Power equipment image semantic segmentation method and system suitable for lightweight architecture

Abstract

The invention discloses a power equipment image semantic segmentation method and system suitable for a lightweight architecture, and the method comprises the steps: enabling a backbone network to generate feature maps of different scales through employing a step-by-step down-sampling encoder; the feature pyramid network performs average pooling on the feature maps with different scales and then splices the feature maps to generate pyramid features; the pyramid features are sent into a plurality of stacked first rectangular self-calibration modules for pyramid feature interaction, and pyramid context features are obtained; feature fusion is carried out on the feature maps of different scales and the pyramid context features of the corresponding scales in a decoder, and a plurality of second rectangular self-calibration modules are utilized to reconstruct the fused features, so that the features are focused on the foreground, and reconstructed features are obtained; and constructing a segmentation head based on semantic prototype guidance, generating semantic prototype embedded category information by the segmentation head, and performing semantic segmentation on the reconstruction features. According to the method, the boundary precision and the classification robustness of image segmentation of the power equipment can be remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and artificial intelligence, specifically to a method and system for semantic segmentation of power equipment images using a lightweight architecture. Background Technology

[0002] Image semantic segmentation, a fundamental task in computer vision, aims to assign a semantic category to each pixel in an input image, and has wide applications in fields such as intelligent inspection, autonomous driving, and remote sensing analysis. In intelligent operation and maintenance of power systems, vision-based semantic segmentation technology can be used to automatically identify key power equipment such as insulators, surge arresters, and transformers, providing support for fault diagnosis and condition assessment.

[0003] However, existing lightweight semantic segmentation models (such as MobileNet) face two major challenges when processing images of power equipment due to limitations in model capacity and computational resources:

[0004] (1) Insufficient boundary modeling capability: Due to the lack of fine perception of spatial details, the model is difficult to accurately depict the edge of the device, resulting in jagged, broken or offset segmentation results; (2) Weak ability to distinguish categories: In scenarios where devices have similar appearances, cluttered backgrounds, or drastic changes in lighting, the model has difficulty effectively embedding and utilizing category semantic information, and is prone to misclassifying devices of different categories as the same category, or confusing the foreground and background.

[0005] Current mainstream methods mostly rely on high-level semantic features for prediction. However, after multiple downsampling, the spatial resolution of high-level features is severely lost. Although some details can be recovered through upsampling or skip connections, they cannot fundamentally reconstruct a high-fidelity spatial structure.

[0006] Therefore, there is an urgent need for a semantic segmentation method that can enhance spatial structure reconstruction capabilities and category semantic guidance mechanisms under a lightweight architecture, so as to improve the boundary accuracy and classification robustness of power equipment image segmentation.

[0007] Invention The technical problem to be solved by this invention is to provide a semantic segmentation method that can simultaneously enhance the spatial structure reconstruction capability and the category semantic guidance mechanism under a lightweight architecture.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A semantic segmentation method for power equipment images applicable to lightweight architectures includes: The backbone network employs a stepwise downsampling encoder to generate feature maps at different scales; The feature pyramid network performs average pooling on feature maps of different scales and then concatenates them to generate pyramid features. The pyramid features are then fed into multiple stacked first rectangular self-calibration modules to interact with each other and obtain pyramid context features. In the future, feature maps of different scales in the encoder and the corresponding pyramid context features will be fused in the decoder, and multiple second rectangle self-calibration modules will be used to reconstruct the fused features so that the features are focused on the foreground and the reconstructed features are obtained. A segmentation head based on semantic prototype guidance is constructed. The segmentation head generates semantic prototype embedding category information and performs semantic segmentation on the reconstructed features.

[0009] In this embodiment, the network architecture of the first rectangular self-calibration module and the second rectangular self-calibration module is the same, both including a rectangular self-calibration attention submodule, a batch normalization submodule, and a multilayer perception submodule; The rectangular self-calibration attention submodule first uses horizontal pooling and vertical pooling to capture the axial global context in two directions, and then uses broadcast addition to perform feature fusion on the rectangular region of interest; next, it uses a self-calibration function to calibrate the region of interest; and then it fuses the calibrated features in the horizontal and vertical directions with the input features through a feature fusion function to obtain weighted attention features. The weighted attention features are sequentially passed through the batch normalization submodule and the multilayer perception submodule to obtain the final attention features.

[0010] In this embodiment, the region of interest is calibrated using a self-calibration function, including: The horizontal shape is calibrated using horizontal bar convolution, adjusting each row of elements to make the horizontal shape closer to the foreground object; the features are normalized using a second batch of normalization submodules, and non-linearity is added using the ReLU activation function; Vertical strip convolution is also used for calibration in the vertical direction, and the principle is the same as that of horizontal strip convolution for calibrating the shape in the horizontal direction.

[0011] In this embodiment, the self-calibration function is expressed by the following formula: ; In the formula, This is a self-calibration function, and its output is attention features. This is the feature obtained after fusing a rectangular region of interest using broadcast addition. Represents vertical strip convolution. Represents horizontal strip convolution. Indicates the length of the strip convolution. This represents the combination of batch normalization and the ReLU activation function. This represents the sigmoid function.

[0012] In this embodiment, the calibrated features in both the horizontal and vertical directions are fused with the input features through a feature fusion function to obtain weighted attention features. This includes: using a 3×3 depthwise convolution to extract local details of the features, and then weighting the calibrated attention features through matrix dot product. The feature fusion function is expressed by the following formula: ; In the formula, For feature fusion function, For a depthwise convolution with a kernel of 3×3, For matrix dot product, For attention features, This is represented as the calibrated feature in both the horizontal and vertical directions.

[0013] In this embodiment, the final attention features are obtained using the following formula: ; In the formula, For the final attention features, For input features, Adding to broadcast, and Indicates horizontal and vertical pooling. For self-calibration function, For feature fusion function, This indicates batch normalization and multi-layer perception.

[0014] In this embodiment, multiple second rectangular self-calibration modules are used to reconstruct the fused features, focusing them on the foreground to obtain reconstructed features, including: The pyramid context features are re-splitened according to the number of channels before merging, resulting in fused features with the same number of channels but downsampled to 1 / 64 of the original resolution. , , ; Fusion features Upsampled to 1 / 32 of the input, and then combined with the feature map after passing through the second rectangular self-calibration module. Performing a dot product, the process is described as follows: ; To reconstruct features, For upsampling, For rectangular self-calibration, Dot product; Reconstruct features Upsample to 1 / 16 of the input and combine with the feature map. The features are added together and then fused with the upsampled features of the same resolution through the second rectangular self-calibration module. Perform a dot product, the process is described as follows: ; Reconstruct features Upsample to 1 / 8 of the input and combine with the feature map. The features are added together, then fused with the same resolution by the second rectangular self-calibration module. Performing a dot product, the process is described as follows: .

[0015] In this embodiment, semantic segmentation of the reconstructed features includes: The reconstructed features are projected onto the class feature space, and then the features in the class space are multiplied by the features in the pixel space to obtain the semantic prototype. A fully connected layer is used to compress the semantic prototype to C×1 dimensions to embed class information and generate a class embedding vector; C is the number of channels. Multiply the class embedding vector by its transpose prototype to obtain the enhanced class embedding vector; this enables the second fully connected layer to capture the context of the attention features in the reconstructed features. The attention features are weighted onto the pixel features to obtain the semantic category features after segmentation.

[0016] In this embodiment, the segmented semantic category features are obtained using the following formula: ; In the formula, For semantic category features, To reconstruct features, as input, For semantic prototypes, For embedding vectors, ReLU represents matrix multiplication, and ReLU represents the ReLU activation function. Represents matrix dot product. This indicates a fully connected layer.

[0017] This invention also provides a semantic segmentation system for power equipment images applicable to a lightweight architecture, which applies the above-described semantic segmentation method for power equipment images, including: The backbone network module is used to generate feature maps of different scales using a progressive downsampling encoder. The pyramid feature extraction module is used to perform average pooling on feature maps of different scales and then stitch them together to generate pyramid features. The pyramid features are then fed into multiple stacked first rectangle self-calibration modules for pyramid feature interaction to obtain pyramid context features. The spatial feature reconstruction module is used to fuse feature maps of different scales in the encoder with the corresponding pyramid context features in the decoder, and to reconstruct the fused features using multiple second rectangle self-calibration modules so that the features are focused on the foreground and the reconstructed features are obtained. The semantic prototype-guided segmentation head module is used to construct a segmentation head based on semantic prototype guidance. The segmentation head generates semantic prototype embedding category information and performs semantic segmentation on the reconstructed features.

[0018] Compared with existing technologies, the advantages of this invention are: This invention proposes a semantic segmentation method for power equipment images based on spatial feature reconstruction, improving the ability to model target locations while embedding category information to enhance the differentiation of target categories. Based on a lightweight architecture constraint, this invention constructs a semantic segmentation model to perform high-precision boundary localization and category discrimination semantic segmentation tasks for power equipment targets in complex backgrounds.

[0019] This invention proposes a rectangular self-calibration module in the pyramid feature extraction module and the spatial feature reconstruction module to improve the position modeling of foreground objects. It captures long-distance axial contextual information through horizontal and vertical strip convolutions and constructs a rectangular receptive field by combining broadcast addition, which effectively restores the spatial structural details lost due to multiple downsampling. At the same time, a self-calibration mechanism is introduced to perform shape-aware adjustment of the region of interest, so that the feature response is more in line with the real geometric contour of the power equipment, thereby significantly improving the boundary segmentation accuracy and alleviating common problems such as jagged edges and breaks.

[0020] A semantic prototype-guided segmentation head module is proposed to embed category information. This module injects category-level semantic priors into the pixel-level feature space by explicitly constructing learnable semantic prototype vectors. It enhances the discrimination ability between foreground and background, and between similar device categories by utilizing a class-pixel interaction mechanism. Under lightweight constraints, it avoids dependence on large classification networks or external knowledge bases and achieves highly robust category discrimination only through internal feature projection and fully connected layers. It is particularly suitable for power inspection scenarios with drastic lighting changes, cluttered backgrounds, or highly similar device appearances, and significantly improves the category discrimination ability of foreground objects.

[0021] This invention significantly improves the boundary accuracy and classification robustness of power equipment image segmentation through multiple improvements. Attached Figure Description

[0022] Figure 1 This is a flowchart of a semantic segmentation method for power equipment images under a lightweight architecture, according to an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of a set of semantic segmentation results according to an embodiment of the present invention.

[0024] Figure 3 The figures shown are a set of experimental results for the segmentation of power transmission lines according to an embodiment of the present invention. Detailed Implementation

[0025] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0026] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0027] Please see Figure 1 As shown, this invention provides a semantic segmentation method for power equipment images applicable to a lightweight architecture, comprising: S10, the backbone network uses a step-by-step downsampling encoder to generate feature maps of different scales.

[0028] In one embodiment of the present invention, the backbone network of the semantic segmentation model employs a progressive downsampling encoder, which generates feature maps at different scales. , , , ], the resolutions are respectively [ , , , ].in, For feature height, The width is the feature width.

[0029] S20, the feature pyramid network performs average pooling on feature maps of different scales and then concatenates them to generate pyramid features; then the pyramid features are fed into multiple stacked first rectangular self-calibration modules for pyramid feature interaction to obtain pyramid context features.

[0030] In one embodiment of the present invention, in the pyramid feature extraction module, in order to ensure the efficiency of the entire network, the decoder removes the largest scale feature. Then, low-scale features are pooled using average pooling. , and downsampling to Size them and stitch them together to generate a pyramid feature. . The features are fed into multiple stacked first rectangular self-calibration modules (RSCMs) for pyramid feature interaction and scale-aware semantic feature extraction. This process can be formulated as follows: ,(1); In the formula, To represent as a feature map Downsampling using average pooling operation times. It has the characteristics of a pyramid context. For rectangular self-calibration.

[0031] S30, the feature maps of different scales in the encoder and the corresponding pyramid context features are fused in the decoder, and multiple second rectangle self-calibration modules are used to reconstruct the fused features so that the features are focused on the foreground and the reconstructed features are obtained.

[0032] In one embodiment of the present invention, a spatial feature reconstruction module is proposed to fuse low-level spatial features from the encoder with corresponding high-level features at the same scale in the decoder. The fused features are then reconstructed using RSCM, thereby making the decoder features more focused on the foreground.

[0033] The Rectangular Self-calibration Module (RSCM) captures the axial global context to model the rectangular key region, and then uses the shape self-calibration function to adjust the attention region to the foreground. The feature pyramid is used to guide spatial feature reconstruction, enabling the reconstructed features to incorporate multi-scale information. Detailed operational aspects of the spatial feature reconstruction module are described below: S31, Pyramid Context Features The channels are re-splittered according to the original number of channels, resulting in a fused feature with the same number of channels but a resolution downsampled to 1 / 64 of the original. , , .

[0034] S32 will fuse features Upsampled to 1 / 32 of the input, and then combined with the feature map after passing through the second rectangular self-calibration module. Performing a dot product, the process is described as follows: ; To reconstruct features, For upsampling, For rectangular self-calibration, It is a dot product.

[0035] S33, reconstruct features Upsample to 1 / 16 of the input and combine with the feature map. The features are added together and then fused with the upsampled features of the same resolution through the second rectangular self-calibration module. Perform a dot product, the process is described as follows: .

[0036] S34 will reconstruct features Upsample to 1 / 8 of the input and combine with the feature map. The features are added together, then fused with the same resolution by the second rectangular self-calibration module. Performing a dot product, the process is described as follows: .

[0037] In one embodiment of the present invention, the first rectangular self-calibration module, the second rectangular self-calibration module, and the first to third rectangular self-calibration modules described above are all rectangular self-calibration (RSCM) modules with the same network architecture. However, for ease of distinction, they all include a rectangular self-calibration attention (RSCA) submodule, a batch normalization (BN) submodule, and a multilayer perception (MLP) submodule.

[0038] In this embodiment, firstly, the rectangular self-calibrating attention submodule uses horizontal and vertical pooling to capture the axial global context in two directions, and then uses broadcast addition to perform feature fusion on the rectangular region of interest. Next, a self-calibration function is used to calibrate the region of interest, including strip convolutions in both the horizontal and vertical directions. Then, the calibrated features in the horizontal and vertical directions are fused with the input features using a feature fusion function to obtain weighted attention features.

[0039] Finally, the weighted attention features are passed sequentially through the batch normalization submodule and the multilayer perception submodule to output the final attention features.

[0040] In this embodiment, the region of interest is calibrated using a self-calibration function, including: using horizontal strip convolution to calibrate the shape in the horizontal direction, adjusting each row of elements to make the horizontal shape closer to the foreground object. Then, a second batch normalization submodule is used to normalize the features, and a ReLU activation function is used to add non-linearity. Vertical strip convolution is then used for calibration in the vertical direction as well. In this way, the convolutions in the two directions can be decoupled, adapting to any shape.

[0041] In this embodiment, the weights of the strip convolutions (horizontal strip convolutions and vertical strip convolutions) are learnable, and the shape of the highlighted region is changed by using two strip convolutions with different weights. Through training, the rectangular self-calibration attention module can learn appropriate weights to adjust the rectangular region towards the foreground target. The rectangular self-calibration attention module models the key rectangular region and adjusts the features through the shape self-calibration function to focus more on the foreground. The shape self-calibration function designed in this invention can be formulated as follows: (2); in, Represents strip convolution. Represents vertical strip convolution. Represents horizontal strip convolution. Indicates the length of the strip convolution. This represents the combination of batch normalization and the ReLU activation function. This represents the sigmoid function. The input features are specifically the features fused from the rectangular region of interest using broadcast addition. Here, they are used as input features, and the output is the attention features.

[0042] In one embodiment of the present invention, the features calibrated in both the horizontal and vertical directions are fused with the input features using a feature fusion function to obtain weighted attention features. This includes: extracting local details of the features using a 3×3 depthwise convolution, and then weighting the calibrated attention features using matrix dot product. The feature fusion function can be formulated as follows: (3); in, For feature fusion function, For a depthwise convolution with a kernel of 3×3, For matrix dot product, The attention features calculated for formula (2) This is represented as the calibrated feature in both the horizontal and vertical directions.

[0043] Finally, the overall framework of the RSCM module can be represented as follows: (4); In the formula, For the final attention features, For input features, Adding to broadcast, and Indicates horizontal and vertical pooling. For self-calibration function, For feature fusion function, This indicates batch normalization and multi-layer perception.

[0044] S40: Construct a segmentation head based on semantic prototype guidance. The segmentation head generates semantic prototype embedding category information and performs semantic segmentation on the reconstructed features.

[0045] In one embodiment of the present invention, a semantic prototype-guided segmentation head (SPGM) is proposed to improve the classification effect of foreground objects. The segmentation head generates semantic prototype embedding category information, which helps to enhance the distinguishability between different categories and improve classification accuracy. The SPGM structure is as follows: Figure 1 As shown, semantic prototype embedding information is generated, class embedding features are projected onto the pixel feature space, and pixel features are weighted to enhance the discriminative power between different classes. The input feature size is... Where D is the number of feature channels, H is the feature height, and W is the feature width.

[0046] In this embodiment, it specifically includes: S41, the reconstructed features are projected onto the class feature space, and then the features in the class space are multiplied by the features in the pixel space to obtain the semantic prototype.

[0047] The semantic prototype reflects the feature distribution of different categories in each image and changes dynamically with the input. The size of the semantic prototype is... Where C represents the number of target categories and D represents the number of feature channels. The semantic prototype generation process can be described as follows: (5); In the formula, Represents matrix multiplication. Represents the semantic prototype. These are the input features, i.e., the reconstructed features. Indicates that features will be reconstructed. Projecting onto the class feature space, the number of channels is projected from D to C.

[0048] S42 uses a fully connected layer to compress the semantic prototype to C×1 dimensions to embed class information and generate class embedding vectors.

[0049] In this embodiment, the segmentation head further embeds category information to enhance the distinguishability between different categories. A fully connected layer is used to compress the prototype to C×1 dimensions to embed class information and generate class embedding vectors, with the Softmax function constraining the value between 0 and 1. The calculation process of the class embedding vector is expressed by the formula: (6); In the formula, Indicates a fully connected layer. Indicates will The number of channels is projected from D to 1. Embedded vector S43, multiply the class embedding vector by its transpose prototype to obtain the enhanced class embedding vector; this enables the fully connected layer 2 to capture the context of the attention features in the reconstructed features.

[0050] S44: Weight the attention features onto the pixel features to obtain the semantic category features after segmentation.

[0051] In this embodiment, to project the class embedding vector into the pixel feature space, the class embedding vector is multiplied by its transpose prototype. With enhanced class embedding, the new attention features have stronger class discriminative power. Then, a fully connected layer is used to capture the context in the attention vector. Finally, the attention features are weighted onto the pixel features to emphasize important features and enhance the feature representation. The entire calculation process is as follows: (7); In the formula, These are semantic category features.

[0052] Please see Figure 2 As shown, this invention is applied to semantic segmentation of images of power equipment, including towers, insulators (with many edges), and wires. Figure 3 In the image, the left image includes the power transmission line as input; the middle image is the manually annotated conductor mask; and the right image is the conductor mask after semantic segmentation according to the present invention. It can be seen that the edge details of the segmentation result of the present invention are better.

[0053] This invention also provides a semantic segmentation system for power equipment images applicable to a lightweight architecture, which applies the above-described semantic segmentation method for power equipment images, including: The backbone network module is used to generate feature maps of different scales using a progressive downsampling encoder.

[0054] The pyramid feature extraction module is used to perform average pooling on feature maps of different scales and then concatenate them to generate pyramid features. The pyramid features are then fed into multiple stacked first rectangle self-calibration modules for pyramid feature interaction to obtain pyramid context features.

[0055] The spatial feature reconstruction module is used to fuse feature maps of different scales in the encoder with the corresponding pyramid context features in the decoder, and to reconstruct the fused features using multiple second rectangle self-calibration modules, so that the features are focused on the foreground and the reconstructed features are obtained.

[0056] The semantic prototype-guided segmentation head module is used to construct a segmentation head based on semantic prototype guidance. The segmentation head generates semantic prototype embedding category information and performs semantic segmentation on the reconstructed features.

[0057] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0058] The above embodiments are merely examples of implementation methods of the invention. The scope of protection of the present invention is not limited to the above embodiments. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A semantic segmentation method for power equipment images applicable to lightweight architectures, characterized in that, include: The backbone network employs a stepwise downsampling encoder to generate feature maps at different scales; The Feature Pyramid Network performs average pooling on feature maps of different scales and then concatenates them to generate pyramid features. The pyramid features are then fed into multiple stacked first rectangular self-calibration modules for pyramid feature interaction to obtain pyramid context features; In the future, feature maps of different scales in the encoder and the corresponding pyramid context features will be fused in the decoder, and multiple second rectangle self-calibration modules will be used to reconstruct the fused features so that the features are focused on the foreground and the reconstructed features are obtained. A segmentation head based on semantic prototype guidance is constructed. The segmentation head generates semantic prototype embedding category information and performs semantic segmentation on the reconstructed features.

2. The image semantic segmentation method for power equipment under a lightweight architecture according to claim 1, characterized in that, The first rectangular self-calibration module and the second rectangular self-calibration module have the same network architecture, both including a rectangular self-calibration attention submodule, a batch normalization submodule, and a multilayer perception submodule; The rectangular self-calibration attention submodule first uses horizontal pooling and vertical pooling to capture the axial global context in two directions, and then uses broadcast addition to perform feature fusion on the rectangular region of interest; next, it uses a self-calibration function to calibrate the region of interest; and then it fuses the calibrated features in the horizontal and vertical directions with the input features through a feature fusion function to obtain weighted attention features. The weighted attention features are sequentially passed through the batch normalization submodule and the multilayer perception submodule to obtain the final attention features.

3. The image semantic segmentation method for power equipment under a lightweight architecture according to claim 2, characterized in that, The region of interest is calibrated using a self-calibration function, including: The horizontal shape is calibrated using horizontal bar convolution, adjusting each row of elements to make the horizontal shape closer to the foreground object; the features are normalized using a second batch of normalization submodules, and non-linearity is added using the ReLU activation function; Vertical strip convolution is also used for calibration in the vertical direction, and the principle is the same as that of horizontal strip convolution for calibrating the shape in the horizontal direction.

4. The image semantic segmentation method for power equipment under a lightweight architecture according to claim 3, characterized in that, The self-calibration function is expressed by the following formula: ； In the formula, This is a self-calibration function, and its output is attention features. This is the feature obtained after fusing a rectangular region of interest using broadcast addition. Represents vertical strip convolution. Represents horizontal strip convolution. Indicates the length of the strip convolution. This represents the combination of batch normalization and the ReLU activation function. This represents the sigmoid function.

5. The image semantic segmentation method for power equipment under a lightweight architecture according to claim 2, characterized in that, The calibrated features in both the horizontal and vertical directions are fused with the input features using a feature fusion function to obtain weighted attention features. This includes: using a 3×3 depthwise convolution to extract local details of the features, and then weighting the calibrated attention features using matrix dot product. The feature fusion function is expressed by the following formula: ； In the formula, For feature fusion function, For a depthwise convolution with a kernel of 3×3, For matrix dot product, For attention features, This is represented as the calibrated feature in both the horizontal and vertical directions.

6. The image semantic segmentation method for power equipment under a lightweight architecture according to claim 2, characterized in that, The final attention features are obtained using the following formula: ； In the formula, For the final attention features, For input features, Adding to broadcast, and Indicates horizontal and vertical pooling. For self-calibration function, For feature fusion function, This indicates batch normalization and multi-layer perception.

7. The image semantic segmentation method for power equipment under a lightweight architecture according to claim 1, characterized in that, Multiple second-rectangle self-calibration modules are used to reconstruct the fused features, focusing them on the foreground, and the reconstructed features are obtained, including: The pyramid context features are re-splitened according to the number of channels before merging, resulting in fused features with the same number of channels but downsampled to 1 / 64 of the original resolution. , , ; Fusion features Upsampled to 1 / 32 of the input, and then combined with the feature map after passing through the second rectangular self-calibration module. Performing a dot product, the process is described as follows: ; To reconstruct features, For upsampling, For rectangular self-calibration, Dot product; Reconstruct features Upsample to 1 / 16 of the input and combine with the feature map. The features are added together and then fused with the upsampled features at the same resolution through the second rectangular self-calibration module. Perform a dot product, the process is described as follows: ; Reconstruct features Upsample to 1 / 8 of the input and combine with the feature map. The features are added together, then fused with the same resolution by the second rectangular self-calibration module. Performing a dot product, the process is described as follows: .

8. The image semantic segmentation method for power equipment under a lightweight architecture according to claim 1, characterized in that, Semantic segmentation of the reconstructed features includes: The reconstructed features are projected onto the class feature space, and then the features in the class space are multiplied by the features in the pixel space to obtain the semantic prototype. A fully connected layer is used to compress the semantic prototype to C×1 dimensions to embed class information and generate a class embedding vector; C is the number of channels. Multiply the class embedding vector by its transpose prototype to obtain the enhanced class embedding vector; this enables the second fully connected layer to capture the context of the attention features in the reconstructed features. The attention features are weighted onto the pixel features to obtain the semantic category features after segmentation.

9. The image semantic segmentation method for power equipment under a lightweight architecture according to claim 8, characterized in that, To obtain the semantic category features after segmentation, use the following formula: ； In the formula, For semantic category features, To reconstruct features, as input, For semantic prototypes, For embedding vectors, ReLU represents matrix multiplication, and ReLU represents the ReLU activation function. Represents matrix dot product. This indicates a fully connected layer.

10. A semantic segmentation system for power equipment images suitable for a lightweight architecture, characterized in that, The image semantic segmentation method for power equipment according to any one of claims 1-9 includes: The backbone network module is used to generate feature maps of different scales using a progressive downsampling encoder. The pyramid feature extraction module is used to perform average pooling on feature maps of different scales and then stitch them together to generate pyramid features. The pyramid features are then fed into multiple stacked first rectangle self-calibration modules for pyramid feature interaction to obtain pyramid context features. The spatial feature reconstruction module is used to fuse feature maps of different scales in the encoder with the corresponding pyramid context features in the decoder, and to reconstruct the fused features using multiple second rectangle self-calibration modules so that the features are focused on the foreground and the reconstructed features are obtained. The semantic prototype-guided segmentation head module is used to construct a segmentation head based on semantic prototype guidance. The segmentation head generates semantic prototype embedding category information and performs semantic segmentation on the reconstructed features.

Images