Lightweight lightning image instance segmentation method

Abstract

The invention belongs to the technical field of high-voltage lightning stroke fault prevention, and more specifically relates to a lightweight lightning image instance segmentation method. The method comprises the steps that firstly, lightning images existing in a target area in thunderstorm days are obtained and cleaned, images which do not meet conditions are removed, and the reserved images should be exposed moderately; then, inputting the cleaned lightning image data into a YOLO-lightning network model, and carrying out instance segmentation on the lightning image data; and finally, outputting contour information of the to-be-identified lightning in the image. According to the method, the problems of insufficient precision and limited feature extraction capability in complex form target instance segmentation in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of high-voltage lightning fault prevention, and more specifically, relates to a lightweight lightning image instance segmentation method. Background Art

[0002] High-voltage transmission lines are the core infrastructure of the power system, and ensuring their safe and stable operation is crucial to the reliability of the power grid. However, since high-voltage transmission lines are often located in a wide and complex natural environment, they are extremely vulnerable to lightning disasters. Lightning is recognized as one of the main causes of transmission line failures and tripping, especially lightning with complex shapes (such as bifurcated lightning), which may cause multiple grounding points, thereby causing multiple lines to trip at the same time. Such faults not only seriously threaten the safety of transmission lines, but may also cause large-scale power outages and affect the continued operation of the power grid.

[0003] With the gradual extension of power transmission lines, especially in complex terrains such as mountainous areas and plateaus, the risk of failures caused by lightning strikes is increasing. This risk is not only due to the complex terrain, but is also closely related to the lightning protection design of the transmission lines. Due to the undulating terrain and changes in the lightning protection angle, lightning is prone to form branch structures, which increases the difficulty and challenge of troubleshooting.

[0004] Existing lightning fault detection methods mainly rely on traditional methods such as video monitoring, lightning location systems and ground inspections. However, these methods can only provide macroscopic information on the occurrence of lightning strikes, and cannot capture the fine features of lightning, especially the specific form of lightning paths and their branches. At the same time, due to the high randomness and complexity of lightning, existing technologies have significant deficiencies in identifying lightning development paths, bifurcations and deformations. This makes the work of preventing and responding to lightning disasters more complicated, and it is impossible to respond to potential threats in a timely and accurate manner.

[0005] The development of deep learning and computer vision technology has brought new opportunities for the accurate identification of lightning paths. As an advanced computer vision method, instance segmentation technology can perform pixel-level segmentation on each lightning channel in an image and capture the complex morphology and branching structure of lightning. However, existing instance segmentation methods still face challenges such as insufficient accuracy and weak feature extraction capabilities when dealing with targets with complex shapes and fine edges such as lightning, especially when dealing with details and complex edges of lightning.

[0006] Therefore, in response to the need for lightning prevention of high-voltage transmission lines, designing an efficient and lightweight lightning instance segmentation model to improve the ability to identify complex lightning paths and branches has become a key technical means to prevent lightning hazards. This model can not only improve the accuracy of lightning warning, but also provide more targeted technical support for lightning protection design and operation and maintenance of transmission lines, thereby greatly reducing the threat of lightning strikes to power grid operations.

[0007] Chinese patent document CN118096654A discloses a road surface disease recognition method based on deep learning, which includes: obtaining a road surface image to be recognized obtained by shooting a target road section; inputting the road surface image to be recognized into a pre-trained road surface disease instance segmentation model, so that the road surface disease instance segmentation model determines the category information and contour information of the road surface disease contained in the road surface image to be recognized, wherein the road surface disease instance segmentation model is based on the Yolov8-seg model, introduces the convolutional attention mechanism into the c2f module, uses the SPPFCSPC module to replace the SPPF module in the original model, and adds the Inner Iou loss function on the basis of the original model; according to the category information and contour information of the road surface disease, the length, width or disease area of ​​the road surface disease is determined. However, this instance segmentation method still faces problems such as insufficient accuracy and weak feature extraction ability when dealing with targets with complex shapes and fine edges such as lightning.

[0008] In view of this, the present invention designs a lightweight lightning image instance segmentation method to solve the problems existing in the prior art. Summary of the invention

[0009] The present invention aims to overcome at least one defect of the above-mentioned prior art and provide a lightweight lightning image instance segmentation method to solve the problems of insufficient accuracy and limited feature extraction capability in the prior art in instance segmentation of complex morphological targets.

[0010] The detailed technical scheme of the present invention is as follows: A lightweight lightning image instance segmentation method, the method comprising: S1. First, obtain lightning images in the target area during thunderstorms, clean them, remove images that do not meet the requirements, and retain images with moderate exposure; S2. Then, the cleaned lightning image data is input into the YOLO-lightning network model to perform instance segmentation on the lightning image data; The YOLO-lightning network model is obtained by lightweight improvement of the YOLOv8-seg network model: the C2f_PBConv module is used to replace the original C2f module, and the lightweight Triple Fusion module is introduced in the Neck part for multi-scale feature fusion; the YOLO-lightning network model specifically includes the Backbone part, the Neck part and the head part; The Backbone part includes Conv, C2f_PBConv and SPPF modules; the Conv module includes a Conv2d two-dimensional convolution layer, a BN layer and a SiLU activation function; the SPPF module includes a Conv module and a MaxPool2d layer; The Neck part includes a C2f_PBConv module, a Triple-Fusion module, a Concatenate and a Conv module; The head part includes a detection head and a segmentation head; S3. Finally, the contour information of the lightning to be identified in the image is output.

[0011] Preferably, according to the present invention, the instance segmentation of the lightning image means that the lightning image to be identified is firstly passed into the Backbone part to obtain a feature map, which is mainly used to extract feature information of the image data; the Neck part fuses the feature map extracted by the Backbone to enhance the representation ability and semantic information of the feature map; the head part uses the output feature map of the Neck part as input to obtain the contour information of the lightning to be identified.

[0012] Preferably according to the present invention, the C2f_PBConv module includes: a feature input Input, a convolution layer Conv, a Split layer, a PBConv_Neck module, a feature concatenation layer Concatenate and an output convolution layer.

[0013] Preferably, according to the present invention, the specific processing flow of the C2f_PBConv module is as follows: S11, the features of the input image are first processed by the convolution layer Conv to perform preliminary feature extraction on the features of the input image; S12, then enter the Split layer for processing, the processed features are divided into two parts, one part of the features is directly passed to the feature concatenation layer Concatenate through the jump connection Shortcut, and the other part enters n repeated PBConv_Neck modules for deep processing; S13, the features processed by n PBConv_Neck modules and the features directly reaching the feature concatenation layer Concatenate through the jump connection are concatenated in Concatenate to retain the multi-scale and multi-channel feature information; S14. Finally, the fused features are compressed and output through the last convolutional layer Conv to generate the segmentation result.

[0014] Preferably, according to the present invention, the PBConv_Neck module replaces the traditional bottleneck layer Bottleneck, including partial batch normalization convolution PBConv, SiLU activation layer, depth convolution Dw_Conv, point-by-point convolution Pw_Conv, Dropout random inactivation layer and residual connection; the input features entering PBConv_Neck through the Split layer pass through PBConv, SiLU activation layer, Dw_Conv, Pw_Conv, Dropout and residual connection in sequence to extract richer local and global features.

[0015] Preferably, according to the present invention, the specific processing flow of the PBConv_Neck module is as follows: S21, the input image features are convolved through some batch normalization convolution PBConv, and the convolution operation is performed on some channels. The features after convolution are batch normalized BatchNormalization to make the feature distribution more stable; S22, after PBConv output, the feature enters the SiLU activation layer and is processed by the SiLU activation function; S23, the features are processed by SiLU activation function and then enter the deep convolution Dw_Conv. Each channel performs convolution operation independently to capture spatial features. S24, then enter the point-by-point convolution Pw_Conv, and combine the features of different channels through 1x1 convolution, which greatly reduces the number of parameters and calculations, and further realizes lightweight design; S25, in order to avoid overfitting of the model, enter the Dropout layer and randomly discard some neurons; S26. Finally, the Add layer is used to merge the features processed by the Dropout layer with the features of the original input of the PBConv_Neck module.

[0016] Preferably, according to the present invention, the TripleFusion module combines deformable convolution and CBAM to improve the flexibility and accuracy of feature extraction by dynamically adjusting the convolution kernel and attention mechanism.

[0017] Preferably, according to the present invention, the specific processing process of the TripleFusion module is as follows: S31, input the lower-level feature Lowerstagelayerfeature, the corresponding-level feature Correspondingstagelayerfeature and the higher-level feature Higherstagelayerfeature from the Backbone part of the YOLO-lightning network into the TripleFusion module; S32. Deformable Convolution performs deformable convolution processing on the lower-stagelayerfeature: Deformable Convolution enhances the ability to capture irregular shapes by dynamically adjusting the position of the convolution kernel; S33, the lower-level features Lowerstagelayerfeature processed by deformable convolution and the corresponding layer features and high-level features enter the feature concatenation layer for feature concatenation to obtain the fused features; S34, the fused features are processed by the CBAM attention mechanism module; The CBAM attention mechanism module weights the importance of different features in channel and spatial dimensions to strengthen the focus on key features; S35. The fused features processed by CBAM are output as fused features Fusionfeature, and the fused features Fusionfeature contain feature information of multiple scales and multiple levels.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention adopts the C2f_PBConv module constructed by PBConv_Neck with partial batch normalization convolution PBConv as the core, and makes a lightweight improvement on the original C2f module. C2f_PBConv is optimized on the basis of C2f. By convolution processing only on some channels and combining batch normalization technology, the amount of calculation is significantly reduced. While reducing the computational complexity of the model, the PBConv module still maintains a high feature extraction accuracy. It is particularly suitable for instance segmentation tasks of complex and variable morphological targets such as lightning, so that the model can maintain efficient operation while reducing the number of parameters.

[0019] (2) The TripleFusion module of the present invention combines the deformable convolution and the convolutional block attention mechanism CBAM. Through the combination of these two lightweight technologies, the accuracy and efficiency are improved. The deformable convolution dynamically adapts to the changes of complex morphological targets such as lightning, effectively processes multi-scale low-level features, and enhances the capture of details. The CBAM module weights important areas through channel and spatial attention mechanisms, further improving the ability to focus on key features. Without significantly increasing parameters, the TripleFusion module expands the receptive field through multi-scale feature fusion, improves the model's feature extraction capability, and ensures the efficient operation of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a structural diagram of the C2f_PBConv module described in the present invention.

[0021] Figure 2 It is a structural diagram of the PBConv module described in the present invention.

[0022] Figure 3 It is a structural diagram of the TripleFusion module described in the present invention.

[0023] Figure 4 This is a structural diagram of the YOLO-lightning network model described in the present invention.

[0024] Figure 5 It is a lightning image obtained after instance segmentation using existing instance segmentation methods.

[0025] Figure 6 The lightning image is obtained after instance segmentation is performed using the instance segmentation method of the present invention. DETAILED DESCRIPTION

[0026] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

[0028] Embodiment 1, This embodiment provides a lightweight lightning image instance segmentation method, the method comprising: First, collect lightning image data and pre-process the data; Then, the preprocessed lightning image data is input into the YOLO-lightning network to perform instance segmentation on the lightning image data; like Figure 4 As shown in the figure, the YOLO-lightning network is obtained by lightweight improvement of the YOLOv8-seg network: the C2f_PBConv module is used to replace the original C2f module, and the lightweight Triple Fusion module is introduced into the Neck part for multi-scale feature fusion. The YOLO-lightning network model specifically includes the Backbone part, the Neck part and the head part; The Backbone part includes Conv, C2f_PBConv and SPPF modules; the Conv module includes a Conv2d two-dimensional convolution layer, a BN layer and a SiLU activation function; the SPPF module includes a Conv module and a MaxPool2d layer; The Neck part includes a C2f_PBConv module, a Triple-Fusion module, a Concatenate and a Conv module; The head part includes a detection head and a segmentation head; The lightning image to be identified is first passed to the Backbone part to extract the feature information of the image data to obtain a feature map; then the Neck part fuses the feature map extracted by the Backbone to enhance the representation ability and semantic information of the feature map; finally, the head part uses the output feature map of the Neck part as input to obtain the contour information of the lightning to be identified.

[0029] Figure 1 The overall architecture of the C2f_PBConv module is shown in Figure 1. It is optimized based on the C2f module in YOLOv8-seg. The C2f_PBConv module includes feature input Input, convolution layer Conv, Split layer, PBConv_Neck module, feature concatenation layer Concatenate, and output convolution layer.

[0030] The specific processing flow of the C2f_PBConv module is as follows: The features of the input image input are first processed by the convolution layer Conv to perform preliminary feature extraction on the input image features; this convolution operation is similar to the traditional C2f module and Bottleneck operation; Then it enters the Split layer for processing. The processed features are divided into multiple parts. One part of the features is directly passed to the Concatenate layer through the skip connection Shortcut, and the other part enters n repeated PBConv_Neck modules for deep processing; After being processed by n PBConv_Neck modules, the features directly passed through the skip connection and processed by the PBConv_Neck module are merged here through the feature concatenation layer Concatenate to ensure that the multi-scale and multi-channel feature information is retained.

[0031] Finally, the fused features are compressed and output through the last convolutional layer Conv to generate the segmentation result.

[0032] The PBConv_Neck module is designed based on partial batch normalization convolution PBConv, aiming to improve the efficiency of feature extraction and reduce the amount of calculation through partial channel convolution and batch normalization technology. The core of its design is to achieve lightweight model by effectively reducing unnecessary calculations while maintaining high feature extraction capabilities. The PBConv_Neck module replaces the traditional bottleneck layer Bottleneck, such as Figure 1 As shown in the right half, it includes PBConv, SiLU activation, Dw_Conv, Pw_Conv, Dropout and residual connection; the input features pass through PBConv, SiLU activation, Dw_Conv, Pw_Conv, Dropout and residual connection in turn to extract richer local and global features.

[0033] The specific processing flow is as follows: First, if Figure 2 As shown in the figure, the input image features are convolved on some channels through partial batch normalization convolution PBConv to avoid redundant processing of all channels. In this way, the amount of calculation is significantly reduced and the overall calculation efficiency is improved. The convolved features will undergo batch normalization BatchNormalization, which can make the feature distribution more stable, effectively prevent the gradient from disappearing or exploding, and accelerate the training of the model. Figure 2 In the figure, Identity represents identity mapping, filters represents filters, and convolution represents convolution.

[0034] After PBConv output, the feature map is processed by the SiLU activation function, which is an improved lightweight activation function that not only retains nonlinear expression capabilities, but also reduces the negative information loss problem that may be caused by traditional activation functions (such as ReLU) by optimizing gradient flow. SiLU helps speed up the training and inference process of the network.

[0035] The feature map is processed by the SiLU activation function and then enters the deep convolution Dw_Conv. Each channel performs the convolution operation independently to capture the spatial features. The introduction of deep convolution significantly reduces the amount of calculation in the convolution operation.

[0036] Then enter the point-by-point convolution Pw_Conv: The features of different channels are combined through 1x1 convolution, which greatly reduces the number of parameters and calculations, and further realizes lightweight design; In order to avoid overfitting of the model, some neurons are randomly discarded in the Dropout layer. This mechanism enhances the generalization ability of the model and makes the model more robust during training and reasoning. The introduction of Dropout can ensure the robustness of the model while maintaining lightweight.

[0037] Finally, the Add layer is used to merge the processed features with the original input. This design ensures the continuity of information transmission through residual connections and effectively alleviates the gradient vanishing problem. At the same time, the introduction of the Add layer does not significantly increase the amount of calculation, but can maintain the integrity and stability of the features during the forward propagation process.

[0038] In the comparison between the C2f_PBConv module and the traditional C2f module, by optimizing the convolution operation, feature extraction method and adding regularization mechanism, the C2f_PBConv module has significant advantages in lightweight: Higher computational efficiency: The traditional C2f module uses conventional convolution to process all channels, resulting in a large amount of computation, especially when processing large-scale inputs, and the computational efficiency is low. The C2f_PBConv module uses multiple PBConv_Neck structures, in which PBConv only performs convolution processing on some channels, avoiding redundant calculations and significantly reducing the amount of computation. The introduction of BatchNormalization further improves computational efficiency and training stability, speeds up inference speed FPS, and better adapts to lightweight design requirements.

[0039] More sophisticated feature extraction: The C2f_PBConv module effectively separates spatial information and channel information by introducing deep separable convolutions Dw_Conv and Pw_Conv. While reducing the number of parameters, the feature extraction accuracy is improved. In contrast, the traditional C2f module uses conventional convolution, which cannot effectively separate space and channels, resulting in insufficient feature extraction accuracy, especially when dealing with complex, multi-branch or multi-scale targets. The performance of traditional modules is poor. The lightweight PBConv_Neck design enables C2f_PBConv to maintain efficient computing performance while improving accuracy.

[0040] Stronger generalization ability: The Dropout layer is introduced in PBConv_Neck, which further enhances the generalization ability of the C2f_PBConv module. Dropout effectively reduces the model's dependence on specific data by randomly discarding some neurons, avoiding the occurrence of overfitting problems. This makes the C2f_PBConv module more robust when dealing with complex and morphologically variable targets, while the traditional C2f module is prone to overfitting problems when dealing with diverse data sets due to the lack of a clear regularization mechanism.

[0041] The core design concept of the TripleFusion module is to enhance the detection and segmentation capabilities of complex morphological targets (such as lightning) by fusing multi-scale features. Figure 3As shown in the figure, the TripleFusion module combines DeformableConvolution and CBAM to improve the flexibility and accuracy of feature extraction by dynamically adjusting the convolution kernel and attention mechanism.

[0042] The specific processing process of the TripleFusion module is as follows: The low-level features Lowerstagelayerfeature, corresponding stagelayerfeature and high-level features Higherstagelayerfeature from the Backbone part of the YOLO-lightning network are input into the TripleFusion module; these features represent levels of different spatial resolutions and semantic information, and fusing these features can capture both local and global information of the target.

[0043] Deformable Convolution performs deformable convolution on the lower-stagelayerfeature: Deformable convolution dynamically adjusts the position of the convolution kernel to enhance the model's ability to capture irregular shapes, and is particularly suitable for processing targets with complex shapes such as lightning. This operation can effectively extract local features in lightning images and capture the shape changes of objects.

[0044] The low-level features processed by deformable convolution, the corresponding-level features, and the high-level features enter the feature concatenation layer for feature concatenation, fusing feature information of different scales together to ensure that the network can simultaneously consider local and global features of multiple scales during the processing process.

[0045] The fused features are processed by the CBAM attention mechanism module; The CBAM attention mechanism module is a module that combines channel attention and spatial attention. It can weight the importance of different features in channel and spatial dimensions, thereby strengthening the focus on key features. Especially for targets with complex backgrounds and more details such as lightning, CBAM can effectively improve the network's attention to key areas, which helps to segment instances more accurately.

[0046] Finally, the fusion features processed by CBAM are output as fusion features. This feature will contain feature information of multiple scales and multiple levels. After being enhanced by the attention mechanism, it can more accurately capture the details and shape changes of the target and be used for subsequent segmentation operations.

[0047] Compared with traditional multi-scale feature fusion methods, the TripleFusion module achieves significant advantages in the following aspects through lightweight design: Dynamic feature capture: Through deformable convolution, the network can dynamically adjust the position of the convolution kernel according to the morphological changes of the target, thereby flexibly capturing the features of irregularly shaped targets. This lightweight mechanism reduces the calculation of redundant features and is particularly suitable for processing targets with complex morphology and many branch changes, such as lightning. Compared with traditional convolution operations, deformable convolution can improve the adaptability to complex targets without significantly increasing the amount of calculation.

[0048] Multi-scale feature fusion: The TripleFusion module combines low-level, mid-level, and high-level feature information through the Concatenate operation to achieve lightweight multi-scale feature fusion. Compared with traditional multi-scale fusion methods, the TripleFusion module avoids repeated calculations, ensuring that while capturing local and global information of the target, redundant operations are reduced. This lightweight feature fusion solution is particularly suitable for processing targets with large ranges and obvious morphological changes, and can improve segmentation accuracy while maintaining model efficiency.

[0049] Enhanced attention mechanism: The introduction of the convolutional block attention mechanism (CBAM) module weights features in two dimensions: channel attention and spatial attention, effectively enhancing the focus on key areas. The weighted operation of CBAM helps the network focus on key features and reduce the processing of unimportant background information, thereby achieving a reasonable allocation of computing resources. This design improves the accuracy of feature extraction in complex scenes and the overall computing efficiency of the model without increasing computational complexity.

[0050] Optimized computational efficiency: Although the TripleFusion module combines multiple feature processing mechanisms, its feature extraction and fusion process is more efficient and flexible by introducing lightweight technologies such as deformable convolution and CBAM. Deformable convolution reduces invalid calculations and ensures that only necessary feature areas are processed, while CBAM optimizes the allocation of computing resources through the attention mechanism and focuses on important features. This lightweight design greatly improves computational efficiency and ensures that the model can still reason quickly when processing complex targets.

[0051] The lightning image finally obtained by instance segmentation is as follows Figure 6 As shown in the figure, it can be seen that the lightning image obtained by instance segmentation using the improved method of the present invention is Figure 5 The lightning image obtained by the unimproved instance segmentation is shown, which enhances the ability to capture the detailed features of lightning.

[0052] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation methods of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A lightweight lightning image instance segmentation method, characterized in that: The method comprises: S1. First, obtain lightning images in the target area during thunderstorms, clean them, remove images that do not meet the requirements, and retain images with moderate exposure; S2. Then, the cleaned lightning image data is input into the YOLO-lightning network model to perform instance segmentation on the lightning image. The YOLO-lightning network model is obtained by lightweight improvement of the YOLOv8-seg network model: the C2f_PBConv module is used to replace the original C2f module, and the lightweight TripleFusion module is introduced in the Neck part for multi-scale feature fusion; the YOLO-lightning network model specifically includes the Backbone part, the Neck part and the head part; The Backbone part includes Conv, C2f_PBConv and SPPF modules; the Conv module includes a Conv2d two-dimensional convolution layer, a BN layer and a SiLU activation function; the SPPF module includes a Conv module and a MaxPool2d layer; The Neck part includes a C2f_PBConv module, a TripleFusion module, a Concatenate and a Conv module; The head part includes a detection head and a segmentation head; S3. Finally, the contour information of the lightning to be identified in the image is output.

2. The lightweight lightning image instance segmentation method according to claim 1, characterized in that: The instance segmentation of lightning images means that the lightning image to be identified is firstly transmitted to the Backbone part to extract the feature information of the image data to obtain a feature map; then the Neck part fuses the feature map extracted by the Backbone to enhance the representation ability and semantic information of the feature map; finally, the head part uses the output feature map of the Neck part as input to obtain the contour information of the lightning to be identified.

3. The lightweight lightning image instance segmentation method according to claim 1, characterized in that: The specific processing flow of the C2f_PBConv module is as follows: S11, the features of the input image are first processed by the convolution layer Conv to perform preliminary feature extraction on the features of the input image; S12, then enter the Split layer for processing, the processed features are divided into two parts, one part of the features is directly passed to the feature concatenation layer Concatenate through the jump connection Shortcut, and the other part enters n repeated PBConv_Neck modules for deep processing; S13, the features processed by n PBConv_Neck modules and the features directly reaching the feature concatenation layer Concatenate through the jump connection are concatenated in Concatenate to retain the multi-scale and multi-channel feature information; S14. Finally, the fused features are compressed and output through the last convolutional layer Conv.

4. The lightweight lightning image instance segmentation method according to claim 3, characterized in that: The PBConv_Neck module includes a partial batch normalization convolution PBConv, a SiLU activation layer, a depth convolution Dw_Conv, a point-by-point convolution Pw_Conv, a Dropout random inactivation layer and a residual connection; the input features entering the PBConv_Neck through the Split layer pass through PBConv, a SiLU activation layer, Dw_Conv, Pw_Conv, Dropout and a residual connection in sequence.

5. The lightweight lightning image instance segmentation method according to claim 4, characterized in that: The specific processing flow of the PBConv_Neck module is as follows: S21, the input image features pass through partial batch normalization convolution PBConv, perform convolution operation on some channels, and the convolved features undergo batch normalization BatchNormalization; S22, after PBConv output, the feature enters the SiLU activation layer and is processed by the SiLU activation function; S23, the features are processed by SiLU activation function and then enter the deep convolution Dw_Conv. Each channel performs convolution operation independently to capture spatial features. S24, then enter the point-by-point convolution Pw_Conv, and combine the features of different channels through 1x1 convolution; S25, enter the Dropout layer to randomly discard some neurons; S26. Finally, the Add layer is used to merge the features processed by the Dropout layer with the features of the original input of the PBConv_Neck module.

6. The lightweight lightning image instance segmentation method according to claim 1, characterized in that: The TripleFusion module combines deformable convolution and CBAM.

7. The lightweight lightning image instance segmentation method according to claim 6, characterized in that: The specific processing process of the TripleFusion module is as follows: S31, input the lower-level feature Lowerstagelayerfeature, the corresponding layer feature Correspondingstagelayerfeature and the higher-level feature Higherstagelayerfeature from the Backbone part of the YOLO-lightning network into the TripleFusion module; S32. Deformable Convolution performs deformable convolution processing on the lower-stagelayerfeature: Deformable Convolution enhances the ability to capture irregular shapes by dynamically adjusting the position of the convolution kernel; S33, the lower-level features Lowerstagelayerfeature processed by deformable convolution and the corresponding layer features and high-level features enter the feature concatenation layer for feature concatenation to obtain the fused features; S34, the fused features are processed by the CBAM attention mechanism module; The CBAM attention mechanism module weights the importance of different features in channel and spatial dimensions to strengthen the focus on key features; S35. The fused features processed by CBAM are output as fused features Fusionfeature, and the fused features Fusionfeature contain feature information of multiple scales and multiple levels.

Images