Aerial transmission line recognition method based on line structure reliable transmission and collaborative correction

Abstract

A kind of aerial transmission line identification method based on line structure credible transmission and collaborative correction. It includes: obtaining the aerial image to be identified, and inputting the encoder to extract multi-level coding features;Perform line width coupling response enhancement on deep coding features and bottleneck features, enhance the line structure representation ability;Use edge prompt, line body prompt and direction estimation feature to generate credible transmission gate, filter shallow coding features, and fuse with current decoding features, realize line structure credible gate jump connection transmission;Extract edge branch features and line body branch features from fused decoding features, realize edge-line collaborative decoding and confidence constraint refinement through bidirectional modulation, obtain refined decoding features;Segmentation prediction is carried out on refined decoding features, and the transmission line identification result is output.The present application aims to improve the problems of insufficient direction adaptability, lack of structure constraint in jump connection, limited scale modeling capability and edge and main body information split in the existing aerial transmission line identification method.

Description

Technical Field

[0001] This invention relates to the field of aerial image analysis technology, specifically to an aerial method for identifying power transmission lines based on reliable transmission of line structures and collaborative correction. Background Technology

[0002] With the development of drone inspections and smart grid construction, the automatic identification of power transmission line targets using aerial images has become an important technical means in power line inspection. However, power transmission lines in aerial images typically exhibit characteristics such as a slender structure, low pixel ratio, diverse directional distribution, and inconsistent local width. They are also easily affected by complex backgrounds such as tree branches, building edges, road edges, mountain textures, shadows, and changes in lighting, posing a significant challenge to high-precision identification.

[0003] While existing deep learning-based image segmentation methods can capture certain semantic features, they still have the following shortcomings in power line recognition tasks: First, conventional convolutional operations often use approximately square receptive fields, which are insufficient for responding to the directions of horizontal, vertical, and long-distance linear structures, making it difficult to effectively handle the directional diversity of power lines in aerial images. Second, traditional encoder-decoder networks directly concatenate shallow features containing rich edge and texture information with decoded features, but shallow features also contain a large number of pseudo-line structures such as tree branches, building outlines, and road edges. Their direct introduction can easily lead to the transmission of background pseudo-lines to the decoder, causing missegmentation or unstable break repair. Third, power line targets are relatively narrow and have significant local scale variations. Single-scale convolution or fixed receptive field structures cannot take into account thin line regions, wider line regions, and sparse linear contextual information, limiting further improvement in segmentation accuracy. Fourth, existing methods usually unify power line boundary prediction and main body region prediction into a single segmentation result, lacking bidirectional structural constraints between the boundary region and the main body of the line, which can easily lead to problems such as blurred boundaries, local adhesion, and thin line breaks. Summary of the Invention

[0004] This invention aims to provide an aerial method for identifying power transmission lines based on reliable transmission of line structures and collaborative correction, in order to improve the problems of insufficient directional adaptability, lack of structural constraints in skip connections, limited scale modeling capabilities, and separation of edge and subject information in existing aerial methods for identifying power transmission lines.

[0005] To address the above technical problems, the specific solution adopted in this invention is as follows: An aerial image to be identified is acquired; the image is then input into a pre-trained recognition model for processing to obtain the power transmission line identification result. The processing method of the aerial image data to be identified by the recognition model includes: line-direction-linewidth coupling response enhancement, line structure trusted gated jump connection transmission, and edge-line body collaborative decoding and confidence constraint refinement. Line-direction-linewidth coupling response enhancement includes inputting the aerial image to be identified into an encoder, extracting multi-level coding features, including shallow coding features, deep coding features, and bottleneck features; performing line-direction-linewidth coupling response enhancement on the deep coding features and bottleneck features respectively; extracting the direction aggregation features and width aggregation features of the power transmission line; and obtaining enhanced features based on the direction aggregation features and width aggregation features. The line structure trusted gated jump connection transmission packet... The process includes inputting enhanced features into the decoder for decoding; at shallow jump connections between the encoder and decoder, generating edge and line cues based on the current decoded features; extracting direction estimation features based on shallow encoded features; combining edge, line cues, and direction estimation features to generate trusted transit gating weights; using these trusted transit gating weights to filter shallow encoded features to obtain trusted jump connection features; fusing these trusted jump connection features with the current decoded features to obtain fused decoded features; edge-line co-decoding and confidence constraint refinement include extracting edge and line branch features from the fused decoded features, performing bidirectional modulation co-correction on these features to obtain co-fused features, and generating refined decoded features based on these co-fused features; performing segmentation prediction on the refined decoded features to output the transmission line identification result.

[0006] As a further optimization of the aerial power transmission line identification method based on line structure reliable transmission and collaborative correction: line-direction-linewidth coupling response enhancement includes: performing local semantic pre-integration on the feature F to be enhanced to obtain preprocessed feature F. p The method is as follows: ,in, This represents the basic convolutional mapping, used for preliminary reorganization of local context in the input features; it also represents the preprocessed features F. p Input the line direction response branch group and the line width response branch group respectively to obtain the response of each direction and the response of each width; generate line direction routing weights and line width routing weights based on the preprocessed features, and perform weighted fusion on the response of each direction and the response of each width to obtain the line direction aggregation feature and the line width aggregation feature. The method is as follows: , ,in, This represents the i-th linear response branch. Let Nd represent the j-th linewidth response branch, Nd represent the number of linewidth response branches, and Nw represent the number of linewidth response branches. and These represent the line-direction routing weight and the line width routing weight, respectively. The enhanced features are obtained by channel-joining the features to be enhanced, the line-direction aggregated features, and the line width aggregated features, followed by fusion mapping and residual connection. The method is as follows: Where [] represents channel dimension splicing operation, Indicates fused convolutional mapping, This indicates the enhanced characteristics of the linear-linewidth coupling response.

[0007] As a further optimization of the aerial power transmission line identification method based on reliable transmission and collaborative correction of line structure: the line direction response branch group includes multiple sets of asymmetric deep convolution branches, including at least 1×5 lateral line direction branches, 5×1 longitudinal line direction branches, 1×9 extended lateral line direction branches and 9×1 extended longitudinal line direction branches, which are used to respond to the local short-range direction and long-range extension structure of the power transmission line respectively; the line width response branch group includes multiple sets of deep convolution branches with different receptive fields, including at least 3×3 fine-grained line width branches, 5×5 neighborhood extended line width branches and 3×3 sparse context line width branches with a hole rate of 2, which are used to perceive the local region of thin line, the neighborhood of wider line body and sparse linear context respectively.

[0008] As a further optimization of the aerial power transmission line identification method based on line structure trusted transmission and collaborative correction, the trusted transmission gating weight generation method is as follows: ,in, This represents the trusted transit gating generation mapping, where G represents the trusted transit gating weight, and X represents the trusted transit gating weight. p S represents the current decoded feature after projection. p The projected shallow jump connection features are represented by E, edge cue, line cue, and T, direction estimation features. The method for obtaining reliable jump connection features is as follows: ,in, Indicates a reliable jump connection feature. This represents element-wise multiplication; the method for obtaining the fused decoded features is as follows: , Indicates by and The initial fusion feature obtained by fusion, Q represents the pseudo-line suppression weight generated by the initial fusion feature, edge cue, and line body cue, and F represents the initial fusion feature. t This indicates the fusion decoding features.

[0009] As a further optimization of the aerial power transmission line identification method based on reliable transmission and collaborative correction of line structure, the bidirectional collaborative correction method is as follows: Edge-to-line modulation is used to convert the edge branch features F... e Structural information injection into line branch feature F l The updated line body features are obtained. The method is as follows: ,in, This indicates the updated line body features. Indicates the modulation weight from the edge to the line body. This represents element-wise multiplication; through line-to-edge modulation, the main information of the updated line features is fed back to the edge branch features to obtain the updated edge features. The method is as follows: ,in, This represents the updated edge features. This represents the modulation weight from the line to the edge.

[0010] As a further optimization of the aerial power transmission line identification method based on reliable transmission and collaborative correction of line structure: the fused decoding features, updated edge features, and updated line body features are collaboratively fused to obtain collaboratively fused features, and then refined decoding features are generated based on the collaboratively fused features. The method is as follows: ,in, This indicates refined decoding features. C represents the collaborative fusion feature, and C represents the updated edge feature. and updated line body features Jointly generated confidence gating This represents the fused decoding features from the input to the final prediction stage.

[0011] As a further optimization of the aerial power transmission line identification method based on reliable transmission and collaborative correction of line structure: the updated edge features and updated line body features are further processed by the edge prediction head and the line body prediction head to obtain edge prediction map and line body prediction map, respectively. The edge prediction map and the line body prediction map are used as auxiliary supervision signals during the training phase to constrain the collaborative expression of the boundary region and the main body region of the power transmission line.

[0012] Beneficial effects: This invention enhances the network's adaptability to changes in transmission line direction and local width fluctuations by using asymmetric linear response branch groups and multi-scale linewidth response branch groups in conjunction with routing weight fusion. It improves the problem that conventional square receptive fields are not adaptable to linear structure directions and that single-scale convolution cannot take into account lines of different thicknesses.

[0013] This invention employs a line-structured trusted gated jump connection transmission method, combining edge hints, line body hints, and direction estimation features to selectively transmit shallow features under structural constraints. This effectively suppresses the propagation of background pseudo-lines such as tree branches and building edges to the decoding end, and improves the problem of background interference amplification caused by direct splicing of traditional jump connections.

[0014] This invention improves the problems of edge blurring, local adhesion and thin line breakage by using edge-line collaborative decoding and confidence constraint refinement, and by using bidirectional modulation of edge branches and line branches to form structural constraints between the boundary and the main body. Attached Figure Description

[0015] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a diagram of the model architecture of the present invention; Figure 3 A comparison chart showing the performance of different methods on test set A; Figure 4 This is a comparison chart of the performance of different methods on test set B. Detailed Implementation

[0016] Example 1

[0017] like Figure 1 and Figure 2 As shown in the figure, this embodiment presents an aerial imagery method for identifying power transmission lines based on reliable transmission of line structures and collaborative correction. First, an aerial image to be identified is acquired. This image is taken by a drone. After acquisition, the corresponding power transmission line annotation mask is preprocessed to obtain a power transmission line image dataset.

[0018] The aerial image to be identified is input into a pre-trained recognition model for processing to obtain the power transmission line identification result.

[0019] The processing methods of the recognition model for the aerial image data to be recognized include: line-to-linewidth coupling response enhancement, line structure reliable gating jump connection transmission, and edge-line body collaborative decoding and confidence constraint refinement.

[0020] Linear-to-linewidth coupling response enhancement includes: S1. Input the aerial image to be identified into the encoder and extract multi-level coding features, which include shallow coding features, deep coding features and bottleneck features; S2. Perform line-to-linewidth coupling response enhancement on deep coding features and bottleneck features respectively, extract the direction aggregation features and width aggregation features of the transmission line, and obtain the enhanced features based on the direction aggregation features and width aggregation features; The wire structure trusted gating jump connection transmission includes: S3. Input the enhanced features into the decoder for decoding. At the shallow jump connection between the encoder and decoder, generate edge cues and line cues based on the current decoded features. Extract the direction estimation features based on the shallow encoding. S4. Generate reliable transitive gating weights based on edge cues, line cues, and direction estimation features; S5. Use trusted transit gating weights to filter shallow coding features to obtain trusted jump connection features. Then fuse trusted jump connection features with the current decoding features to obtain fused decoding features. Edge-line co-decoding and confidence constraint refinement include: S6. Extract edge branch features and line branch features from the fused decoding features respectively, and perform bidirectional modulation collaborative correction on the edge branch features and line branch features to obtain collaborative fusion features. Then, generate refined decoding features based on the collaborative fusion features. S7. Perform segmentation and prediction on the refined decoding features, and output the power transmission line identification results.

[0021] In this embodiment, two aerial image datasets of power transmission lines are used for experimental verification, denoted as dataset A and dataset B, respectively. Dataset A contains 200 aerial images of power transmission lines, with 160 images in the training set, 20 in the validation set, and 20 in the test set. Dataset B contains 858 aerial images of power transmission lines, with 686 images in the training set, 86 in the validation set, and 86 in the test set.

[0022] Before inputting the data into the encoder, the aerial images and their corresponding masks are preprocessed. For the images, they are converted to RGB format and uniformly scaled to a preset input size (256×256 pixels). For the masks, they are converted to single-channel grayscale images, and thresholding is used to set the transmission line area to 1 and the background area to 0.

[0023] During the training phase, further enhancements such as random horizontal flipping, random vertical flipping, random rotation in multiples of 90°, and perturbations of brightness, contrast, and saturation are employed to improve the model's adaptability to different shooting angles, lighting conditions, and background environments. Subsequently, the images are converted into tensors and normalized according to preset mean and standard deviation. Since power lines in aerial images typically appear as slender, low-pixel-ratio structures, and may exhibit horizontal, vertical, tilted, or intersecting distributions in different scenes, random orientation enhancements and brightness perturbations allow the model to encounter more linear target samples under varying directions and lighting conditions during training, thereby reducing the model's dependence on a single scene.

[0024] In this embodiment, the power transmission line identification network adopts an encoder-decoder architecture. The encoder is used to extract multi-level coded features from the input image. The multi-level coded features, from shallow to deep, include shallow coded features, deep coded features, and bottleneck features.

[0025] S1. Input the preprocessed aerial image to be identified into the encoder to extract multi-level coding features. Among them, shallow coding features retain high spatial resolution and rich edge and texture information; deep coding features and bottleneck features have high semantic abstraction ability and low spatial resolution.

[0026] S2. Perform line-to-linewidth coupling response enhancement on the deep coding features and bottleneck features respectively, extract the direction aggregation features and width aggregation features of the transmission line, and obtain the enhanced features based on the direction aggregation features and width aggregation features.

[0027] Linear-to-linewidth coupling response enhancement includes: performing local semantic pre-integration on the feature F to be enhanced to obtain preprocessed feature F. p , is represented as: ,in, This represents the basic convolutional mapping, used for preliminary reorganization of local context in the input features; it also represents the preprocessed features F. p Input the line direction response branch group and the line width response branch group respectively to obtain the response in each direction and the response in each width.

[0028] The linear response branch group includes multiple sets of asymmetric deep convolution branches, including at least 1×5 lateral linear branches, 5×1 longitudinal linear branches, 1×9 extended lateral linear branches, and 9×1 extended longitudinal linear branches, which are used to respond to the local short-range direction and long-range extension structure of the transmission line, respectively.

[0029] The linewidth response branch group includes multiple sets of deep convolution branches with different receptive fields, including at least a 3×3 fine-grained linewidth branch, a 5×5 neighborhood expansion linewidth branch, and a 3×3 sparse context linewidth branch with a hole ratio of 2, which are used to perceive the local region of the fine line, the neighborhood of the wider line body, and the sparse linear context, respectively.

[0030] Based on the preprocessed features, line-direction routing weights and line-width routing weights are generated. The responses in each direction and each width are then weighted and fused to obtain the line-direction aggregated features and line-width aggregated features, as follows: , In the formula, This represents the i-th linear response branch. Let Nd represent the j-th linewidth response branch, Nd represent the number of linewidth response branches, and Nw represent the number of linewidth response branches. and These represent the line-direction route weight and the line width route weight, respectively.

[0031] The enhanced features are obtained by channel-joining the features to be enhanced, the linear aggregated features, and the linewidth aggregated features, followed by fusion mapping and residual connection, and are represented as follows: In the formula, [] represents the channel dimension splicing operation. Indicates fused convolutional mapping, This indicates the enhanced characteristics of the linear-linewidth coupling response.

[0032] S3. Input the enhanced features into the decoder for decoding. At the shallow jump connection between the encoder and decoder, generate edge cueing, line cueing and direction estimation features based on the current decoded features and the shallow encoded features to be transmitted.

[0033] In this embodiment, the current decoded feature is first upsampled to the same spatial size as the shallow skip features through bilinear interpolation, and then channel projection is performed through 1×1 convolution to generate the projected current decoded feature X. p and the shallow encoded features S after projection p .

[0034] Generate edge hint maps and line hint maps based on the current decoded features after projection: , Where E is the edge tooltip, L is the line tooltip, and f is the line tooltip. e and f l These represent edge hint generation mapping and line hint generation mapping, respectively, and σ represents the Sigmoid function.

[0035] The orientation estimation feature T is generated based on the projected shallow coding features. In this embodiment, the horizontal and vertical gradients are calculated using the Sobel operator, and three channel features—gradient magnitude, orientation cosine, and orientation sine—are constructed. , , , , In the formula, and Let M be the horizontal gradient and C be the vertical gradient, respectively; M be the gradient magnitude; C be the direction cosine feature; and S be the direction sine feature. For the stable term (in this embodiment, ε can be taken as 1 × 10), -8 The orientation estimation feature T is composed of three channels: gradient magnitude M, orientation cosine C, and orientation sine S, and is used to characterize the gradient orientation information of each pixel in the shallow features.

[0036] Subsequently, in S4, the projected current decoded features, projected shallow encoded features, edge cue maps, line cue maps, and orientation estimation features are concatenated to generate trusted transitive gating weights. ,in, This represents the trusted transit gating generation map, where G represents the trusted transit gating weight. This represents the current decoded features after projection. E represents the shallow jump connection feature after projection, L represents the edge cue, T represents the line cue, and T represents the orientation estimation feature.

[0037] S5. Filter the shallow coding features using trusted transitive gating weights to obtain trusted skip-connection features. The expression for the trusted skip-connection feature is: In the formula, Indicates a reliable jump connection feature. This indicates element-wise multiplication.

[0038] The trusted skip connection feature is fused with the current decoding feature to obtain the fused decoding feature. The expression for the fused decoding feature is: F s Indicates by X p With S r The initial fusion feature obtained by fusion, Q represents the pseudo-line suppression weight generated by the initial fusion feature, edge cue, and line body cue, and F represents the initial fusion feature. t This indicates the fusion decoding features.

[0039] Through the aforementioned line structure trusted gating jump connection transmission, shallow features do not enter the decoding end unconditionally, but are constrained by edge, line body and direction information, which can effectively suppress the propagation of background pseudo-lines such as tree tree edges, building outlines, and road edges to the decoding end.

[0040] S6. Extract edge branch features and line branch features from the fused decoding features respectively, and perform bidirectional modulation collaborative correction on the edge branch features and line branch features to obtain collaborative fusion features. Then, generate refined decoding features based on the collaborative fusion features.

[0041] Edge features and line features are calculated using edge projection branch and line projection branch: , , of which F e F1 represents edge features, and F1 represents line body features.

[0042] Then, through edge-to-line modulation, the edge branch feature F is... e Structural information injection into line branch feature F l The updated line body features are obtained. Its expression is: In the formula, This indicates the updated line body features. Indicates the modulation weight from the edge to the line body. This indicates element-wise multiplication.

[0043] By modulating from line body to edge, the main information of the updated line body features is fed back to the edge branch features to obtain the updated edge features. Its expression is: In the formula, This represents the updated edge features. This represents the modulation weight from the line to the edge.

[0044] Through the aforementioned bidirectional modulation, a bidirectional structural constraint is formed between the edge branches and the line body branches. The edge information assists in the positioning of the main body region, while the main body region information in turn constrains the boundary continuity.

[0045] The fused decoding features, updated edge features, and updated line features are collaboratively fused to obtain collaboratively fused features. Based on these collaboratively fused features, refined decoding features are generated, with the following expression: In the formula, This indicates refined decoding features. C represents the collaborative fusion feature, and C represents the updated edge feature. and updated line body features The jointly generated confidence gating, F t This represents the fused decoding features from input to the final prediction stage. The confidence gating expression is: ,in, This indicates a confidence-gated generation mapping.

[0046] The updated edge features and line features are then processed by the edge prediction head and line prediction head to obtain edge prediction maps and line prediction maps, respectively. The edge prediction maps and line prediction maps serve as auxiliary supervision signals during the training phase to constrain the collaborative representation of the transmission line boundary region and the main body region.

[0047] S7. Perform segmentation prediction on the refined decoded features and output the power transmission line identification result. The refined decoded features are input into the segmentation prediction layer, which includes an output convolutional kernel and a bias term to map the refined decoded features into a single-channel response map. The single-channel response map is normalized to the [0,1] interval using the Sigmoid activation function to obtain the power transmission line segmentation probability map Ys, which satisfies: In the formula, This represents the decoded features after edge-line co-refinement. This represents the segmentation prediction convolution kernel. The asterisk (*) represents the bias term, and the asterisk (*) represents the convolution operation. This represents the Sigmoid function. This represents the transmission line segmentation probability map. The transmission line segmentation probability map is compared with a preset threshold. If the pixel probability value is greater than or equal to the threshold, the pixel is determined to belong to the transmission line category; otherwise, it is determined to be the background category. A binarized transmission line recognition result map is output.

[0048] The power transmission line recognition network was trained using a power transmission line image dataset to obtain a trained power transmission line recognition model.

[0049] This embodiment employs a combination of main segmentation loss, edge auxiliary loss, and line auxiliary loss to optimize the model. The main segmentation loss consists of binary cross-entropy loss and soft Dice loss, used to constrain the consistency between the final transmission line prediction result and the actual labeled mask.

[0050] The principal segmentation loss is expressed as: Where P represents the final segmentation prediction result and Y represents the actual transmission line annotation mask.

[0051] The soft Dice loss is expressed as: , where s represents the smoothing term.

[0052] The edge-assisted loss is expressed as: ,in, Y represents the edge prediction result. e This represents the edge label obtained by differential dilation and erosion of the real mask.

[0053] The line auxiliary loss is expressed as: ,in, This indicates the prediction result for the line body.

[0054] The total loss function in this embodiment is: ,in, and These represent the weights of the edge-assisted loss and the line-assisted loss, respectively.

[0055] To verify the effectiveness of the method of the present invention, comparative experiments and ablation experiments were conducted on datasets A and B, respectively. The comparative experiments were used to verify the overall performance advantage of the method of the present invention compared with existing semantic segmentation methods; the ablation experiments were used to verify the actual contributions of the line-width enhancement module, the trusted gating transfer module, and the collaborative correction module to the power transmission line identification performance.

[0056] During the evaluation phase, Intersection over Union (IoU), Precision, Recall, and F1 score are used to quantitatively evaluate the power transmission line identification results. Considering that power transmission line targets in aerial images typically have low foreground pixel ratios, slender shapes, and sensitivity to local breaks, this embodiment first performs global statistics on the pixel-level prediction results of all images in the test set when calculating the evaluation indicators. The number of true positive pixels, false positive pixels, false negative pixels, and true negative pixels are accumulated separately. Then, based on the accumulated global pixel statistics, each evaluation indicator is calculated uniformly.

[0057] The intersection-over-union ratio, precision, recall, and F1 score satisfy the following relationships: , , , , where TP represents the number of pixels predicted as power lines and actually being power lines, FP represents the number of pixels predicted as power lines but actually being background, FN represents the number of pixels predicted as background but actually being power lines, TN represents the number of pixels predicted as background and actually being background, and ε represents a stabilizing term used to avoid a denominator of zero.

[0058] Comparative experiments were conducted to verify the overall performance advantages of this invention compared to existing semantic segmentation methods. The comparison methods included TransNeXt, DeepLabv3+, OffSeg, EDAFormer, UNetMamba, UNet++, and SegFormer. All methods were tested using the same test set partitioning and the same evaluation metrics.

[0059] The segmentation results of different methods on dataset A are as follows: Figure 3 As shown in Table 1, the test results demonstrate that on dataset A, the method of this invention achieves an IoU of 78.08% and an F1 score of 87.69%, outperforming the listed comparison methods. Compared to the UNet++ method, the method of this invention improves IoU by approximately 6.19 percentage points and F1 score by approximately 4.04 percentage points; compared to the SegFormer method, the method improves IoU by approximately 13.33 percentage points and F1 score by approximately 9.09 percentage points. These results indicate that in scenarios with relatively small data scale and low foreground pixel proportion of power transmission lines, this invention effectively enhances the structural response capability and overall recognition effect of slender power transmission line targets through enhanced line-to-linewidth coupling response, reliable gating jump connection transmission of line structure, and edge-line body collaborative decoding and confidence constraint refinement.

[0060] Table 1. Test results of the present invention on dataset A

[0061] The segmentation results of different methods on dataset B are as follows: Figure 4 As shown in Table 2, the test results demonstrate that on dataset B, the method of this invention achieves an IoU of 79.00% and an F1 score of 88.27%, maintaining superior performance in terms of comprehensive evaluation metrics. Although some comparative methods have slightly higher precision metrics, the method of this invention achieves better results in both IoU and F1 scores, indicating that this invention can better maintain the integrity and continuity of the power transmission line target while controlling missegmentation. For power transmission line recognition tasks, a high precision metric alone does not necessarily represent optimal segmentation quality, as excessive suppression of foreground prediction may lead to missed detections in thin line regions; IoU and F1 scores can more comprehensively reflect the balance between foreground recognition integrity and misdetection control.

[0062] Table 2 Test results of the present invention on dataset B

[0063] To further verify the contribution of each component module of this invention to the power transmission line identification performance, ablation experiments were conducted on datasets A and B, respectively. The experimental results are shown in Tables 3 and 4. The ablation experiments used the basic network without the introduction of each innovative module as the baseline. On this basis, line-to-linewidth coupling response enhancement (C1), line structure trusted gated jump connection transfer (C2), and edge-line body collaborative decoding and confidence constraint refinement (C3) were added in sequence. At the same time, edge / line body supervised loss was enabled. All experiments used the same dataset partitioning, training strategy, and evaluation metrics to ensure the comparability of the experimental results.

[0064] Table 3 Ablation experimental results of the present invention on dataset A

[0065] Table 4 Ablation experimental results of the present invention on dataset B

[0066] In the table, DWAMFE, TASF, and ELCR represent line-direction-linewidth coupling response enhancement, line structure trusted gating jump connection propagation, edge-line body collaborative decoding, and confidence constraint refinement, respectively. Table 3 shows that on dataset A, the model performance continuously improves with the addition of each innovative step. The IoU of the basic network is 74.91%, which increases to 76.98% after adding line-direction-linewidth coupling response enhancement, indicating that this step enhances the expressive power of transmission line direction and width information, improving the feature response effect of slender targets. Further introducing line structure trusted gating jump connection propagation further increases the IoU to 77.12%, demonstrating that trusted gating of shallow encoded features effectively suppresses background pseudo-line interference and improves the fusion quality of encoded and decoded features. Finally, by incorporating edge-line collaborative decoding and confidence constraint refinement, the IoU reached 78.08%, the Precision increased to 86.22%, and the F1 score reached 87.69%. This demonstrates that the collaborative optimization of the edge region and the line region can further improve the positioning accuracy and overall continuity of the transmission line boundary, enabling the model to achieve the best comprehensive performance.

[0067] As shown in Table 4, on dataset B, with the gradual introduction of each innovative step, the IoU gradually increased from 77.79% to 79.00%, and the F1 score increased from 87.51% to 88.27%, indicating that each component module can stably improve model performance. Specifically, the line-direction-linewidth coupling response enhancement module enhances the feature representation capability of transmission lines with different directions and widths; the line structure trusted gating jump connection transfer module effectively reduces mismatches caused by complex backgrounds and improves the effective utilization rate of shallow structural information; the edge-line body collaborative decoding and confidence constraint refinement module further improves the consistency between the transmission line edge and the main body region, improving the overall recognition accuracy and continuity. Experimental results show that the methods of this invention all have good effectiveness and synergy, jointly promoting the improvement of transmission line recognition performance.

[0068] Example 2

[0069] This embodiment provides an aerial power transmission line identification system based on reliable transmission and collaborative correction of line structure, including a data acquisition module, an coded feature extraction module, a line direction-line width enhancement module, a reliable gating transmission module, a collaborative correction module, and a segmentation output module.

[0070] The data acquisition module is used to acquire aerial images to be identified; The coding feature extraction module is used to input the aerial image to be identified into the encoder and extract multi-level coding features; The line-width enhancement module is used to perform line-width coupling response enhancement on deep coding features and bottleneck features respectively, extract the direction aggregation features and width aggregation features of the transmission line, and obtain the enhanced features based on the direction aggregation features and width aggregation features; The trusted gating transfer module is used to generate edge cueing, line cueing, and direction estimation features at the shallow jump connection between the encoder and decoder based on the current decoding features and the shallow coding features to be transferred. It also generates trusted transfer gating weights based on the edge cueing, line cueing, and direction estimation features, filters the shallow coding features using the trusted transfer gating weights to obtain trusted jump connection features, and fuses the trusted jump connection features with the current decoding features to obtain fused decoding features. The collaborative correction module is used to extract edge branch features and line branch features from the fused decoding features, and to perform bidirectional modulation collaborative correction on the edge branch features and line branch features to obtain collaborative fusion features. Then, refined decoding features are generated based on the collaborative fusion features. The segmentation output module is used to segment and predict the refined decoded features and output the power transmission line identification results.

[0071] Example 3

[0072] This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the aerial photography identification method for power transmission lines based on line structure trusted transmission and collaborative correction as described in Embodiment 1.

[0073] Example 4

[0074] This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the aerial photography identification method for power transmission lines based on reliable transmission and collaborative correction of line structure as described in Embodiment 1.

[0075] Example 5

[0076] This embodiment provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the steps of the aerial photography identification method for power transmission lines based on line structure trusted transmission and collaborative correction as described in Embodiment 1.

Claims

1. A method for aerial identification of power transmission lines based on reliable transmission and collaborative correction of line structures, characterized in that: Acquire aerial images to be identified, input the aerial images to be identified into a pre-trained recognition model for processing, and obtain the power transmission line identification results; The processing methods of the recognition model for the aerial image data to be recognized include: line-to-linewidth coupling response enhancement, line structure reliable gating jump connection transfer, and edge-line volume collaborative decoding and confidence constraint refinement. Line-to-linewidth coupling response enhancement involves inputting the aerial image to be identified into the encoder, extracting multi-level coding features, including shallow coding features, deep coding features, and bottleneck features, performing line-to-linewidth coupling response enhancement on the deep coding features and bottleneck features respectively, extracting the direction aggregation features and width aggregation features of the transmission line, and obtaining the enhanced features based on the direction aggregation features and width aggregation features. The line structure trusted gated jump connection transmission includes inputting enhanced features into the decoder for decoding; at the shallow jump connection between the encoder and decoder, generating edge cues and line body cues based on the current decoded features; extracting direction estimation features based on the shallow encoded features; combining the edge cues, line body cues, and direction estimation features to generate trusted transmission gate weights; using the trusted transmission gate weights to filter the shallow encoded features to obtain trusted jump connection features; and fusing the trusted jump connection features with the current decoded features to obtain fused decoded features. Edge-line body collaborative decoding and confidence constraint refinement include extracting edge branch features and line body branch features from the fused decoding features, performing bidirectional modulation collaborative correction on the edge branch features and line body branch features to obtain collaborative fusion features, and then generating refined decoding features based on the collaborative fusion features; The refined decoding features are segmented and predicted to output the power transmission line identification results.

2. The aerial power transmission line identification method based on reliable transmission and collaborative correction of line structure according to claim 1, characterized in that: Linear-to-linewidth coupling response enhancement includes: Local semantic pre-integration is performed on the feature F to be enhanced to obtain preprocessed feature F. p The method is as follows: ,in, This represents the basic convolutional mapping, used for the initial reorganization of local context in the input features; Preprocessing feature F p Input the line direction response branch group and the line width response branch group respectively to obtain the response in each direction and the response in each width. Based on the preprocessed features, line-direction routing weights and line-width routing weights are generated. The responses in each direction and each width are then weighted and fused to obtain line-direction aggregate features and line-width aggregate features. The method is as follows: , ,in, This represents the i-th linear response branch. Let Nd represent the j-th linewidth response branch, Nd represent the number of linewidth response branches, and Nw represent the number of linewidth response branches. and These represent the line-direction route weight and the line width route weight, respectively. The enhanced features are obtained by channel-joining the features to be enhanced, the linear aggregated features, and the linewidth aggregated features, followed by fusion mapping and residual connection. The method is as follows: Where [] represents channel dimension splicing operation, Indicates fused convolutional mapping, This indicates the enhanced characteristics of the linear-linewidth coupling response.

3. The aerial power transmission line identification method based on reliable transmission and collaborative correction of line structure according to claim 2, characterized in that: The linear response branch group includes multiple sets of asymmetric deep convolution branches, including at least 1×5 lateral linear branches, 5×1 longitudinal linear branches, 1×9 extended lateral linear branches and 9×1 extended longitudinal linear branches, which are used to respond to the local short-range direction and long-range extension structure of the transmission line, respectively. The linewidth response branch group includes multiple sets of deep convolution branches with different receptive fields, including at least a 3×3 fine-grained linewidth branch, a 5×5 neighborhood expansion linewidth branch, and a 3×3 sparse context linewidth branch with a hole ratio of 2, which are used to perceive the local region of the fine line, the neighborhood of the wider line body, and the sparse linear context, respectively.

4. The aerial power transmission line identification method based on reliable transmission and collaborative correction of line structure according to claim 1, characterized in that: The trusted transitive gating weight generation method is as follows: ,in, This represents the trusted transit gating generation mapping, where G represents the trusted transit gating weight, and X represents the trusted transit gating weight. p S represents the current decoded feature after projection. p E represents the shallow jump connection feature after projection, L represents the edge cue, T represents the line cue, and T represents the direction estimation feature. The method for obtaining reliable jump connection features is as follows: ,in, Indicates a reliable jump connection feature. This represents element-wise multiplication; The method for obtaining the fused decoding features is as follows: , Indicates by and The initial fusion feature obtained by fusion, Q represents the pseudo-line suppression weight generated by the initial fusion feature, edge cue, and line body cue, and F represents the initial fusion feature. t This indicates the fusion decoding features.

5. The aerial power transmission line identification method based on reliable transmission and collaborative correction of line structure according to claim 1, characterized in that: The bidirectional collaborative correction method is as follows: by edge-to-line modulation, the edge branch feature F is... e Structural information injection into line branch feature F l The updated line body features are obtained. The method is as follows: ,in, This indicates the updated line body features. Indicates the modulation weight from the edge to the line body. This represents element-wise multiplication; By modulating from line body to edge, the main information of the updated line body features is fed back to the edge branch features to obtain the updated edge features. The method is as follows: ,in, This represents the updated edge features. This represents the modulation weight from the line to the edge.

6. The aerial power transmission line identification method based on reliable transmission and collaborative correction of line structure according to claim 5, characterized in that: The method involves collaboratively fusing the fused decoding features, updated edge features, and updated line features to obtain collaboratively fused features, and then generating refined decoding features based on these collaboratively fused features. ,in, This indicates refined decoding features. C represents the collaborative fusion feature, and C represents the updated edge feature. and updated line body features Jointly generated confidence gating This represents the fused decoding features from the input to the final prediction stage.

7. The aerial power transmission line identification method based on reliable transmission and collaborative correction of line structure according to claim 6, characterized in that: The updated edge features and line features are then processed by the edge prediction head and line prediction head to obtain edge prediction maps and line prediction maps, respectively. The edge prediction maps and line prediction maps serve as auxiliary supervision signals during the training phase to constrain the collaborative representation of the transmission line boundary region and the main body region.

Images