function, the weight coefficient 0.5 is determined by grid search, L prune represents the pruning regularization loss function.

6. The method for detecting tea plant diseases and insect pests in a complex environment based on the improved YOLOv11 model according to claim 1, characterized in that: The specific methods of S3 are: S3.1 Input the training set data, output the fused feature map through the network model, and save the weight file; The S3.2 network model uses pre-trained weights to infer the validation set and obtain preliminary performance indicators; S3.3 Based on the performance indicators, the model hyperparameters are adjusted and the optimal weights are obtained after multiple rounds of iterations; S3.4 uses the optimal weights to test the model generalization ability on the test set to verify the model effectiveness and deployment feasibility.

Images