Plant leaf disease image segmentation method based on boundary perception guidance

Abstract

The invention provides a plant leaf disease image segmentation method based on boundary perception guidance. The method comprises the following steps: S1, obtaining a plant leaf sample image and a label image; s2, constructing an image segmentation model based on boundary perception guidance, wherein the image segmentation model comprises an encoder, a multi-scale cross fusion decoder and a boundary perception guidance unit; the encoder inputs a sample image and extracts features of different dimensions of the sample image, the output end of the encoder is connected to the multi-scale cross fusion decoder for feature processing, the output features of the multi-scale cross fusion decoder are input to the boundary perception guide unit, and the boundary perception guide unit outputs a segmentation result; s3, inputting the sample image and the label image into an image segmentation model to train the image segmentation model; and S4, obtaining a to-be-identified plant leaf image, and inputting the to-be-identified plant leaf image into the trained image segmentation model to obtain a disease part segmentation result of the to-be-identified plant leaf image.

Description

Technical Field

[0001] The invention relates to an image segmentation method, and in particular to a plant leaf disease image segmentation method based on boundary perception guidance. Background Art

[0002] Plant leaf diseases are one of the key factors affecting plant cultivation, such as grape leaves, tea leaves, etc., especially for some economic crops, which particularly affect their yields. Therefore, timely and accurate segmentation of diseased areas and identification of disease types are crucial for assessing the severity of the disease and formulating control strategies.

[0003] In the prior art, the disease assessment of plant leaves mainly relies on experience to subjectively judge the disease type and its severity. However, this method is overly dependent on agricultural technicians, which is not only cumbersome and time-consuming, but also has low reliability in disease identification. With the development of technology, automatic segmentation algorithms for leaf disease images based on computer vision technology have gradually attracted attention. Such algorithms provide higher accuracy and efficiency for disease assessment by accurately segmenting the diseased area. However, complex and uncontrollable environments affect the accuracy of segmentation. For example, complex backgrounds and light interference cause blurred boundaries of diseased areas and differences in the size and shape of diseased areas, which increases the difficulty of segmentation and leads to low accuracy. Moreover, the performance improvement of existing segmentation models is usually accompanied by an increase in model scale and computational cost.

[0004] Therefore, in order to solve the above technical problems, it is urgent to propose a new technical means. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a plant leaf disease image segmentation method based on boundary perception guidance. The multi-scale cross-fusion decoder can effectively extract multi-scale features and capture global information by combining multi-size strip convolution kernels and cross-axis attention mechanism, thereby improving the segmentation effect of disease areas of different sizes. Moreover, the boundary perception guidance unit can strengthen the learning of boundary information features and enhance the sensitivity to boundary information, thereby effectively improving the recognition ability of the fuzzy edges of plant leaf diseases and ensuring the accuracy of the final disease area image segmentation.

[0006] The present invention provides a plant leaf disease image segmentation method based on boundary perception guidance, comprising the following steps:

[0007] S1. Obtain plant leaf sample images and label images;

[0008] S2. Constructing an image segmentation model based on boundary perception guidance, wherein the image segmentation model includes an encoder, a multi-scale cross-fusion decoder and a boundary perception guidance unit;

[0009] The encoder inputs a sample image and extracts features of different dimensions of the sample image, the output end of the encoder is connected to a multi-scale cross-fusion decoder for feature processing, the output features of the multi-scale cross-fusion decoder are input to a boundary perception guidance unit, and the boundary perception guidance unit outputs a segmentation result;

[0010] S3. Input the sample image and the label image into the image segmentation model to train the image segmentation model until the training is completed;

[0011] S4. Obtain the image of the plant leaf to be identified, and input the image of the plant leaf to be identified into the trained image segmentation model to obtain the segmentation result of the diseased part of the image of the plant leaf to be identified.

[0012] Furthermore, the encoder extracts features from the sample image in four stages, where:

[0013] The image feature of the first stage is X1, and its resolution is

[0014] The image feature of the second stage is X2, and its resolution is

[0015] The image feature of the third stage is X3, and its resolution is

[0016] The image feature of the fourth stage is X4, and its resolution is

[0017] Where: H is the height of the sample image, and W is the width of the sample image.

[0018] Further, the multi-scale cross-fusion decoder includes 3 multi-scale cross-attention units and 3 dual-path channel attention units;

[0019] Image feature X1, image feature X2 and image feature image X3 are input into the first multi-scale cross-attention unit for processing, image feature X2, image feature X3 and image feature image X4 are input into the second multi-scale cross-attention unit for processing, image feature X3 and image feature X4 are input into the third multi-scale cross-attention unit for processing, the output feature of the third multi-scale cross-attention unit is added element by element to the output feature of the second multi-scale cross-attention unit and then input into the third dual-path channel attention unit for processing, the output feature of the third dual-path channel attention unit is added element by element to the output feature of the first multi-scale cross-attention unit and then input into the second dual-path channel attention unit for processing, the output feature of the second dual-path channel attention unit is added element by element to the image feature X1 and then input into the first dual-path channel attention unit for processing, and the output features of the first dual-path channel attention unit, the second dual-path channel attention unit and the third dual-path channel attention unit are input into the boundary perception guidance unit for processing.

[0020] Furthermore, the three multi-scale cross-attention units have the same structure;

[0021] The multi-scale cross attention unit includes Concat module, 1×1 convolution module I, multi-scale cross attention module, local attention module and 1×1 convolution module II;

[0022] The input end of the Concat module is the input end of the multi-scale cross-attention unit. The Concat module is used to concatenate and fuse the input features. The output end of the Concat module is connected to the input end of the 1×1 convolution module I. The output end of the 1×1 convolution module I is connected to the input ends of the multi-scale cross-attention module and the local attention module respectively. The output features of the multi-scale cross-attention module and the local attention module are added element by element and input into the 1×1 convolution module II. The output end of the 1×1 convolution module II is the output end of the multi-scale cross-attention unit.

[0023] Furthermore, the local attention module includes a 1×1 convolution module III, a 3×3 lightweight convolution module, and a 1×1 convolution module IV;

[0024] The input end of the 1×1 convolution module III is the input end of the local attention module, and the output end of the 1×1 convolution module III is connected to the input end of the 3×3 lightweight convolution module. The output features of the 3×3 lightweight convolution module are added element-by-element with the output features of the 1×1 convolution module II and input into the 1×1 convolution module IV. The output features of the 1×1 convolution module IV are processed by the Sigmoid activation function to generate weight map features. The weight map features are multiplied element-by-element with the output features of the 1×1 convolution module II, and then added element-by-element with the output features of the 1×1 convolution module II to obtain the output features of the local attention module.

[0025] Furthermore, the loss function of the image segmentation model is:

[0026] L=L seg +L b ;

[0027] L seg =L Bce (P,B)+L Dice (P,B);

[0028] L b =L Bce (P b ,B b )+L Dice (P b ,B b );

[0029] Where: P is the image segmentation model prediction image, B is the label image, P b is the boundary prediction result of the sample image, B b is the boundary segmentation map of the label image, L Bce represents the binary cross entropy loss function, L Dice Represents the Dice loss function.

[0030] Further, the boundary perception guidance unit convolution feature fusion block and boundary guidance feature propagation module;

[0031] The input end of the convolution feature fusion module is connected to the output end of the multi-scale cross-fusion decoder, the output end of the convolution feature fusion module is connected to the output end of the boundary guidance feature propagation module, and the boundary guidance feature propagation module outputs the image segmentation result.

[0032] Beneficial effects of the present invention: Through the present invention, the multi-scale cross-fusion decoder can effectively extract multi-scale features and capture global information by combining multi-size strip convolution kernels and cross-axis attention mechanism, thereby improving the segmentation effect of diseased areas of different sizes, and the boundary perception guidance unit can strengthen the learning of boundary information features and enhance the sensitivity to boundary information, ultimately effectively improving the ability to recognize the fuzzy edges of plant leaf diseases and ensuring the accuracy of the final diseased area image segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0034] Figure 1 It is a structural schematic diagram of the present invention.

[0035] Figure 2 It is a schematic diagram of the structure of the multi-scale cross-fusion decoder of the present invention.

[0036] Figure 3 Schematic diagram of the local attention module structure of the present invention.

[0037] Figure 4 Schematic diagram of the multi-scale cross-attention unit structure of the present invention.

[0038] Figure 5 It is a schematic diagram of the structure of the convolutional feature fusion module of the present invention.

[0039] Figure 6 It is a schematic diagram of the structure of the boundary guidance feature propagation module of the present invention.

[0040] Figure 7 Schematic diagram of the structure of the multi-scale cross-attention module of the present invention.

[0041] Figure 8 Schematic diagram of the dual-path channel attention unit structure of the present invention. DETAILED DESCRIPTION

[0042] The present invention is further described in detail below:

[0043] The present invention provides a plant leaf disease image segmentation method based on boundary perception guidance, comprising the following steps:

[0044] S1. Obtain plant leaf sample images and label images;

[0045] S2. Constructing an image segmentation model based on boundary perception guidance, wherein the image segmentation model includes an encoder, a multi-scale cross-fusion decoder and a boundary perception guidance unit;

[0046] The encoder inputs a sample image and extracts features of different dimensions of the sample image. The output end of the encoder is connected to a multi-scale cross-fusion decoder for feature processing. The output features of the multi-scale cross-fusion decoder are input to a boundary-aware guidance unit, and the boundary-aware guidance unit outputs a segmentation result. The encoder uses a lightweight backbone network MobileNetV2 based on CNN to extract features. The MobileNetV2 network uses an inverse residual structure and a linear bottleneck layer design to improve the information expression ability of the model in a low-dimensional space, and uses a deep separable convolution to reduce the number of parameters and the amount of calculation.

[0047] S3. Input the sample image and the label image into the image segmentation model to train the image segmentation model until the training is completed;

[0048] S4. Obtain the image of the plant leaf to be identified, and input the image of the plant leaf to be identified into the trained image segmentation model to obtain the segmentation result of the diseased part of the plant leaf image to be identified. Through the present invention, the multi-scale cross-fusion decoder can effectively extract multi-scale features and capture global information by combining multi-size strip convolution kernels and cross-axis attention mechanism, thereby improving the segmentation effect of diseased areas of different sizes, and the boundary perception guidance unit can strengthen the learning of boundary information features and enhance the sensitivity to boundary information, ultimately effectively improving the recognition ability of the fuzzy edges of plant leaf diseases, and ensuring the accuracy of the final diseased area image segmentation.

[0049] In this embodiment, the encoder divides the sample image into four stages to extract features, where:

[0050] The image feature of the first stage is X1, and its resolution is

[0051] The image feature of the second stage is X2, and its resolution is

[0052] The image feature of the third stage is X3, and its resolution is

[0053] The image feature of the fourth stage is X4, and its resolution is

[0054] Where: H is the height of the sample image, and W is the width of the sample image.

[0055] In this embodiment, Figure 2 As shown: the multi-scale cross-fusion decoder includes 3 multi-scale cross-attention units and 3 dual-path channel attention units;

[0056] Image feature X1, image feature X2 and image feature image X3 are input into the first multi-scale cross attention unit for processing, image feature X2, image feature X3 and image feature image X4 are input into the second multi-scale cross attention unit for processing, image feature X3 and image feature X4 are input into the third multi-scale cross attention unit for processing, the output feature of the third multi-scale cross attention unit is added element by element to the output feature of the second multi-scale cross attention unit and then input into the third dual-path channel attention unit for processing, the output feature of the third dual-path channel attention unit is added element by element to the output feature of the first multi-scale cross attention unit and then input into the second dual-path channel attention unit. The output features of the second dual-path channel attention unit are processed in the first dual-path channel attention unit, and the output features of the first dual-path channel attention unit, the second dual-path channel attention unit, and the third dual-path channel attention unit are input into the boundary perception guidance unit for processing. Through the above structure, three dual-path channel attention units (DualBranchChannelAttention, DBCA for short) respectively perform adaptive weighting on the feature map after multi-scale aggregation to reduce redundant information and enhance effective features, thereby improving the ability to restore image details. Among them, the first multi-scale cross attention unit is referred to as MSCM1, the second multi-scale cross attention unit is MSCM2, and the third multi-scale cross attention unit is referred to as MSCM3. The first dual-path channel attention unit is referred to as DBCA1, the second dual-path channel attention unit is referred to as DBCA2, and the third dual-path channel attention unit is referred to as DBCA3.

[0057] In this embodiment, Figure 4 As shown: the three multi-scale cross-attention units have the same structure;

[0058] The multi-scale cross attention unit includes Concat module, 1×1 convolution module I, multi-scale cross attention module, local attention module and 1×1 convolution module II;

[0059] The input of the Concat module is the input of the multi-scale cross attention unit. The Concat module is used to concatenate and fuse the input features. The output of the Concat module is connected to the input of the 1×1 convolution module I. The output of the 1×1 convolution module I is connected to the input of the multi-scale cross attention module and the local attention module respectively. The output features of the multi-scale cross attention module and the local attention module are added element by element and input into the 1×1 convolution module II. The output of the 1×1 convolution module II is the output of the multi-scale cross attention unit. The multi-scale cross attention unit combines the multi-scale strip convolution operation and the axial attention mechanism. The axial attention decomposes the self-attention into two independent attention operations of the horizontal axis and the vertical axis. Among them, the multi-scale cross attention module is an existing technology, and its structure is as follows: Figure 7 As shown, the processing process is:

[0060] For a given feature map F, whose shape is H×W×C, F is encoded by three parallel one-dimensional convolutions, and the output results are added and then sent to a 1×1 convolution layer for further processing. The formula is as follows:

[0061]

[0062] Among them, Conv1Dx i (·) represents the one-dimensional convolution along the x-axis, Norm() is the layer normalization, and Fx is the output result. Following the settings of the SegNeXt network, the kernel size of the one-dimensional convolution is set to 1×7, 1×11, and 1×21. For the y-axis branch output Fy, it can be expressed as:

[0063]

[0064] Secondly, in order to better utilize the multi-scale convolutional features in two spatial directions, the network performs cross-attention calculations between Fx and Fy, treating Fx as a key and value matrix and Fy as a query matrix. The calculation process is as follows:

[0065] F T =C att_x (Fy,Fx,Fx), (3)

[0066] Among them, C att_x Represents the multi-head cross attention of the x-axis. For the branch in the y-axis direction, a similar method is used to encode the context along the y-axis. The calculation process is as follows:

[0067] F B =C att_y (Fx,Fy,Fy),(4)

[0068] Among them, C att_y represents the multi-head cross attention along the y-axis.

[0069] The dual-path channel attention unit adopts the existing technology, specifically: Figure 8 As shown:

[0070] First, a 3×3 convolution kernel with a stride of 1 is used to align the input feature map. Perform convolution operation, then perform batch normalization and ReLU6 activation function operation to output it, which is expressed as Then, on the two branches, Conduct SE Attention and SP Attention Operation, the output feature map is concatenated. Finally, a 3×3 convolution operation is performed to further extract features, and the final output features are obtained after batch normalization and ReLU6 activation function. Attention The calculation of is as follows:

[0071]

[0072] Among them, the vector z c The length of the feature map The number of channels C is obtained by It is obtained by performing global average pooling. and is a fully connected layer.

[0073] Since some features are difficult to extract and require extra attention, SP is added to the DBCA module. Attention Branches. Different categories in the image have obvious correlation in channel sparsity. By analyzing channel sparsity, it can be found that the features in some channels are common, while the features in other channels are relatively rare. SP Attention Image feature distribution can provide meaningful distinction information for different categories. The feature map of the kth channel Sparsity is calculated as follows:

[0074] Γ k =1-Ω k ,(7)

[0075]

[0076] Among them, Γ k is the sparsity of the kth channel. In order to emphasize the importance of channels associated with less common features, the DBCA module redefines the channel weights as follows:

[0077]

[0078] Where C represents the number of channels in the feature map. δ is a very small non-zero value, which is set to 0.0001 in the experiment. Next, the weight coefficient ω is used to k Further processing is performed to adjust its scale and reshape the result into a shape suitable for channel-by-channel multiplication with the input feature map. Finally, by weighting the input feature map, the representation of important channels is enhanced and the influence of unimportant channels is suppressed, thereby guiding the network to focus on more significant features.

[0079] To further improve the feature expression capability, feature maps of different resolutions are fused through local attention blocks, which enhances the interaction between global and local features while retaining local detail information. Figure 3 As shown:

[0080] The local attention module includes a 1×1 convolution module III, a 3×3 lightweight convolution module, and a 1×1 convolution module IV;

[0081] The input end of the 1×1 convolution module III is the input end of the local attention module, and the output end of the 1×1 convolution module III is connected to the input end of the 3×3 lightweight convolution module. The output features of the 3×3 lightweight convolution module are added element-by-element with the output features of the 1×1 convolution module II and input into the 1×1 convolution module IV. The output features of the 1×1 convolution module IV are processed by the Sigmoid activation function to generate weight map features. The weight map features are multiplied element-by-element with the output features of the 1×1 convolution module II, and then added element-by-element with the output features of the 1×1 convolution module II to obtain the output features of the local attention module.

[0082] In this embodiment, the loss function of the image segmentation model is:

[0083] L=L seg +L b ;

[0084] L seg =L Bce (P,B)+L Dice (P,B);

[0085] L b =L Bce (P b ,B b )+L Dice (P b ,B b );

[0086] Where: P is the image segmentation model prediction image, B is the label image, P b is the boundary prediction result of the sample image, B b is the boundary segmentation map of the label image, L Bcerepresents the binary cross entropy loss function, L Dice Represents the Dice loss function, where both the binary cross entropy loss function and the Dice loss function use existing function expressions.

[0087] In this embodiment, the boundary perception guidance unit convolution feature fusion block and boundary guidance feature propagation module;

[0088] The input end of the convolution feature fusion module (denoted by its English abbreviation CFFB in the figure, and its full name in English is Convolution Feature Fusion Block) is connected to the output end of the multi-scale cross fusion decoder, and the output end of the convolution feature fusion module is connected to the output end of the boundary guidance feature propagation module (denoted by its English abbreviation BFPB in the figure, and its full name in English is Boundary-guided Feature Propagation Block), and the boundary guidance feature propagation module outputs the image segmentation result. Figure 5 and Figure 6 As shown, the specific structures of the two modules of the boundary sensing and guiding unit are implemented by using the existing technology, specifically:

[0089] The convolutional feature fusion block first transforms the output of each stage in the decoder d 1 , d 2 , d 3 , d 4 Upsample to the original image size, concatenate, and pass the concatenated feature map to a 3×3 convolution layer for processing. In order to further enhance the fused feature expression, the module uses global average pooling, 1×1 convolution layer, and ReLU activation function to generate a feature weight vector. Then, the weight vector is multiplied by the output feature of the convolution sequence to obtain the output result X.

[0090] Subsequently, the boundary-guided feature propagation block combines the feature map X with the binary boundary feature map FB to generate a boundary-aware feature map Among them, the boundary feature map FB is composed of d 1 It is obtained by Laplace operator. In this process, FB is used to guide feature propagation. The feature propagation direction is divided into horizontal direction and vertical direction. In the horizontal direction, feature propagation is performed from left to right and from right to left respectively, generating two feature maps and In the vertical direction, feature propagation is also performed from top to bottom and from bottom to top to generate two feature maps. and Specifically, for each row in the horizontal direction, the pixel index range belonging to the same segment is found by scanning row by row, and then the average value of the pixel values ​​in the range is calculated, and the feature propagation is performed in sequence. The propagation in the vertical direction is completed column by column, and the pixel index range in the same segment is also found, and the feature is updated according to the average value of the pixels in the range. The specific formula is as follows:

[0091]

[0092] Finally, the propagation feature maps in four directions and The module outputs the feature map X out Defined as:

[0093] Through the above, it is ensured that each pixel can receive information from different directions within the same segment, thereby improving the similarity and consistency of features while maintaining the distinction between different objects. This multi-directional information propagation mechanism based on boundary perception can effectively enhance the boundary perception ability of objects in segmentation tasks while improving the accuracy of feature representation.

[0094] In the figure, ⊕ represents element-by-element addition, ⊙ represents element-by-element multiplication, represents the sparsity of the feature map, Represents matrix multiplication.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A plant leaf disease image segmentation method based on boundary perception guidance, characterized in that: The following steps are involved: S1. Obtain plant leaf sample images and label images; S2. Constructing an image segmentation model based on boundary perception guidance, wherein the image segmentation model includes an encoder, a multi-scale cross-fusion decoder and a boundary perception guidance unit; The encoder inputs a sample image and extracts features of different dimensions of the sample image, the output end of the encoder is connected to a multi-scale cross-fusion decoder for feature processing, the output features of the multi-scale cross-fusion decoder are input to a boundary perception guidance unit, and the boundary perception guidance unit outputs a segmentation result; S3. Input the sample image and the label image into the image segmentation model to train the image segmentation model until the training is completed; S4. Obtain the image of the plant leaf to be identified, and input the image of the plant leaf to be identified into the trained image segmentation model to obtain the segmentation result of the diseased part of the image of the plant leaf to be identified.

2. The plant leaf disease image segmentation method based on boundary perception guidance according to claim 1 is characterized by: The encoder extracts features from the sample image in four stages, where: The image feature of the first stage is X1, and its resolution is The image feature of the second stage is X2, and its resolution is The image feature of the third stage is X3, and its resolution is The image feature of the fourth stage is X4, and its resolution is Where: H is the height of the sample image, and W is the width of the sample image.

3. The plant leaf disease image segmentation method based on boundary perception guidance according to claim 2 is characterized by: The multi-scale cross-fusion decoder includes 3 multi-scale cross-attention units and 3 dual-path channel attention units; Image feature X1, image feature X2 and image feature image X3 are input into the first multi-scale cross-attention unit for processing, image feature X2, image feature X3 and image feature image X4 are input into the second multi-scale cross-attention unit for processing, image feature X3 and image feature X4 are input into the third multi-scale cross-attention unit for processing, the output feature of the third multi-scale cross-attention unit is added element by element to the output feature of the second multi-scale cross-attention unit and then input into the third dual-path channel attention unit for processing, the output feature of the third dual-path channel attention unit is added element by element to the output feature of the first multi-scale cross-attention unit and then input into the second dual-path channel attention unit for processing, the output feature of the second dual-path channel attention unit is added element by element to the image feature X1 and then input into the first dual-path channel attention unit for processing, and the output features of the first dual-path channel attention unit, the second dual-path channel attention unit and the third dual-path channel attention unit are input into the boundary perception guidance unit for processing.

4. The plant leaf disease image segmentation method based on boundary perception guidance according to claim 3 is characterized by: The 3 multi-scale cross-attention units have the same structure; The multi-scale cross attention unit includes Concat module, 1×1 convolution module I, multi-scale cross attention module, local attention module and 1×1 convolution module II; The input end of the Concat module is the input end of the multi-scale cross-attention unit. The Concat module is used to concatenate and fuse the input features. The output end of the Concat module is connected to the input end of the 1×1 convolution module I. The output end of the 1×1 convolution module I is connected to the input ends of the multi-scale cross-attention module and the local attention module respectively. The output features of the multi-scale cross-attention module and the local attention module are added element by element and input into the 1×1 convolution module II. The output end of the 1×1 convolution module II is the output end of the multi-scale cross-attention unit.

5. The plant leaf disease image segmentation method based on boundary perception guidance according to claim 4 is characterized in that: The local attention module includes a 1×1 convolution module III, a 3×3 lightweight convolution module, and a 1×1 convolution module IV; The input end of the 1×1 convolution module III is the input end of the local attention module, and the output end of the 1×1 convolution module III is connected to the input end of the 3×3 lightweight convolution module. The output features of the 3×3 lightweight convolution module are added element-by-element with the output features of the 1×1 convolution module II and input into the 1×1 convolution module IV. The output features of the 1×1 convolution module IV are processed by the Sigmoid activation function to generate weight map features. The weight map features are multiplied element-by-element with the output features of the 1×1 convolution module II, and then added element-by-element with the output features of the 1×1 convolution module II to obtain the output features of the local attention module.

6. The plant leaf disease image segmentation method based on boundary perception guidance according to claim 1, characterized in that: The loss function of the image segmentation model is: L=L seg +L b ; L seg =L Bce (P,B)+L Dice (P,B); L b =L Bce (P b ,B b )+L Dice (P b ,B b ); Where: P is the image segmentation model prediction image, B is the label image, P b is the boundary prediction result of the sample image, B b is the boundary segmentation map of the label image, L Bce represents the binary cross entropy loss function, L Dice Represents the Dice loss function.

7. The plant leaf disease image segmentation method based on boundary perception guidance according to claim 1, characterized in that: The boundary perception guidance unit convolution feature fusion block and boundary guidance feature propagation module; The input end of the convolution feature fusion module is connected to the output end of the multi-scale cross-fusion decoder, the output end of the convolution feature fusion module is connected to the output end of the boundary guidance feature propagation module, and the boundary guidance feature propagation module outputs the image segmentation result.

Images