Transmission tower corrosion detection method based on deep learning

Abstract

The invention belongs to the technical field of intelligent operation and maintenance of electric power facilities, and particularly relates to a transmission tower corrosion detection method based on deep learning, which comprises the following steps: firstly, constructing a transmission tower image data set D1, and carrying out classified marking according to a corrosion grade; then, constructing a structure perception feature extraction backbone network GQ-SAB, and inputting any transmission tower image B1 into the network to obtain a tower structure perception feature map B5, a tower structure perception feature map B11, a tower structure perception feature map B8 and a tower structure perception feature map B12; based on this, through an asymmetric double-path architecture of a structure-space decoupling enhancement module in the backbone network, linear and grid geometric features of the tower are extracted in a targeted manner, background noise is suppressed, accurate structured perception of the tower body is realized, and the problems of false alarm and missing alarm caused by feature confusion of a universal convolutional network are solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power facilities, specifically a method for detecting corrosion of transmission towers based on deep learning. Background Technology

[0002] Against the backdrop of the ongoing development of the national smart grid and the increasing prevalence of drone inspection technology, efficient, accurate, and automated condition monitoring of critical infrastructure such as transmission towers is a core requirement for ensuring the safe and stable operation of the power grid. Transmission towers are exposed to the elements for extended periods, and corrosion of their metal components is one of the most significant hidden dangers affecting their structural safety. Traditional corrosion detection methods have several insurmountable shortcomings: The manual inspection mode is outdated: it relies on maintenance personnel to climb or use binoculars for visual inspection, which is not only labor-intensive and inefficient, but also exposes workers to huge safety risks such as falls from heights, and the inspection results are easily affected by subjective factors.

[0003] Cloud-based identification solutions suffer from high latency and network dependence: Existing drone inspections mostly employ a model of on-site data collection and back-end analysis. Image data needs to be transmitted back to the cloud for calculation by large models, resulting in long delays in analysis results and hindering real-time on-site early warning. Furthermore, many power poles are located in remote areas with poor network signals, making data transmission a significant bottleneck.

[0004] The general detection model has insufficient adaptability: First, the model structure is large and computationally intensive, making it impossible to achieve real-time inference on drones or handheld devices with limited computing power; second, the general model lacks the ability to understand the unique hollow and linear geometric structure of power transmission towers, and is easily interfered with by complex backgrounds such as tree branches and mountain outlines, resulting in insufficient detection accuracy.

[0005] Therefore, the power inspection industry urgently needs a new type of lightweight intelligent inspection technology that is specifically designed for tower corrosion detection tasks, can operate efficiently on front-end equipment, and can fundamentally understand the target structure to combat complex background interference.

[0006] While traditional automated detection methods based on conventional image processing techniques, such as thresholding for rust color or edge detection for rust patches, have improved efficiency to some extent, these methods are extremely sensitive to changes in lighting, shadows, and background interference. In the complex outdoor environment where power transmission towers are located, the color of rust is easily confused with background elements such as soil and dead leaves. Furthermore, the mottled texture and irregular edges of rust are difficult for traditional operators to effectively capture, resulting in poor robustness and a high false detection rate. In recent years, with the rapid development of deep learning technology, especially the successful application of convolutional neural networks (CNNs) in object detection, end-to-end learning of multi-level feature extraction patterns has provided a highly promising technical path for rust detection of power transmission towers.

[0007] Despite significant progress in general object detection, deep learning technology still faces technical bottlenecks due to the unique nature of the task when applied directly to the specific task of detecting corrosion on power transmission towers.

[0008] Existing models lack prior understanding of the tower geometry, leading to severe confusion between foreground and background. General convolutional networks cannot effectively distinguish between detected objects and background, resulting in high false positive and false negative rates. Feature fusion mechanisms suffer from information overload, causing the loss of subtle corrosion features. Finally, conventional detection heads employ a flattened search strategy, indiscriminately searching across the entire high-resolution image, which is inefficient. Therefore, designing a deep learning model that can distinguish foreground from background, prevent the loss of key features, and efficiently utilize hierarchical relationships for collaborative reasoning, tailored to specific structures, is the core direction of current research.

[0009] Based on the above, the existing technology has the following problems: 1. Existing methods lack sufficient understanding of the structured features of transmission towers: The symmetric convolution kernels used in general convolutional networks are difficult to effectively extract the unique hollow and linear geometric features of transmission towers, resulting in serious feature confusion and a large number of false alarms and false negatives in the detection results.

[0010] 2. Existing methods have a crude feature fusion mechanism, which leads to the submergence of corrosion features: When fusing multi-scale features, existing models generally use simple feature splicing. Strong semantic features used to identify towers and weak semantic features used to identify corrosion are mixed indiscriminately. During the fusion process, corrosion feature signals are easily diluted or submerged by tower structural feature signals, making it difficult to identify corrosion on towers and causing missed detections.

[0011] 3. Existing methods rely on a single dimension for acquiring contextual information, resulting in poor performance in detecting diverse corrosion morphologies: Corrosion of transmission towers can progress from initial point corrosion to linear spreading corrosion and then to severe sheet corrosion, with significant differences in color and contrast. Existing models mostly rely on max pooling, such as SPPF, when aggregating contextual information. However, this mechanism cannot effectively capture the features of corrosion areas for small-area, weak-feature initial corrosion or dark corrosion that is difficult to distinguish from the tower's shadow, leading to insufficient ability to identify various key corrosion morphologies.

[0012] 4. The existing detection mode lacks hierarchy, resulting in a loss of accuracy: The existing detection head adopts a flat search strategy, searching for tiny rust targets indiscriminately throughout the entire image space. It does not take advantage of the strong spatial prior relationship that rust must exist on the tower, making it difficult for the model to detect small rust targets. Summary of the Invention

[0013] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A deep learning-based method for detecting corrosion on power transmission towers includes the following specific steps: S1, construct the power transmission tower image dataset D1, and classify and label it according to the corrosion level; divide the dataset into training set, validation set and test set according to the proportion; S2, construct the structure-aware feature extraction backbone network GQ-SAB, input any transmission tower image B1 into the structure-aware feature extraction backbone network, perform feature extraction, and obtain tower structure-aware feature map B5, tower structure-aware feature map B11, tower structure-aware feature map B8, and tower structure-aware feature map B12. S3. Construct a dynamic routing adaptive multi-scale feature fusion network GQ-DRN. Input B5, B11, B8, and B12 into the dynamic routing adaptive multi-scale feature fusion network to perform feature fusion and obtain corrosion feature enhancement fusion map N8, corrosion feature enhancement fusion map N9, and corrosion feature enhancement fusion map N12. S4. Construct the carrier-defect detection bidirectional gated collaborative reasoning head network GQ-COHEAD, input N8, N9, and N12 into the carrier-defect detection bidirectional gated collaborative reasoning head network, perform collaborative reasoning and prediction, and obtain the tower corrosion information detection results. S5. Construct the SACD-Net network model for sensing corrosion detection and early warning of power transmission tower structures; integrate the three parts GQ-SAB, GQ-DRN and GQ-CoHead to construct the complete SACD-Net model proposed in this invention; and input the collected images of power transmission towers into the model for supervised learning and optimization training. S6 deploys the trained tower structure perception corrosion detection and early warning network model into the power inspection and analysis system to identify and detect corrosion targets on transmission towers, and finally outputs hierarchically labeled images of towers and corrosion locations, corrosion level categories and confidence scores.

[0014] As a preferred embodiment of the deep learning-based method for detecting corrosion of power transmission towers according to the present invention, the specific steps of step S1 are as follows: Multiple types of images are acquired to obtain image data of the tower to be inspected; Based on location selection and corrosion level, the acquired image data is manually annotated in a two-level manner to obtain the transmission tower image dataset D1, which contains multiple corrosion levels.

[0015] As a preferred embodiment of the deep learning-based method for detecting corrosion of power transmission towers according to the present invention, the specific steps of S2 are as follows: Input any transmission tower image B1∈D1 into the CBS module for preliminary feature extraction and downsampling to generate tower structure perception feature map B2; In path one, B2 is first input into the C2f module to perform deep feature extraction, resulting in the tower structure perception feature map B3; then, B3 is input into the first SSDE module to extract the geometric features of the tower and effectively suppress background interference, resulting in the tower structure perception feature map B4. In path two, B2 is input into the second SSDE module for processing to obtain the tower structure perception feature map B5; Concat B4 and B5 along the channel dimension to obtain the tower structure perception feature map B6; perform Conv 5*5 convolution on B6 to obtain the tower structure perception feature map B11. Input B6 into the CBS module to obtain the tower structure perception feature map B7, and then input B7 into the third SSDE module to obtain the tower structure perception feature map B8. First, input B8 into the C2f module for deep feature extraction to obtain the tower structure perception feature map B9; then input B9 into the efficient multi-scale attention module (EMA) for processing to obtain the tower structure perception feature map B10; then input B10 into the alternating multi-scale pooling module (AMP) for processing to obtain the tower structure perception feature map B12.

[0016] As a preferred embodiment of the deep learning-based method for detecting corrosion of power transmission towers according to the present invention, the specific steps of S3 are as follows: In path one, a downsampling operation is performed on B5 to obtain the rust feature enhancement fusion image N1; then, N1 is input into the C2f module for deep feature extraction to obtain the rust feature enhancement fusion image N2. In path two, B12 is upsampled to obtain the rust feature enhancement fusion image N4; N4 and B8 are concatted along the channel dimension to obtain the rust feature enhancement fusion image N5; N5 is input into the C2f module for feature extraction to obtain the rust feature enhancement fusion image N6. In path three, B5 is input into the coordinate attention mechanism module (CA) to perform feature enhancement operation, resulting in the rust feature enhancement fusion map N3. N3 is then subjected to Conv 3*3 convolution operation to obtain the rust feature enhancement fusion map N10. Concat N2, N6 and B11 along the channel dimension to obtain the rust feature enhancement fusion image N7; N7 and N10 are input into the first CGDR module for context-guided dynamic routing fusion to obtain the rust feature enhancement fusion map N8; then, N5 and N8 are input into the second CGDR module to obtain the rust feature enhancement fusion map N9; and N9 is subjected to a Conv 1*1 convolution operation to obtain the rust feature enhancement fusion map N11. In path four, B12 and N11 are input into the third CGDR module for final fusion to obtain the rust feature enhancement fusion map N12.

[0017] As a preferred embodiment of the deep learning-based method for detecting corrosion of transmission towers according to the present invention, the specific steps of S4 are as follows: In branch one, N8 is input into the first BGC-HEAD module to obtain the tower corrosion prediction feature map D1; non-maximum suppression operation is performed on D1 to obtain the tower corrosion prediction feature map D4; In branch two, N9 is input into the second BGC-HEAD module to obtain the tower corrosion prediction feature map D2; non-maximum suppression operation is performed on D2 to obtain the tower corrosion prediction feature map D5; In branch three, N12 is input into the third BGC-HEAD module to obtain the tower corrosion prediction feature map D3; non-maximum suppression operation is performed on D3 to obtain the tower corrosion prediction feature map D6; Finally, D4, D5, and D6 are input into the weighted frame fusion (WBF) module to obtain the tower corrosion information detection results.

[0018] Compared with existing technologies: 1. To address the shortcomings of existing technologies in understanding the structural features of poles and towers, this invention utilizes an asymmetric dual-path architecture of the Structure-Space Decoupling Enhancement Module (SSDE) in the backbone network to specifically extract the linear and mesh geometric features of poles and towers while suppressing background noise. This achieves accurate structural perception of the pole and tower body and solves the problems of false positives and false negatives caused by feature confusion in general convolutional networks. 2. To address the problem of existing feature fusion mechanisms being too crude and causing corrosion features to be submerged, this invention uses a context-guided dynamic routing module (CGDR) to guide and adaptively enhance low-level corrosion detail features with high-level semantic features, effectively preventing the loss of weak corrosion features during the fusion process and improving the preservation and recognition capabilities of corrosion features. 3. To address the problem that existing technologies have a single dimension for obtaining contextual information and poor performance in detecting various forms of corrosion, this invention uses an alternating multi-scale pooling module (AMP) to aggregate saliency and contextual features in parallel, thereby obtaining more comprehensive multi-scale information and effectively improving the detection capability for various forms of corrosion targets, from point-like early stage to sheet-like heavy stage. 4. To address the issue of accuracy loss due to the lack of hierarchy in existing detection technologies, this invention utilizes a bidirectional gated collaborative detection head module (BGC-HEAD) to achieve hierarchical and efficient reasoning by leveraging the spatial prior relationship between towers and corrosion. This avoids inefficient use of computing resources and can also verify the accuracy of tower identification in reverse, ultimately improving the accuracy, efficiency, and reliability of transmission tower corrosion detection. Attached Figure Description

[0019] Figure 1 This is a flowchart of the technical solution of the present invention; Figure 2 This is a diagram of the backbone network architecture for structure-aware feature extraction in this invention. Figure 3 This is a structural-spatial decoupling enhancement module architecture diagram of the present invention; Figure 4 This is an architecture diagram of the structural feature extraction submodule of the present invention; Figure 5 This is a diagram of the alternating multi-scale pooling module architecture of the present invention; Figure 6 This is a diagram of the dynamic routing adaptive multi-scale feature fusion network architecture of the present invention; Figure 7 This is an architecture diagram of the context-guided dynamic routing module of the present invention; Figure 8 This is a diagram of the bidirectional gated collaborative inference head network architecture for carrier-defect detection in this invention. Figure 9 This is a diagram illustrating the architecture of the bidirectional gated collaborative detection head module of the present invention. Figure 10 This is a diagram of the SACD-Net architecture for the tower structure sensing corrosion detection and early warning network of the present invention; Figure 11 This is an image of a power transmission tower according to the present invention; Figure 12 This is an image output from the transmission tower detection of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0021] This invention provides a deep learning-based method for detecting corrosion on power transmission towers. Please refer to [link / reference]. Figures 1-12 The specific steps are as follows: S1, construct the power transmission tower image dataset D1, and classify and label it according to the corrosion level; divide the dataset into training set, validation set and test set according to the proportion; The specific steps of S1 are as follows: By using drone aerial photography or handheld smart terminal devices, multi-angle, high-resolution images of various types of in-service transmission towers, such as angle steel towers and steel pipe towers, are collected under different seasons, lighting conditions, and weather conditions to obtain image data of the towers to be inspected. The collected image data was manually annotated in a two-tiered manner: First, the overall location of each transmission tower in the image was selected and labeled as a tower; second, within the tower's annotation box, all corroded areas were further selected and classified according to the corrosion level, resulting in a transmission tower image dataset D1 containing multiple corrosion levels. The corrosion grades are based on the visual assessment standard for steel surface corrosion grades in the national standard GB / T 8923.1-2011, and are formulated in combination with the operation and maintenance experience of the power industry. This invention divides the corrosion state into three grades: Grade I: initial point or spot corrosion; Grade II: corrosion area expands, with flowing rust or flaky corrosion; Grade III: severe corrosion, with delamination, peeling or obvious material loss.

[0022] Example: To ensure the model can handle diverse real-world conditions, the dataset was constructed with priority given to sample collection in high-risk environments. For example, specialized drone inspections were conducted in typical areas such as coastal high-salt-fog regions, heavy industrial acid deposition zones, and old power lines to acquire a large number of high-resolution tower images. Additionally, some image data from the Public Power Inspection Challenge was integrated as a supplement. The final original image dataset contains 12,000 images, covering different types of towers, such as angle steel towers and steel pipe towers, in different seasons, under different lighting conditions, and in complex scenarios with partial occlusion, such as vegetation cover. The dataset includes approximately 5,000 images mainly showing Grade I corrosion, approximately 4,000 images showing Grade II corrosion, approximately 2,000 images showing Grade III corrosion, and approximately 1,000 images of towers without obvious corrosion. The annotation work is performed using the web-based open-source annotation platform CVAT (Computer Vision Annotation Tool). The annotation strategy employs a nested bounding box approach: first, a tower bounding box is used to precisely define the visible main structure of the tower in the image; then, within the tower bounding box, labels are used to meticulously annotate all identifiable corroded areas. This nested annotation method clearly defines the hierarchical relationship between the carrier and the defect. All annotation information is ultimately exported as a COCO JSON format file to support complex object detection tasks. Finally, the completed dataset D1, consisting of 12,000 images of transmission towers with various corrosion levels, was randomly divided into training, validation, and test sets in a ratio of 80%:10%:10% for subsequent model training, validation, and evaluation.

[0023] Among them, CVAT (Computer Vision Annotation Tool) is an open-source, web-based interactive image and video annotation tool developed by Intel. Its core purpose is to provide efficient dataset annotation support for the field of computer vision. It can be adapted to various mainstream annotation tasks such as object detection, image segmentation, and key point annotation, and helps to quickly build and optimize algorithm training data.

[0024] Nested bounding boxes are a labeling strategy used to represent hierarchical or subordinate relationships between objects. Here, the rust box is labeled inside the tower box, clearly indicating that the rust is a defect attached to the tower, which helps the model learn this spatial constraint relationship.

[0025] COCO JSON format is a dataset annotation format widely used for object detection, segmentation, and keypoint detection tasks. It stores image information, class labels, bounding box coordinates, segmentation masks, and other information in the form of JSON files and is supported by many mainstream deep learning frameworks.

[0026] S2, construct the structure-aware feature extraction backbone network GQ-SAB (Geometric Quality-Structure-Aware Backbone), input any transmission tower image B1 into the structure-aware feature extraction backbone network, perform feature extraction, and obtain tower structure-aware feature maps B5, B11, B8, and B12. The GQ-SAB network integrates multiple convolution types and attention mechanisms, enabling accurate extraction of transmission tower features, elimination of background noise, focusing on key corrosion areas, and enhanced perceptual capabilities. This allows the backbone network to perform preliminary extraction of tower geometry and initial corrosion perception when outputting multi-scale features. The architecture of the GQ-SAB network is as follows: Figure 2 As shown.

[0027] The specific steps of S2 are as follows: Input any transmission tower image B1∈D1 into the CBS module for preliminary feature extraction and downsampling to generate tower structure perception feature map B2; In path one, B2 is first input into the C2f module to perform deep feature extraction, resulting in the tower structure perception feature map B3; then, B3 is input into the first SSDE module to extract the geometric features of the tower and effectively suppress background interference, resulting in the tower structure perception feature map B4. In path two, B2 is input into the second SSDE module for processing to obtain the tower structure perception feature map B5; The SSDE module adopts an asymmetric dual-path architecture, which uses asymmetric convolution and dilated convolution to specifically extract the linear and mesh geometric features of the tower, capture the macroscopic outline of the target, achieve deep suppression of background noise, and improve the model's ability to distinguish between hollow targets such as towers and complex backgrounds.

[0028] The structure of the Structure-Spatial Decoupling Enhancement (SSDE) module is as follows: Figure 3 The steps of the SSDE module are shown below: [Represented as A1 and A2] A1: Denote the input feature map of the SSDE module as E1. Perform a split operation on E1 to obtain the tower perception enhancement feature map E2 and the tower perception enhancement feature map E7 respectively.

[0029] In path one, E2 is input into the Structural Feature Extraction (SFE) submodule to extract fine tower geometric features, resulting in tower perception enhancement feature map E3. Next, E3 is input into the CBS submodule for depth refinement, generating tower perception enhancement feature map E4. Then, E4 is sequentially input into the bottleneck layer for channel compression, resulting in tower perception enhancement feature map E5. E5 is then input into the SAM submodule for spatial noise reduction, resulting in tower perception enhancement feature map E6. Finally, max pooling is performed on E6 to obtain tower perception enhancement feature map E12.

[0030] In path two, E7 is first passed through the CBS submodule to obtain the tower perception enhancement feature map E8. Then, E8 is downsampled by max pooling to extract the macroscopic contour and contextual features of the target, resulting in the tower perception enhancement feature map E9. E4 and E9 are concatenated to obtain the tower perception enhancement feature map E10. E10 is then input into the bottleneck layer for channel compression and information fusion to obtain the tower perception enhancement feature map E11. Finally, E11 is input into the spatial attention submodule (SAM) for spatial noise reduction and enhancement to obtain the tower perception enhancement feature map E13.

[0031] A2: Concat E12 and E13 to obtain the tower perception enhancement feature map E14.

[0032] The architecture diagram of the Structural Feature Extraction (SFE) submodule is as follows: Figure 4 As shown.

[0033] In branch one, a Conv 3*3 convolution operation is performed on E2 to extract common local features, resulting in the tower structure extraction feature map F1. In branch two, a Conv 3*1 convolution operation is performed on E2 to initially extract the linear features of the tower structure in the vertical direction, resulting in the tower structure extraction feature map F2; subsequently, a Conv 1*3 convolution operation is performed on F2 to further fuse the horizontal information, resulting in the tower structure extraction feature map F4. In branch three, a Conv 1*3 convolution operation is performed on E2 to initially extract the linear features of the tower structure in the horizontal direction, resulting in the tower structure extraction feature map F3; subsequently, a Conv 3*1 convolution operation is performed on F3 to further fuse the vertical information, resulting in the tower structure extraction feature map F5. In branch four, E2 is subjected to a Dilated Conv 5*5 dilated convolution operation with an inflation rate of 5 to expand the receptive field and capture the contextual information of the sparse structure, resulting in the tower structure extraction feature map F6. Concat F1, F4, F5, and F6 along the channel dimension to obtain E3. E3 contains features from different branches that are general, linear, and contextual.

[0034] Concat B4 and B5 along the channel dimension to obtain the tower structure perception feature map B6; perform Conv 5*5 convolution on B6 to obtain the tower structure perception feature map B11. Input B6 into the CBS module to obtain the tower structure perception feature map B7, and then input B7 into the third SSDE module to obtain the tower structure perception feature map B8. First, input B8 into the C2f module for deep feature extraction to obtain the tower structure perception feature map B9; then input B9 into the efficient multi-scale attention module (EMA) for processing to obtain the tower structure perception feature map B10; then input B10 into the alternating multi-scale pooling module (AMP) for processing to obtain the tower structure perception feature map B12.

[0035] The AMP module adopts a parallel dual-stream heterogeneous pyramid architecture, which can simultaneously acquire salient details and contextual statistical information of corrosion, helping to improve the network's detection accuracy for corrosion targets of different shapes and sizes, such as point corrosion and sheet corrosion.

[0036] The architecture of Alternating Multi-scale Pooling (AMP) is as follows: Figure 5 As shown.

[0037] First, perform a Conv 3*3 convolution operation on B10 to obtain a multimorphic corrosion aggregate feature map A1. Then, input A1 into the BN submodule and perform a normalization operation to obtain a multimorphic corrosion aggregate feature map A2. A2 is then activated by the ReLU function to generate a multimorphic corrosion aggregate feature map A3. In path one, A3 first undergoes an average pooling operation to obtain a multimorphic corrosion aggregated feature map A4; A4 then undergoes a max pooling operation to obtain a multimorphic corrosion aggregated feature map A5; A5 then undergoes an average pooling operation to obtain a multimorphic corrosion aggregated feature map A6; A4, A5, and A6 are concatenated along the channel dimension to obtain a multimorphic corrosion aggregated feature map A7, which incorporates multi-level hybrid pooling information; finally, A7 is input into the CBS submodule to obtain a multimorphic corrosion aggregated feature map A8.

[0038] In path two, A3 first undergoes a max pooling operation to obtain a multimorphic corrosion aggregated feature map A9; then, A9 undergoes an average pooling operation to obtain a multimorphic corrosion aggregated feature map A10; finally, A10 undergoes a max pooling operation to obtain a multimorphic corrosion aggregated feature map A11; A9, A10, and A11 are concatenated along the channel dimension to obtain a multimorphic corrosion aggregated feature map A12, which incorporates multi-level hybrid pooling information; finally, A12 is input into the CBS submodule to obtain a multimorphic corrosion aggregated feature map A13.

[0039] Finally, A8 and A13 are concatted along the channel dimension to obtain the output of the entire module, namely the multimorphic corrosion aggregate feature map B12.

[0040] Example: Take any transmission tower image B1 ∈ Transmission Tower Image Dataset D1, with a size of 640×640 pixels and 3 channels, and input it into the Structure-Aware Feature Extraction Backbone Network GQ-SAB; First, input B1 into the CBS module for preliminary feature extraction and downsampling to obtain the tower structure perception feature map B2, which has a size of 320×320 pixels and 64 channels; In path one, B2 first passes through the C2f module to obtain the tower structure perception feature map B3, which has a size of 320×320 pixels and 128 channels; then, B3 is input into the first SSDE module for processing to obtain the tower structure perception feature map B4, which has a size of 320×320 pixels and 128 channels. In path two, B2 is input into the second SSDE module for processing to obtain the tower structure perception feature map B5, which has a size of 320×320 pixels and 128 channels. Concatenate B4 and B5 to obtain tower structure perception feature map B6, which has a size of 320×320 pixels and 256 channels. Input B6 into the CBS module to obtain tower structure perception feature map B7, which has a size of 160×160 pixels and 512 channels; then input B7 into the third SSDE module for processing to obtain tower structure perception feature map B8, which has a size of 160×160 pixels and 512 channels. Finally, B8 is input into the C2f module to obtain the tower structure perception feature map B9, which has a size of 80×80 pixels and 1024 channels. B9 is then processed by the efficient multi-scale attention module (EMA) to obtain the tower structure perception feature map B10, which also has a size of 80×80 pixels and 1024 channels. B10 is then input into the alternating multi-scale pooling module (AMP) to obtain the tower structure perception feature map B12, which has a size of 80×80 pixels and 1024 channels. At this point, the GQ-SAB backbone network has completed the processing of the input image, generating tower structure perception feature maps B5, B11, B8, and B12.

[0041] C2f (Cross Stage Partial Faster) is an efficient feature extraction module introduced in the YOLOv8 model. Its core principle is to borrow the cross-stage partial connection idea of ​​CSPNet (Cross Stage Partial Network), split the input feature stream into two branches and perform cross-stage feature fusion, optimizing the gradient propagation path while retaining key feature information. The core purpose of this module is to obtain richer gradient flow information and more comprehensive feature representation while reducing the amount of computation and parameters, thereby improving the model's feature extraction efficiency and detection performance. It is one of the core components of the YOLOv8 model to achieve efficient inference and accurate detection.

[0042] The Efficient Multi-Scale Attention (EMA) module is a lightweight feature enhancement module. Its core principle is to build feature dependencies across spatial dimensions in the channel dimension. Through a lightweight computing architecture, it efficiently captures contextual information at different scales, achieving adaptive aggregation of multi-scale features. The core purpose of this module is to significantly enhance the semantic expressive power and multi-scale feature perception capability of feature maps while controlling the amount of computation and parameters. This helps the model accurately capture the detailed features of the target and the global contextual relationships, making it suitable for feature enhancement in deep learning vision tasks.

[0043] The bottleneck layer is a classic structural unit in convolutional neural networks. Its core principle is to embed 1×1 convolution operations, which flexibly compress or adjust the number of feature channels while maintaining the spatial resolution of the feature map, and at the same time realize the fusion and recombination of cross-channel information. The core purpose of this structural unit is to optimize the compactness of feature representation and enhance the correlation of cross-channel features while reducing the computational cost and parameter count of the model and improving inference efficiency. It is often used in deep convolutional networks to alleviate the gradient vanishing problem and improve feature transfer efficiency, and is a key component for building efficient deep networks.

[0044] The Spatial Attention Module (SAM) is a commonly used spatial attention mechanism module, as a component of the Convolutional Block Attention Module (CBAM). Its implementation involves first performing average pooling and max pooling along the channel dimensions of the input feature map to compress channel information. Then, the two are concatenated and passed through a convolutional layer to learn spatial correlations. Finally, a spatial attention weight map is generated using the sigmoid function and multiplied back onto the original feature map. Its core purpose is to generate a spatial weight map by analyzing the spatial information of the feature map, thereby enhancing the feature responses of key regions in the image and suppressing irrelevant background regions.

[0045] S3. Construct a dynamic routing adaptive multi-scale feature fusion network GQ-DRN (Guided Quality-driven Routing Net). Input B5, B11, B8, and B12 into the dynamic routing adaptive multi-scale feature fusion network to perform feature fusion and obtain corrosion feature enhancement fusion map N8, corrosion feature enhancement fusion map N9, and corrosion feature enhancement fusion map N12. The GQ-DRN network can dynamically and adaptively modulate based on high-level semantics, selectively amplifying feature responses related to weak corrosion targets and preventing them from being diluted or lost when fused with strong tower structure features. The architecture of GQ-DRN is as follows: Figure 6 As shown.

[0046] In the GQ-DRN network, B5, B11, B8, and B12 are processed through multiple parallel paths: In path one, a downsampling operation is performed on B5 to obtain the rust feature enhancement fusion image N1; then, N1 is input into the C2f module for deep feature extraction to obtain the rust feature enhancement fusion image N2. In path two, B12 is upsampled to obtain the rust feature enhancement fusion image N4; N4 and B8 are concatted along the channel dimension to obtain the rust feature enhancement fusion image N5; N5 is input into the C2f module for feature extraction to obtain the rust feature enhancement fusion image N6. In path three, B5 is input into the coordinate attention mechanism module (CA) to perform feature enhancement operation, resulting in the rust feature enhancement fusion map N3. N3 is then subjected to Conv 3*3 convolution operation to obtain the rust feature enhancement fusion map N10. Concat N2, N6 and B11 along the channel dimension to obtain the rust feature enhancement fusion image N7; N7 and N10 are input into the first CGDR module for context-guided dynamic routing fusion to obtain the rust feature enhancement fusion map N8; then, N5 and N8 are input into the second CGDR module to obtain the rust feature enhancement fusion map N9; and N9 is subjected to a Conv 1*1 convolution operation to obtain the rust feature enhancement fusion map N11. In path four, B12 and N11 are input into the third CGDR module for final fusion to obtain the rust feature enhancement fusion map N12.

[0047] The architecture of the Context-Guided Dynamic Routing (CGDR) module is as follows: Figure 7 As shown.

[0048] The CGDR module achieves adaptive channel weighting of low-level detail features by multiplying dynamic routing weights with the rust adaptive fusion feature map, thereby enabling high-level semantics to weight low-level detail features. Simultaneously, the CGDR module introduces residual compensation to generate smooth and stable base features, which are then fused with the weighted features. This guides the fusion process of high-level semantics and low-level detail features, helping to amplify subtle rust and suppress irrelevant information.

[0049] This module receives two corrosion feature enhancement fusion maps as input: let the input of branch one and branch two be the corrosion feature enhancement fusion map R1, and let the input of branch three be the corrosion feature enhancement fusion map R2.

[0050] In branch one, a Conv 1*1 convolution operation is performed on R1 to extract preliminary features and generate the rust adaptive fusion feature map R3; then R3 is smoothed by average pooling (Avgpool) to enhance its contextual information and obtain the rust adaptive fusion feature map R4. In branch two, a DWConv 3*3 operation is performed on R1 to extract preliminary spatial features, resulting in the corrosion adaptive fusion feature map R5. R5 then undergoes multi-dimensional feature decomposition through three parallel convolutional branches: a Conv 1*1 convolution operation is performed on R5 to obtain the corrosion adaptive fusion feature map R6; a Conv 3*3 convolution operation is performed on R5 to obtain the corrosion adaptive fusion feature map R7; and a Dilated Conv 5*5 dilated convolution operation with a dilation rate of 3 is performed on R5 to obtain the corrosion adaptive fusion feature map R8. Subsequently, R6, R7, and R8 are concatenated along the channel dimension to obtain the corrosion adaptive fusion feature map R9. In branch three, R2 first undergoes a global average pooling (Avgpool) operation to compress its spatial information, resulting in the corrosion adaptive fusion feature map R10. Then, R10 is input into the multilayer perceptron (MLP) module for deep nonlinear transformation to learn the complex relationships between channels, yielding the corrosion adaptive fusion feature map R11. Finally, the Softmax activation function is applied to R11 to generate the corrosion adaptive fusion feature map R12. Multiply R12 and R9 element-wise to obtain the rust adaptive fusion feature map R13 after dynamic routing modulation; perform residual compensation addition (ADD) operation on R13 and R4 to obtain the rust adaptive fusion feature map R14.

[0051] Example: The four pole structure perception feature maps output in step S2 are input into the Dynamic Routing Adaptive Multi-Scale Feature Fusion Network (GQ-DRN). These four pole structure perception feature maps are: B5, 320×320 pixels, 128 channels; B11, 160×160 pixels, 256 channels; B8, 160×160 pixels, 512 channels; and B12, 80×80 pixels, 1024 channels.

[0052] In path one, a downsampling operation is performed on B5 with a step size of 2 to obtain a rust feature enhancement fusion image N1 with a size of 160×160 pixels and 128 channels; N1 is input into the C2f module for deep feature extraction to obtain a rust feature enhancement fusion image N2 with a size of 160×160 pixels and 256 channels. In path two, UPsample is performed on B12 to obtain the rust feature enhancement fusion image N4, with a size of 160×160 pixels and 512 channels. N4 is then concatenated with B8 to obtain the rust feature enhancement fusion image N5, with a size of 160×160 pixels and 1024 channels. Feature image N5 is then input into the C2f module for feature extraction to obtain the rust feature enhancement fusion image N6, with a size of 160×160 pixels and 1024 channels. In path three, B5 is input to the coordinate attention mechanism (CA) module for feature enhancement, resulting in a rust feature enhancement fusion image N3 with a size of 320×320 pixels and 128 channels. A Conv3*3 convolution operation is then performed on N3 with a stride of 2 and an output of 512 channels. After adjusting the size and channels, a rust feature enhancement fusion image N10 with a size of 60×160 pixels and 512 channels is obtained. N2, N6 and B11 are concatted along the channel dimension to obtain the rust feature enhancement fusion image N7, which has a size of 160×160 pixels and 1536 channels; N7 and N10 are input into the first CGDR module for intelligent fusion. This yields a rust feature enhanced fusion map N8, with dimensions of 160×160 pixels and 512 channels. N5 and N8 are input into the second CGDR module for intelligent fusion. This yields a rust feature enhancement fusion image N9, with dimensions of 160×160 pixels and 1024 channels. A Conv1*1 convolution operation is then performed on N9 to obtain a rust feature enhancement fusion image N11, also with dimensions of 160×160 pixels and 1024 channels. In path four, N11 and B12 are input into the third CGDR module for intelligent fusion. The resulting rust feature enhanced fusion image N12 has a size of 80×80 pixels and 1024 channels.

[0053] Among them, the Coordinate Attention (CA) module is an attention mechanism focused on capturing spatial location information and orientation sensitivity of feature maps. Its core principle is to decompose channel attention into two parallel one-dimensional feature encoding processes, capturing spatial orientation dependencies along the horizontal and vertical directions respectively, thereby generating an attention weight map that simultaneously contains precise location information and channel relationships. The core purpose of this mechanism is to help the model more accurately locate target positions and identify target features. Especially in visual tasks that require clear spatial orientation relationships, it can effectively enhance the spatial perception ability of feature maps and improve the detection and recognition accuracy of the model.

[0054] Multi-Layer Perceptron (MLP) is a basic form of feedforward neural network. Its core structure consists of one or more fully connected layers (or equivalent 1×1 convolutional layers), which are connected by nonlinear activation functions to introduce nonlinearity. The core purpose of this network is as a controller or feature converter to learn complex nonlinear relationships between input features and map them to a new feature space or generate control signals. It is widely used in machine learning tasks such as classification, regression, and feature transformation.

[0055] Depthwise Separable Convolution (DWConv) is an efficient architecture for reducing the computational cost and parameter count of convolution operations. It decomposes the standard convolution operation into two independent steps: first, spatial filtering is performed using depthwise convolution without changing the number of channels; then, pointwise convolution is performed to linearly combine and increase / decrease the dimensionality between channels. Its core purpose is to approximate the performance of standard convolution with significantly less computation and fewer parameters while maintaining feature extraction capabilities, thereby improving the model's inference efficiency. This is particularly suitable for lightweight network designs in resource-constrained scenarios such as mobile devices.

[0056] S4. Construct the GQ-COHEAD (Gated Quality-enhanced Collaborative Head) bidirectional gated collaborative inference head network for carrier-defect detection. Input N8, N9, and N12 into the carrier-defect detection bidirectional gated collaborative inference head network to perform collaborative inference and prediction, and obtain the tower corrosion information detection results; the architecture is as follows. Figure 8 As shown.

[0057] The GQ-COHEAD is a network architecture that predicts and outputs results from a corrosion feature enhancement fusion map. The corrosion feature enhancement fusion map output by the GQ-DRN network is fed into three BGC-HEAD modules. The bidirectional gating collaboration mechanism within BGC-HEAD is used to decouple and collaboratively infer the detection tasks of towers and corrosion, thereby improving the ability to identify towers and corrosion. Furthermore, a weighted bounding box fusion module (WBF) is used for cross-scale fusion, resulting in a more accurate unique outcome.

[0058] The specific steps of S4 are as follows: In branch one, N8 is input into the first BGC-HEAD module to obtain the tower corrosion prediction feature map D1; non-maximum suppression (NMS) operation is performed on D1 to obtain the tower corrosion prediction feature map D4; In branch two, N9 is input into the second BGC-HEAD module to obtain the tower corrosion prediction feature map D2; non-maximum suppression (NMS) operation is performed on D2 to obtain the tower corrosion prediction feature map D5; In branch three, N12 is input into the third BGC-HEAD module to obtain the tower corrosion prediction feature map D3; Non-maximum suppression (NMS) operation is performed on D3 to obtain the tower corrosion prediction feature map D6; Finally, D4, D5, and D6 are input into the weighted bounding box fusion (WBF) module to obtain the tower corrosion information detection results. The tower corrosion information detection results are detection images that include the tower and corrosion coordinates, its corrosion level, and confidence level.

[0059] The BGC-HEAD (Bi-directional Gated Co-detection Head) module features a bidirectional gated co-detection head. The module architecture diagram is shown below. Figure 9 The steps of the BGC-HEAD module are shown as B1, B2, etc.

[0060] The BGC-HEAD module employs a dual-stream parallel feature processing architecture and introduces a cross-gated modulation mechanism. The key feature of this design is that the two feature maps not only extract features independently but also generate their own gating signals, which are then cross-fused into each other. This bidirectional gating mechanism enables reasoning from locating the tower to determining corrosion.

[0061] Let the input of the BGC-HEAD module be the rust feature enhancement fusion image H1; B1. Perform a Conv 3*3 convolution operation on H1 to obtain the basic feature map H2; B2. Process H2 in parallel using two paths: In path one, H2 is subjected to a dilated Conv3 * 3 convolution with a dilation rate of 1 to obtain the corrosion structure feature map H3. Then, H3 is subjected to a dilated Conv3 * 3 convolution with a dilation rate of 3 to obtain a larger receptive field, resulting in the corrosion structure feature map H5. Next, H5 is subjected to a Conv1 * 1 convolution to obtain the corrosion structure feature map H7. Finally, a sigmoid activation function is applied to H7 to obtain the corrosion gating signal H8. In path two, a Conv1*3 asymmetric convolution operation is performed on H2 to obtain the tower structure feature map H4; then, a Conv3*1 asymmetric convolution operation is performed on H4 to efficiently extract the linear structure of the overall contour, resulting in the tower structure feature map H6; next, a Conv1*1 convolution operation is performed on H6 to obtain the tower structure feature map H10; finally, the Sigmoid activation function is applied to H10 to obtain the tower gating signal H11. B3. Multiply H8 and H6 element by element to obtain the tower positioning collaborative feature map H9; then, add H9 and H6 element by element to obtain the corrosion location prediction feature map H13. B4. Multiply H11 and H5 element-wise to obtain the corrosion detection collaborative feature map H12; then, add H12 and H5 element-wise to obtain the overall bounding box prediction feature map H14. B5. Input H13 and H14 into the weighted frame fusion submodule (WBF) to obtain the tower corrosion prediction feature map H15.

[0062] Example: In this example, when the input to the BGC-HEAD module is a rust feature enhancement fusion image N8 with a size of 160×160 pixels and 512 channels, N8 is denoted as H1. The specific process is as follows: First, H1 is input into a Conv3*3 convolutional layer to perform a convolution operation, resulting in a basic feature map H2 with a size of 160×160 pixels and 512 channels. In path one, H2 is subjected to a DilatedConv3*3 dilated convolution with a dilation rate of 1 to obtain the rust structure feature map H3, which has a size of 160×160 pixels and 512 channels. Then, H3 is subjected to a DilatedConv3*3 dilated convolution with a dilation rate of 3 to obtain the rust structure feature map H5, which has a size of 160×160 pixels and 512 channels. Next, H5 is subjected to a Conv1*1 convolution to obtain the rust structure feature map H7, which has a size of 160×160 pixels and 512 channels. Finally, a Sigmoid activation function is applied to H7 to generate the rust gate signal H8, which has a size of 160×160 pixels and 512 channels. In path two, a Conv1*3 asymmetric convolution operation is performed on H2 to obtain the tower structure feature map H4, with a size of 160×160 pixels and 512 channels; then, a Conv3*1 asymmetric convolution operation is performed on H4 to obtain the tower structure feature map H6, with a size of 160×160 pixels and 512 channels; next, a Conv1*1 convolution operation is performed on H6 to obtain the tower structure feature map H10, with a size of 160×160 pixels and 512 channels; finally, the Sigmoid activation function is applied to H10 to generate the tower gating signal H11, with a size of 160×160 pixels and 512 channels. Multiply H8 and H6 element-wise to obtain the tower positioning collaborative feature map H9, which has a size of 160×160 pixels and 512 channels; then, add H9 and H6 element-wise to obtain the corrosion location prediction feature map H13, which has a size of 160×160 pixels and 512 channels. Element-wise multiplication of H11 and H5 yields the corrosion detection collaborative feature map H12, with a size of 160×160 pixels and 512 channels; then, element-wise addition of H12 and the unmodulated H5 yields the overall bounding box prediction feature map H14, with a size of 160×160 pixels and 512 channels. Input H13 and H14 into the weighted box fusion submodule (WBF) to obtain the tower corrosion prediction feature map H15, which is the tower corrosion prediction feature map D1, with a size of 160×160 pixels and 512 channels. Similarly, inputting N9 into the second BGC-HEAD module yields the tower corrosion prediction feature map D2, with a size of 160×160 pixels and 1024 channels. Inputting N12 into the third BGC-HEAD module yields the tower corrosion prediction feature map D3, with a size of 80×80 pixels and 1024 channels. Performing non-maximum suppression (NMS) operation on D1 yields tower corrosion prediction feature map D4; performing NMS operation on D2 yields tower corrosion prediction feature map D5; performing NMS operation on D3 yields tower corrosion prediction feature map D6. The Intersection over Union (IoU) threshold for NMS is set to 0.6. For all bounding boxes in each result set that predict the same class and have an overlap of more than 0.6, only the predicted box with the highest confidence score is retained, and the remaining overlapping redundant boxes are suppressed. Finally, D4, D5, and D6 are input into the Weighted Boundary Fusion (WBF) module for final accurate fusion. The WBF's Intersection over Union (IoU) threshold is set to 0.7 for clustering overlapping boxes, and the confidence threshold is set to 0.01 for filtering low-scoring boxes. The WBF module first groups predicted boxes from different scales that may point to the same rust based on the spatial overlap of the bounding boxes (IoU > 0.7) and the rust level. Then, for each cluster, the WBF module calculates a new, weighted average fused bounding box coordinate [x', y', w', h'] and a fused confidence score Score', using the confidence scores of all boxes within the cluster as weights. Finally, the WBF module outputs the tower corrosion information detection results, which are detection images containing the tower and corrosion coordinates, its corrosion level and confidence level.

[0063] Non-Maximum Suppression (NMS) is a core post-processing operation for object detection. Its core principle is to retain the prediction box with the highest confidence through an iterative screening mechanism and suppress all redundant prediction boxes whose overlap with the box exceeds a preset threshold. The core purpose of this operation is to eliminate highly overlapping redundant boxes in the object detection results, ensuring that each detected target corresponds to only one unique bounding box, thereby improving the simplicity and accuracy of the detection results. It is a key step in the object detection process to remove duplicate predictions.

[0064] Weighted Box Fusion (WBF) is a post-processing technique for object detection. Its core principle is to cluster multiple overlapping predicted bounding boxes pointing to the same target from different scale feature layers or different detection models. Using the confidence scores of each predicted box as weights, a new fused bounding box and its corresponding new confidence score are calculated through weighted averaging. The core purpose of this module is to integrate multi-source prediction information, effectively reduce the localization error of a single predicted box, and improve the accuracy and robustness of object detection results. It is especially suitable for detection tasks with multi-scale targets or severely overlapping predicted boxes in complex scenes.

[0065] S5, the SACD-Net model for tower structure-aware corrosion detection and early warning; integrating GQ-SAB, GQ-DRN, and GQ-CoHead to construct the complete SACD-Net model (Structure-Aware Corrosion Detection Network) proposed in this invention; and inputting the collected transmission tower images into the model for supervised learning and optimization training. The architecture of the SACD-Net tower structure-aware corrosion detection and early warning network is as follows. Figure 10 As shown.

[0066] (1) Input any transmission tower image B1 into the structure-aware feature extraction backbone network GQ-SAB for feature extraction to obtain tower structure-aware feature map B5, tower structure-aware feature map B11, tower structure-aware feature map B8 and tower structure-aware feature map B12. (2) Input B5, B11, B8, and B12 into the dynamic routing adaptive multi-scale feature fusion network GQ-DRN to perform feature fusion and obtain corrosion feature enhancement fusion map N8, corrosion feature enhancement fusion map N9, and corrosion feature enhancement fusion map N12. (3) Input N8, N9, and N12 into the carrier-defect detection bidirectional gated collaborative inference head network GQ-COHEAD to output the tower corrosion information detection result. The tower corrosion information detection result is a detection image containing the tower and corrosion coordinates, its corrosion level, and confidence level. Subsequently, based on the training set and validation set divided in step S1, the data is input into the SACD-Net model for supervised learning and optimization training until the model's performance metrics on the validation set converge, ultimately resulting in a trained detection model capable of accurately identifying towers and corrosion.

[0067] Example: Based on the training set of 9600 images and the validation set of 1200 images prepared in step S2, supervised learning and optimization training were performed on the constructed SACD-Net model. The experimental environment was based on the PyTorch deep learning framework, with a batch size of 32 and a total of 300 training epochs. AdamW was used as the optimizer, with an initial learning rate of 0.001, and a cosine annealing strategy was used to periodically and dynamically adjust the learning rate.

[0068] A Composite Loss function is employed, with CIoU regression loss and Focal classification loss used for overall tower frame prediction, and CIoU regression loss and Focal classification loss used for rust location prediction, targeting rust level. End-to-end parameter optimization of the entire SACD-Net model is achieved through backpropagation of this composite loss function. During training, the model's performance metrics are evaluated on the validation set after each epoch until the model's performance metrics on the validation set no longer show significant improvement. Finally, the model parameters that perform best on the validation set during training are saved as a weight file, resulting in the trained tower structure-aware rust detection and early warning network model.

[0069] S6 deploys the trained Tower Structure Perception Corrosion Detection and Early Warning Network (SACD-Net) model into the power inspection and analysis system to identify and detect corrosion targets on transmission towers, and finally outputs hierarchically labeled images of towers and corrosion locations, corrosion level categories and confidence scores.

[0070] Example: A corrosion inspection task was carried out on power transmission towers in a certain area, and the corrosion level was assessed to facilitate subsequent prevention and management work. A drone was used to inspect each power transmission tower in the area, and images of the towers were collected as follows. Figure 11 As shown.

[0071] Using this image as input to the tower structure corrosion detection and early warning network, the output detection results are as follows: Figure 12 As shown.

[0072] exist Figure 12 The following information is marked in the text: Corrosion grade: Grade II corrosion Bounding box coordinates: (40, 70, 70, 80) Confidence score: 0.92.

[0073] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for detecting corrosion of power transmission towers based on deep learning, characterized in that, The specific steps are as follows: S1, construct the power transmission tower image dataset D1, and classify and label it according to the corrosion level; divide the dataset into training set, validation set and test set according to the proportion; S2, construct the structure-aware feature extraction backbone network GQ-SAB, input any transmission tower image B1 into the structure-aware feature extraction backbone network, perform feature extraction, and obtain tower structure-aware feature map B5, tower structure-aware feature map B11, tower structure-aware feature map B8, and tower structure-aware feature map B12. S3. Construct a dynamic routing adaptive multi-scale feature fusion network GQ-DRN. Input B5, B11, B8, and B12 into the dynamic routing adaptive multi-scale feature fusion network to perform feature fusion and obtain corrosion feature enhancement fusion map N8, corrosion feature enhancement fusion map N9, and corrosion feature enhancement fusion map N12. S4. Construct the carrier-defect detection bidirectional gated collaborative reasoning head network GQ-COHEAD, input N8, N9, and N12 into the carrier-defect detection bidirectional gated collaborative reasoning head network, perform collaborative reasoning and prediction, and obtain the tower corrosion information detection results. S5. Construct the SACD-Net network model for sensing corrosion detection and early warning of power transmission tower structures; integrate the three parts GQ-SAB, GQ-DRN and GQ-CoHead to construct the complete SACD-Net model proposed in this invention; and input the collected images of power transmission towers into the model for supervised learning and optimization training. S6 deploys the trained tower structure perception corrosion detection and early warning network model into the power inspection and analysis system to identify and detect corrosion targets on transmission towers, and finally outputs hierarchically labeled images of towers and corrosion locations, corrosion level categories and confidence scores.

2. The method for detecting corrosion of transmission towers based on deep learning according to claim 1, characterized in that, The specific steps of S1 are as follows: Multiple types of images are acquired to obtain image data of the tower to be inspected; Based on location selection and corrosion level, the acquired image data is manually annotated in a two-level manner to obtain the transmission tower image dataset D1, which contains multiple corrosion levels.

3. The method for detecting corrosion of transmission towers based on deep learning according to claim 2, characterized in that, The specific steps of S2 are as follows: Input any transmission tower image B1∈D1 into the CBS module for preliminary feature extraction and downsampling to generate tower structure perception feature map B2; In path one, B2 is first input into the C2f module to perform deep feature extraction, resulting in the tower structure perception feature map B3; then, B3 is input into the first SSDE module to extract the geometric features of the tower and effectively suppress background interference, resulting in the tower structure perception feature map B4. In path two, B2 is input into the second SSDE module for processing to obtain the tower structure perception feature map B5; By concatenating B4 and B5 along the channel dimension, we obtain the tower structure perception feature map B6; by performing a Conv 5*5 convolution operation on B6, we obtain the tower structure perception feature map B11. Input B6 into the CBS module to obtain the tower structure perception feature map B7, and then input B7 into the third SSDE module to obtain the tower structure perception feature map B8. First, input B8 into the C2f module for deep feature extraction to obtain the tower structure perception feature map B9; then input B9 into the efficient multi-scale attention module for processing to obtain the tower structure perception feature map B10; finally, input B10 into the alternating multi-scale pooling module for processing to obtain the tower structure perception feature map B12.

4. The method for detecting corrosion of transmission towers based on deep learning according to claim 1, characterized in that, The specific steps of S3 are as follows: In path one, a downsampling operation is performed on B5 to obtain the rust feature enhancement fusion image N1; then, N1 is input into the C2f module for deep feature extraction to obtain the rust feature enhancement fusion image N2. In path two, B12 is upsampled to obtain corrosion feature enhancement fusion image N4; N4 and B8 are spliced ​​along the channel dimension to obtain corrosion feature enhancement fusion image N5; N5 is input into the C2f module for feature extraction to obtain corrosion feature enhancement fusion image N6. In path three, B5 is input into the coordinate attention mechanism module to perform feature enhancement operation, resulting in rust feature enhancement fusion map N3. Conv 3*3 convolution operation is performed on N3 to obtain rust feature enhancement fusion map N10. N2, N6 and B11 are stitched together along the channel dimension to obtain the rust feature enhancement fusion image N7; N7 and N10 are input into the first CGDR module for context-guided dynamic routing fusion to obtain the rust feature enhancement fusion map N8; then, N5 and N8 are input into the second CGDR module to obtain the rust feature enhancement fusion map N9; and N9 is subjected to a Conv 1*1 convolution operation to obtain the rust feature enhancement fusion map N11. In path four, B12 and N11 are input into the third CGDR module for final fusion to obtain the rust feature enhancement fusion map N12.

5. The method for detecting corrosion of transmission towers based on deep learning according to claim 1, characterized in that, The specific steps of S4 are as follows: In branch one, N8 is input into the first BGC-HEAD module to obtain the tower corrosion prediction feature map D1; non-maximum suppression operation is performed on D1 to obtain the tower corrosion prediction feature map D4; In branch two, N9 is input into the second BGC-HEAD module to obtain the tower corrosion prediction feature map D2; non-maximum suppression operation is performed on D2 to obtain the tower corrosion prediction feature map D5; In branch three, N12 is input into the third BGC-HEAD module to obtain the tower corrosion prediction feature map D3; non-maximum suppression operation is performed on D3 to obtain the tower corrosion prediction feature map D6; Finally, D4, D5, and D6 are input into the weighted frame fusion module to obtain the tower corrosion information detection results.

Images