Power transmission tower angle steel corrosion evaluation method and system based on image segmentation

Abstract

The application discloses a power transmission line tower angle steel corrosion evaluation method and system based on image segmentation, and particularly relates to the technical field of tower angle steel corrosion evaluation, and the method comprises the following steps: realizing multi-scale adaptive segmentation of a corrosion area through a double-branch encoder and a cross-scale attention mechanism; utilizing a conditional generative adversarial network to combine a multi-scale perception loss and a curriculum learning strategy to perform sample enhancement and generation; extracting corrosion geometry and distribution features based on a segmentation result, and combining angle steel structure priori, historical data and environmental factors to perform weighted risk scoring and safety cycle prediction. The system comprises four modules of image segmentation, data enhancement, feature extraction and risk evaluation. The application improves the segmentation precision and model generalization capability of small target corrosion, and realizes closed-loop intelligent operation and maintenance support from detection to risk evaluation.

Description

Technical Field

[0001] This invention relates to the field of corrosion assessment technology for angle steel of transmission towers, and more specifically, to a method and system for corrosion assessment of angle steel of transmission line towers based on image segmentation. Background Technology

[0002] Transmission line towers are exposed to complex natural environments for extended periods, and their angle steel structures are susceptible to corrosion from factors such as humidity, salt spray, and industrial pollution, seriously threatening the safe and stable operation of the power grid. Traditional corrosion detection mainly relies on manual inspections and visual judgment, which suffers from low efficiency, strong subjectivity, and significant safety hazards. In recent years, vision-based detection methods have been gradually applied to tower inspections; however, conventional image processing technologies are greatly affected by lighting, angle, and background interference in identifying corroded areas, and it is difficult to achieve accurate segmentation and quantitative assessment of corrosion morphology, failing to meet the needs of intelligent and refined operation and maintenance.

[0003] With the development of deep learning technology, image segmentation methods based on convolutional neural networks have shown potential in defect detection. However, in the scenario of corrosion detection of angle steel in power transmission lines, the following challenges still exist: First, the corrosion areas are diverse in shape and scale, and small-area point corrosion is easily missed; second, real corrosion samples are scarce and annotation costs are high, resulting in insufficient generalization ability of the model; third, existing methods mainly focus on identifying the presence of corrosion, lacking a comprehensive assessment system for corrosion location, density, development trend and its correlation with structural safety, making it difficult to directly support maintenance decisions.

[0004] To address the aforementioned issues, existing research has attempted data augmentation using multi-scale networks, attention mechanisms, or generative adversarial networks. However, limitations remain, such as insufficient feature fusion, inadequate realism of generated samples, and a single evaluation dimension. Therefore, there is an urgent need for an integrated method and system capable of achieving high-precision corrosion segmentation, possessing adaptive sample augmentation capabilities, and combining structural priors with environmental factors for multi-dimensional risk assessment. This would improve the automation and intelligence of transmission line tower corrosion detection and provide a reliable basis for operation and maintenance decisions. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for assessing corrosion of angle steel in transmission line towers based on image segmentation.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A corrosion assessment method for angle steel of transmission line towers based on image segmentation includes the following steps:

[0008] B1. Multi-scale adaptive segmentation: A dual-branch encoder is constructed to extract the overall contour and high-frequency detail features of the angle steel image. Through cross-scale feature interaction and adaptive weight loss function, pixel-level segmentation of the eroded area is achieved.

[0009] B2. Sample Augmentation and Generation: Based on conditional generative adversarial networks, virtual images with diverse erosion morphologies are generated, and multi-scale perceptual loss and curriculum learning strategies are combined to improve the realism and complexity of the generated samples.

[0010] B3. Corrosion Quantification and Risk Assessment: Extract the geometric and distribution characteristics of the corrosion area from the segmentation results, combine the prior knowledge of angle steel structure to perform weighted risk scoring, and calculate the corrosion expansion trend and remaining safe period based on historical data and environmental factors, and output the risk assessment results.

[0011] Specifically, in the dual-branch encoder of B1

[0012] The second branch connects a deformable convolutional layer after the last bottleneck layer of the MobileNetV3 network. The offset of the deformable convolutional layer is generated by a 3×3 convolutional layer based on the current feature map.

[0013] During training, the weights of the offset layer are initialized to zero, and the learning rate is set to 0.1 times the base learning rate.

[0014] Specifically, the cross-scale feature interaction process includes:

[0015] The deep feature map from the first branch is upsampled by 2 times, and after matching its spatial size with the shallow feature map from the second branch, they are stitched together in the channel dimension.

[0016] The concatenated feature map is input into a spatial attention submodule. Two single-channel feature maps are obtained by global average pooling and global max pooling along the channel dimension. After concatenation, a spatial weight map is generated by a 7×7 convolutional layer, and the weight values ​​are normalized using the Sigmoid function.

[0017] The normalized weight map is multiplied element-wise with the shallow feature map, and then added to the deep feature map to obtain the output.

[0018] Specifically, in B2:

[0019] The generator of the conditional generative adversarial network takes a normal angle steel image and a randomly generated erosion region mask as conditional inputs, and adopts a U-Net structure;

[0020] The discriminator uses a PatchGAN structure to distinguish between real and fake local regions of an image.

[0021] Specifically, the multi-scale perceptual loss is calculated using a pre-trained VGG-19 network as a fixed feature extractor;

[0022] Feature maps of three selected feature layers—middle, upper-middle, and deep—are extracted respectively. The mean square error of the generated image and the real erosion image on each feature map is calculated, and the results are weighted and summed according to preset weights to obtain the multi-scale perceptual loss value.

[0023] Specifically, the course learning strategy includes three stages:

[0024] In the first stage, the ratio of generated images to real images is 1:3. The discriminator distinguishes between 64×64 pixel blocks. When the discrimination accuracy is below 60% for 10 consecutive training cycles, it enters the second stage.

[0025] In the second stage, the mixing ratio is adjusted to 1:1, the discriminator input block size is adjusted to 128×128 pixels, and the third stage is entered when the discrimination accuracy is below 55% for 10 consecutive training cycles.

[0026] In the third stage, the mixing ratio is adjusted to 3:1, the discriminator judges the scale of the entire image, and a historical image data fine-tuning generator is introduced.

[0027] Specifically, in the weighted risk score:

[0028] The risk factor of the location of the corrosion zone is predetermined based on the location of the angle steel component;

[0029] The environmental severity factor is assigned a value based on the environmental type of the inspection area;

[0030] The criteria for determining the remaining safety period are as follows:

[0031] If the calculated result is greater than 10 years, it is considered low risk; if it is between 5 and 10 years, it is considered medium risk; and if it is less than or equal to 5 years, it is considered high risk.

[0032] The image segmentation-based corrosion assessment system for angle steel of transmission line towers includes the following modules:

[0033] The image segmentation module is used to perform the multi-scale adaptive segmentation step and output the segmentation mask of the eroded region;

[0034] The data augmentation module is used to perform the sample augmentation and generation steps, and generate and output the augmented training samples;

[0035] The corrosion feature extraction and analysis module is used to perform the corrosion quantification step and extract geometric and distribution features from the segmentation results.

[0036] The risk assessment and prediction module is used to perform the risk assessment steps and output a quantitative risk score, expansion trend, and maintenance recommendations.

[0037] The technical effects and advantages of this invention are as follows:

[0038] At the technical implementation level, the accuracy and robustness of corrosion detection are significantly improved by constructing a multi-scale adaptive segmentation network and an intelligent sample augmentation mechanism. The dual-branch encoder, combined with deformable convolution and attention mechanisms, effectively integrates overall contour and detailed features, enhancing the segmentation ability for small targets and irregular corrosion regions. At the same time, data augmentation is performed using conditional generative adversarial networks and multi-scale perceptual loss, which alleviates the problem of scarce training data while ensuring the authenticity and diversity of generated samples, enabling the model to adapt to complex and ever-changing actual corrosion morphologies.

[0039] At the engineering application level, a closed-loop decision support system from corrosion detection to risk assessment has been achieved. By introducing prior knowledge of angle steel structures, historical data, and environmental factors, a multi-dimensional quantitative assessment model covering corrosion degree, location risk, and development trend has been constructed. This model can accurately calculate the corrosion propagation rate and remaining safety period, and output a visualized risk map and graded maintenance recommendations. This method effectively overcomes the limitations of traditional manual inspections, which are characterized by strong subjectivity and low efficiency, and provides a reliable technical means for preventive maintenance and intelligent operation and maintenance of transmission line towers. Attached Figure Description

[0040] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] like Figure 1 As shown, the steps of the image segmentation-based corrosion assessment method for angle steel of transmission line towers are as follows:

[0043] Step 1: Multi-scale adaptive segmentation, specifically including the following steps:

[0044] A dual-branch encoder is constructed. Branch 1 uses ResNet-50 to extract the overall contour and texture features of the angle steel, while branch 2 uses lightweight MobileNetV3 to extract high-frequency detail features. Deformable convolution is used to adaptively focus on the eroded edge region.

[0045] The design incorporates a cross-scale feature interaction module, introducing a spatial attention mechanism during the decoding stage to weightedly fuse deep semantic features with shallow detail features, thereby enhancing the salience of small eroded regions.

[0046] The deformable convolution is implemented as follows: after the last bottleneck layer in the MobileNetV3 branch, a deformable convolution v2 layer is added. The offset of this layer is generated by a 3×3 ordinary convolutional layer, whose input is the current feature map and whose output has 18 channels (corresponding to the x and y offsets of the 3×3 convolutional kernel). During training, the weights of this offset layer are initialized to zero, and the learning rate is set to 0.1 times the base learning rate to ensure the stability of the network in the initial stage of training.

[0047] The specific operation steps of the cross-scale feature interaction module are as follows:

[0048] The deep feature maps from the ResNet-50 branch (denoted as...) ) is upsampled by 2 times, so that its spatial size is the same as that of the shallow feature map from the MobileNetV3 branch (denoted as ). )match.

[0049] Will and Perform splicing along the channel dimension;

[0050] The concatenated feature maps are input into a spatial attention submodule: First, global average pooling and global max pooling are performed along the channel dimension to obtain two single-channel feature maps; second, the two feature maps are concatenated and fed into a 7×7 convolutional layer to generate a spatial weight map W; finally, the weight values ​​of W are normalized to the [0,1] interval using the Sigmoid function.

[0051] The normalized spatial weight map W is compared with the original before stitching. The feature maps are multiplied element by element to achieve feature weighting and highlight the detailed information of small erosion areas.

[0052] The weighted shallow features are added to the deep features, and this sum is used as the output of this decoding stage.

[0053] The output layer employs a multi-supervised loss function, combining Dice Loss and Focal Loss, to optimize the pixel-level classification sensitivity for eroded regions of small targets.

[0054] Simultaneously, an adaptive weight adjustment mechanism based on the proportion of the eroded region is introduced, which enhances the model's focus on positive samples when the eroded region is extremely small, and balances the overall segmentation quality when the eroded region is large; the specific process is as follows:

[0055] set up: The proportion of pixels in the eroded region of the j-th image in the training batch, i.e., the proportion of positive samples;

[0056] and The preset corrosion percentage thresholds are set to 0.01 and 0.20, respectively.

[0057] and These are the dynamic weights of Focal Loss and Dice Loss for the j-th image, respectively;

[0058] Dynamic weight calculation process:

[0059] First, calculate the erosion ratio response factor for each image. : (when >0); where To prevent numerical instability, the minimum value is used;

[0060] based on Calculate the normalized weights:

[0061]

[0062]

[0063] in: =0.5 is the baseline weighting coefficient; =1.2 is the strengthening coefficient; =2.0 is the decay exponent; for Images (uncorrosive), using fixed weights: ;

[0064] Final loss function:

[0065]

[0066] in The total number of images in a single training batch. The Focal Loss value is calculated based on the j-th image. The Dice Loss value is calculated based on the j-th image. The value of Lovasz-SoftmaxLoss is used to directly optimize the intersection-over-union (IoU) ratio of the segmentation results. The total loss function value for training the segmentation network. =0.2 is Fixed weights.

[0067] Step 2, Sample Enhancement and Generation, includes the following steps:

[0068] A conditional generative adversarial network is constructed. The generator takes a normal angle steel image and a random corrosion mask as input and generates an image with virtual corrosion texture through a U-Net structure. The discriminator adopts a PatchGAN structure to determine the authenticity of local images.

[0069] The generator is trained using real corrosion images to generate multi-morphic corrosion patterns (such as dot-like, sheet-like, and stripe-like patterns), and the diversity is enhanced by changes in lighting and perspective.

[0070] Multi-scale perceptual loss is introduced into the generator's optimization objective; a pre-trained deep convolutional neural network is used as a feature perceptron, forcing the eroded regions generated by the generator to be similar to real data not only at the pixel level and in adversarial discrimination, but also at the high-level semantic feature level; the process is as follows:

[0071] Multi-scale feature extractor construction: Select a convolutional neural network (e.g., VGG-19) pre-trained on a large image dataset (e.g., ImageNet) as a fixed feature extractor, denoted as . Three feature layers with different receptive fields and semantic depths are selected to form a feature layer set. , These correspond to the middle, upper-middle, and deep layers of the network, respectively.

[0072] Multi-scale perceptual loss calculation: Let the image output by the generator be... The corresponding real corrosion image is For the selected set Each feature layer in Extract respectively and The feature map of this layer is denoted as and The perceptual loss of this layer is the mean square error between the two feature maps.

[0073]

[0074] in, , , These are the height, width, and number of channels of the feature map, respectively. Denotes the Frobenius norm;

[0075] Multiscale sensing loss The weighted sum of the losses at each layer:

[0076]

[0077] in, For the first The layer weight coefficients are set to , , To place greater emphasis on the consistency of deep semantic features;

[0078] Extension of the generator's total loss function: The loss of the original generator typically includes adversarial loss. and possible L1 reconstruction loss After introducing multi-scale perceptual loss, the overall optimization objective of the generator is... Updated to:

[0079]

[0080] in, and To balance the weighting hyperparameters of various losses; based on experiments, the following settings were configured. = 10 (emphasizing pixel-level structural alignment). Setting the feature extractor to 1 (introducing appropriate feature-level constraints) yields good results. During training, the feature extractor... Parameters remain frozen, only through The gradient is used to update the generator parameters;

[0081] Integration with curriculum learning strategies: During each training phase, the generator must minimize the multi-scale perceptual loss between its output image and the real erosion image while trying to deceive the discriminator of the corresponding phase. This ensures that even in the early stages of curriculum learning (such as generating dotted erosion), the generated texture has realistic erosion visual features, thus providing higher-quality augmented samples for the segmentation model.

[0082] By mixing generated images with real images in proportion and employing a course-based learning strategy, the difficulty of generated samples is gradually increased to improve the model's adaptability to complex erosion patterns.

[0083] The specific implementation of the course learning strategy is divided into three stages:

[0084] Phase 1 (Basic Morphological Learning): The ratio of generated images to real images is 1:3. The generator's training objective focuses on learning to generate isolated dot-like erosion patterns. The discriminator only distinguishes between randomly cropped 64×64 pixel blocks in the image. When the discriminator's accuracy in classifying generated samples in the mixed dataset remains consistently below 60% for 10 consecutive epochs, the process proceeds to the next phase.

[0085] Phase 2 (Intermediate Morphology Learning): The blending ratio is adjusted to 1:1. The generator begins learning to generate sheet-like and small connected eroded regions. The discriminator's input block size is adjusted to 128×128 pixels to consider greater contextual information. The advancement condition is that the discriminator's accuracy is below 55% for 10 consecutive epochs.

[0086] Phase 3 (Complex Morphology Learning): The blending ratio is adjusted to 3:1 to increase the diversity of generated data. The generator learns to generate complex erosions such as stripes, meshes, and combinations thereof. The discriminator input block is restored to the full image scale (256×256 pixels). The generator is fine-tuned using historical image data to simulate the expansion of erosion over time. Training continues until the generative adversarial loss (Wasserstein distance) converges and its fluctuation is less than 5%.

[0087] Step 3, Corrosion Quantification and Risk Assessment, includes the following steps:

[0088] Geometric features of the eroded region are extracted based on the segmentation results, including area ratio, length of the maximum connected region, and spatial density of eroded points.

[0089] By introducing prior knowledge about angle steel structures (such as critical stress areas and welded areas), a weighted risk score is applied to the corrosion area, and a higher risk coefficient is assigned to the corrosion of critical parts.

[0090] A corrosion development trend index is constructed, the corrosion expansion rate is calculated by comparing historical images, and the remaining safe period is predicted by combining environmental data (such as humidity and salinity), and a visualized risk assessment map is output.

[0091] The weighted risk score is calculated using the following formula:

[0092]

[0093] in:

[0094] This represents the total risk value of the current angle steel component.

[0095] n represents the total number of independent corrosion areas identified in the image of the angle steel component;

[0096] For the first The pixel area of ​​each eroded region;

[0097] The risk factor for the location of the corrosion zone is determined by referring to a table based on the component location (e.g., k=1.8 at the connection node between the main member and the diagonal member, k=1.0 in the ordinary web member area, and k=2.0 within 5mm around the bolt hole).

[0098] This represents the spatial density of corrosion points, i.e., the number of corrosion cluster centers per unit area.

[0099] This is the density weighting coefficient, with an empirical value set to 0.3.

[0100] Corrosion spread rate and remaining safety period Calculated using the following model:

[0101] Expansion rate : ,in The standardized corrosion area percentages are for time points t (this inspection) and t−1 (the previous inspection), respectively. The time interval between two inspections (in years). The environmental severity factor is 1.5 for humid industrial areas, 1.0 for ordinary outdoor areas, and 0.7 for arid areas.

[0102] Remaining security period : ,in This is the maximum allowable corrosion area threshold according to the design specifications for this type of angle steel (usually 15% of the component's cross-sectional area loss rate). This represents the current (at the time of this inspection) percentage of the actual corroded area (already standardized). ≤0 or When the lifespan is >10 years, the output risk is low, and routine inspections are recommended; when 5 < When the timeframe is ≤10 years, the output risk is medium, and closer monitoring is recommended; when... If the lifespan is ≤5 years, the output is high-risk, and planned maintenance is recommended.

[0103] The image segmentation-based corrosion assessment system for angle steel of transmission line towers includes the following modules:

[0104] The image segmentation module employs a multi-scale adaptive fusion segmentation network. It extracts the overall contour and high-frequency detail features of the angle steel using a dual-branch encoder (ResNet-50 and MobileNetV3), and combines deformable convolution and cross-scale attention mechanisms to achieve accurate segmentation of the eroded region. The module outputs a pixel-level erosion mask, supporting highly sensitive recognition of small-area, irregular erosion morphologies.

[0105] The data augmentation module generates diverse virtual erosion images based on a conditional generative adversarial network (cGAN). By introducing multi-scale perceptual loss and a curriculum learning strategy, it gradually improves the realism and complexity of the generated samples. The augmented images are then mixed with real data, effectively enhancing the segmentation model's generalization ability and robustness to various erosion morphologies.

[0106] The corrosion feature extraction and analysis module extracts the geometric and distribution features of the corroded areas from the segmentation results, including quantitative indicators such as corrosion area ratio, maximum connected region length, and spatial density. Simultaneously, it incorporates prior knowledge of the angle steel structure (such as key load-bearing components) to perform positional weighting on the corroded areas, providing structured data input for subsequent risk assessment.

[0107] The risk assessment and prediction module calculates the corrosion propagation rate and remaining safe period based on corrosion characteristics and historical data. It also combines environmental factors (humidity, salinity, etc.) to quantify the risk score and output a visualized risk assessment map.

[0108] The above formulas are all dimensionless calculations. Dimensionless calculations can be performed using various methods such as standardization, which will not be elaborated here. The formulas are derived from software simulations based on a large amount of collected data, and the preset parameters in the formulas can be set by those skilled in the art according to the actual situation.

[0109] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0110] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0111] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0112] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0113] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0114] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0115] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0116] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for assessing corrosion of angle steel in transmission line towers based on image segmentation, characterized in that, Includes the following steps: B1. Multi-scale adaptive segmentation: A dual-branch encoder is constructed to extract the overall contour and high-frequency detail features of the angle steel image. Through cross-scale feature interaction and adaptive weight loss function, pixel-level segmentation of the eroded area is achieved. B2. Sample Augmentation and Generation: Based on conditional generative adversarial networks, virtual images with diverse erosion morphologies are generated, and multi-scale perceptual loss and curriculum learning strategies are combined to improve the realism and complexity of the generated samples. B3. Corrosion Quantification and Risk Assessment: Extract the geometric and distribution characteristics of the corrosion area from the segmentation results, combine the prior knowledge of angle steel structure to perform weighted risk scoring, and calculate the corrosion expansion trend and remaining safe period based on historical data and environmental factors, and output the risk assessment results.

2. The method for assessing corrosion of angle steel in transmission line towers based on image segmentation according to claim 1, characterized in that, In the dual-branch encoder of B1: The second branch connects a deformable convolutional layer after the last bottleneck layer of the MobileNetV3 network. The offset of the deformable convolutional layer is generated by a 3×3 convolutional layer based on the current feature map. During training, the weights of the offset layer are initialized to zero, and the learning rate is set to 0.1 times the base learning rate.

3. The method for assessing corrosion of angle steel in transmission line towers based on image segmentation according to claim 2, characterized in that, The cross-scale feature interaction process includes: The deep feature map from the first branch is upsampled by 2 times, and after matching its spatial size with the shallow feature map from the second branch, they are stitched together in the channel dimension. The concatenated feature map is input into a spatial attention submodule. Two single-channel feature maps are obtained by global average pooling and global max pooling along the channel dimension. After concatenation, a spatial weight map is generated by a 7×7 convolutional layer, and the weight values ​​are normalized using the Sigmoid function. The normalized weight map is multiplied element-wise with the shallow feature map, and then added to the deep feature map to obtain the output.

4. The method for assessing corrosion of angle steel in transmission line towers based on image segmentation according to claim 1, characterized in that, In B2: The generator of the conditional generative adversarial network takes a normal angle steel image and a randomly generated erosion region mask as conditional inputs, and adopts a U-Net structure; The discriminator uses a PatchGAN structure to distinguish between real and fake local regions of an image.

5. The method for assessing corrosion of angle steel in transmission line towers based on image segmentation according to claim 4, characterized in that, The multi-scale perceptual loss is calculated using a pre-trained VGG-19 network as a fixed feature extractor. Feature maps of three selected feature layers—middle, upper-middle, and deep—are extracted respectively. The mean square error of the generated image and the real erosion image on each feature map is calculated, and the results are weighted and summed according to preset weights to obtain the multi-scale perceptual loss value.

6. The method for assessing corrosion of angle steel in transmission line towers based on image segmentation according to claim 5, characterized in that, The course learning strategy specifically includes three stages: In the first stage, the ratio of generated images to real images is 1:

3. The discriminator distinguishes between 64×64 pixel blocks. When the discrimination accuracy is below 60% for 10 consecutive training cycles, it enters the second stage. In the second stage, the mixing ratio is adjusted to 1:1, the discriminator input block size is adjusted to 128×128 pixels, and the third stage is entered when the discrimination accuracy is below 55% for 10 consecutive training cycles. In the third stage, the mixing ratio is adjusted to 3:1, the discriminator judges the scale of the entire image, and a historical image data fine-tuning generator is introduced.

7. The method for assessing corrosion of angle steel in transmission line towers based on image segmentation according to claim 1, characterized in that, In the weighted risk score: The risk factor of the location of the corrosion zone is predetermined based on the location of the angle steel component; The environmental severity factor is assigned a value based on the environmental type of the inspection area; The criteria for determining the remaining safety period are as follows: If the calculated result is greater than 10 years, it is considered low risk; if it is between 5 and 10 years, it is considered medium risk; and if it is less than or equal to 5 years, it is considered high risk.

8. A system applied to the image segmentation-based corrosion assessment method for angle steel of transmission line towers according to any one of claims 1-7, characterized in that, Includes the following modules: The image segmentation module is used to perform the multi-scale adaptive segmentation step and output the segmentation mask of the eroded region; The data augmentation module is used to perform the sample augmentation and generation steps, and generate and output the augmented training samples; The corrosion feature extraction and analysis module is used to perform the corrosion quantification step and extract geometric and distribution features from the segmentation results. The risk assessment and prediction module is used to perform the risk assessment steps and output a quantitative risk score, expansion trend, and maintenance recommendations.

Images