High-precision lightweight stay cable apparent disease segmentation method and system based on deep learning

Abstract

The invention discloses a high-precision lightweight stay cable apparent disease segmentation method and system based on deep learning. Relates to the technical field of cable-stayed bridge health detection and the technical field of computer vision. 2, carrying out image preprocessing and image enhancement on the apparent disease image of the stay cable to obtain an apparent disease data set of the stay cable; 3, constructing an apparent disease segmentation model of the stay cable; and 4, inputting the stay cable apparent disease data set into the stay cable apparent disease segmentation model, and outputting a disease segmentation result. According to the method, the apparent disease of the stay cable can be segmented more quickly and accurately.

Description

Technical Field

[0001] The present invention relates to the fields of cable-stayed bridge health detection technology and computer vision technology, and more specifically to a high-precision lightweight cable-stayed cable apparent disease segmentation method and system based on deep learning. Background Art

[0002] Among long-span bridges, cable-stayed bridges are widely used in practical engineering due to their large span capacity, excellent wind and seismic stability, and aesthetically pleasing appearance. As the core component of the coordinated force-bearing system of the main girder and tower of a cable-stayed bridge, the degradation of its mechanical properties and the evolution of damage directly affect the structural safety of the bridge throughout its lifecycle. During long-term service, the coupling of multiple factors such as material fatigue, environmental corrosion, and extreme loads can easily induce apparent cracks and scratches. If not promptly inspected and maintained, the reduction in the effective cross-sectional area of ​​the cable and the resulting stress redistribution will lead to a significant decrease in load-bearing capacity, posing potential safety risks.

[0003] Currently, visual information-based methods for detecting apparent defects in cable-stayed cables mainly include manual detection methods, machine learning methods based on traditional image processing, and automatic detection methods based on deep learning. Manual detection methods mainly rely on manual visual identification and recording of apparent defects in cable-stayed cables. While flexible and intuitive, they are highly subjective, inefficient, and risky, making them difficult to meet the needs of rapid detection of apparent defects in large-scale cable-stayed cables. Machine learning methods based on traditional image processing mainly rely on manual feature extraction algorithms such as edge detection, grayscale analysis, and texture analysis, combined with classifiers such as support vector machines, K-nearest neighbors, and random forests to identify apparent defects in cable-stayed cables. While these methods are user-friendly for small-scale samples and highly interpretable, their feature extraction relies on manual experience, has poor adaptability, and is difficult to cope with complex backgrounds and lighting changes. Automatic detection methods based on deep learning use neural networks to automatically identify apparent defects in cable-stayed cables. These methods offer advantages such as end-to-end training, automatic feature extraction, strong robustness, and the ability to adapt to complex environments. However, they rely heavily on high-quality labeled data, have high model complexity, limited real-time performance, and require optimization for model lightweightness and deployability.

[0004] Automatic detection methods based on deep learning mainly include object detection networks and semantic segmentation networks. Existing detection methods still have room for improvement in detection accuracy, computational efficiency, and real-time performance.

[0005] Therefore, how to design a high-precision lightweight cable-stayed cable apparent defect segmentation method and system based on deep learning to detect cable-stayed cable apparent defects more quickly and accurately is a problem that technical personnel in this field urgently need to solve. Summary of the Invention

[0006] In view of this, the present invention provides a high-precision lightweight cable-stayed cable apparent defect segmentation method and system based on deep learning, which can segment the cable-stayed cable apparent defects more quickly and accurately.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a high-precision method for segmenting apparent defects of lightweight stay cables based on deep learning, comprising:

[0008] Step 1: Obtain the apparent damage image of the cable;

[0009] Step 2: Perform image preprocessing and image enhancement methods on the cable-stayed apparent damage images to obtain the cable-stayed apparent damage dataset;

[0010] Step 3: Construct the apparent damage segmentation model of the cable;

[0011] Step 4: Input the cable-stayed cable apparent disease dataset into the cable-stayed cable apparent disease segmentation model and output the disease segmentation results.

[0012] Preferably, in step 2, the image is preprocessed by using histogram equalization and Gaussian filtering, and then the image enhancement method of rotation, scaling and cropping is used to expand the number of defect images to obtain a cable-stayed cable apparent defect dataset.

[0013] Preferably, the cable-stayed cable apparent defect segmentation model includes a high-precision semantic segmentation network and a lightweight semantic segmentation network, and the high-precision semantic segmentation network and the lightweight semantic segmentation network work together through knowledge distillation and a generative adversarial mechanism.

[0014] Preferably, the high-precision semantic segmentation network includes a baseline semantic segmentation network, an SFF semantic feature fusion extraction network, and a SEA squeeze enhanced axial attention network; the corresponding SFF semantic feature fusion extraction network is fused between the encoder and decoder of the baseline semantic segmentation network, and the SEA squeeze enhanced axial attention network is introduced into the decoder of the baseline semantic segmentation network;

[0015] The SFF semantic feature fusion extraction network is used to fuse the feature maps of the cable-stayed cable apparent disease dataset at different scales, and the SEA squeeze-enhanced axial attention network is used to enhance the local detail perception capability.

[0016] The SFF semantic feature fusion extraction network first fuses the feature maps of different scales output by the encoder, and the fused feature maps are passed as input to the SEA squeeze enhanced axial attention network.

[0017] Preferably, the SEA squeezed enhanced axial attention network includes SAA squeezed axial attention and DEK detail enhancement kernel, which captures global semantics through SAA axial attention and restores local texture in combination with DEK detail enhancement kernel;

[0018] The SAA squeeze axial attention is used to squeeze the input features horizontally and vertically respectively, broadcasting the global semantics through multi-head attention, and the output feature map size remains unchanged;

[0019] The DEK detail enhancement kernel concatenates the original features with the SAA output, extracts local details through deep convolution, generates enhancement weights and multiplies them with the SAA features to restore edge local information.

[0020] Preferably, the high-precision semantic segmentation network serves as the teacher network, and transfers feature expression capabilities to the lightweight semantic segmentation network through knowledge distillation. The lightweight semantic segmentation network serves as the student network, and learns the feature expression of the teacher network through knowledge distillation.

[0021] Preferably, the lightweight semantic segmentation network is trained through knowledge distillation to learn the high-precision semantic segmentation network feature extraction capability, and the adversarial training of the lightweight semantic segmentation network is achieved through generative adversarial methods;

[0022] Knowledge distillation training uses the KL divergence loss function to enable the student network to learn the probability distribution of the teacher network;

[0023] Adversarial training introduces a discriminator to distinguish the outputs of the teacher network and the student network, and optimizes the student network through adversarial training.

[0024] Preferably, a high-precision lightweight stay-cable apparent defect segmentation system based on deep learning comprises:

[0025] Image acquisition module: obtains the apparent damage image of the inclined cable;

[0026] Preprocessing module: Perform image preprocessing and image enhancement methods on the apparent damage images of the inclined cables to obtain the apparent damage dataset of the inclined cables;

[0027] Model building module: build the apparent damage segmentation model of the cable;

[0028] Output module: Input the cable apparent disease dataset into the cable apparent disease segmentation model and output the disease segmentation results.

[0029] It can be seen from the above technical solutions that, compared with the prior art, the present invention provides a high-precision method and system for segmenting apparent defects of lightweight stay cables based on deep learning, which has the following beneficial technical effects:

[0030] 1. The present invention uses a training set of high-quality cable-stayed cable apparent disease images to train a cable-stayed cable apparent disease semantic segmentation model. After the training is completed, the cable-stayed cable apparent disease images of the validation set are segmented, which can achieve high-precision segmentation of cable-stayed cable apparent diseases.

[0031] 2. The semantic segmentation model of apparent defects of the inclined cable used in the method of the present invention integrates the proposed SFF semantic feature fusion module, which enhances the feature expression ability and improves the accuracy of defect boundary segmentation; introduces the SEA extrusion enhanced axial attention mechanism, strengthens the ability to perceive local details, improves the overall detection accuracy, and realizes the refined segmentation of apparent defects of the inclined cable.

[0032] 3. The lightweight semantic segmentation model for apparent defects of cable-stayed cables used in the method of the present invention proposes a joint training strategy of knowledge distillation and generative adversarial learning, which guides the lightweight student network to learn the feature expression ability of the teacher network, and achieves high-precision and high-efficiency segmentation of apparent defects of cable-stayed cables.

[0033] 4. The high-precision lightweight cable-stayed cable apparent defect segmentation method based on deep learning in the present invention can achieve faster and more accurate cable-stayed cable apparent defect segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0035] Figure 1 Flowchart provided for the present invention;

[0036] Figure 2 A schematic diagram of the process of constructing a cable-stayed cable apparent disease dataset provided by the present invention;

[0037] Figure 3 Schematic diagram of the high-precision lightweight stay-cable apparent defect segmentation model architecture provided by the present invention;

[0038] Figure 4 Schematic diagram of the semantic feature fusion extraction network structure provided by the present invention;

[0039] Figure 5 Schematic diagram of the high-precision semantic segmentation network structure provided by the present invention. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0041] like Figure 1 As shown, the embodiment of the present invention discloses a high-precision method for segmenting apparent defects of lightweight stay cables based on deep learning, comprising:

[0042] Step 1: Obtain the apparent damage image of the cable;

[0043] Step 2: Perform image preprocessing and image enhancement methods on the cable-stayed apparent damage images to obtain the cable-stayed apparent damage dataset;

[0044] Step 3: Construct the apparent damage segmentation model of the cable;

[0045] Step 4: Input the cable-stayed cable apparent disease dataset into the cable-stayed cable apparent disease segmentation model and output the disease segmentation results.

[0046] Specifically, in step 2, the image is preprocessed using histogram equalization and Gaussian filtering, and then the image enhancement method of rotation, scaling and cropping is used to expand the number of defect images to obtain the apparent defect data set of the inclined cable.

[0047] In another specific embodiment, Figure 2 As shown in the figure, the images of apparent cracks and scratches on the inclined cables are collected by portable mobile devices. To address the common problems of uneven illumination and noise interference in field-collected images, histogram equalization and Gaussian filtering are used for preprocessing to improve the clarity and feature recognizability of the defect images. In addition, image enhancement methods such as rotation, scaling, and cropping are used to expand the number of defect images.

[0048] To address the common problems of uneven illumination and noise interference in on-site image collection, image preprocessing includes histogram equalization and image filtering. The images of apparent cracks and scratches on the inclined cables collected by portable mobile devices have problems such as uneven overall brightness distribution, low image contrast, and unclear image details, which seriously affect the model's effective learning of disease characteristics. Therefore, histogram equalization is used to process the apparent disease images of inclined cables with uneven illumination, which can effectively improve the problem of uneven image brightness distribution, enhance image contrast and detail clarity, and enhance the recognizability of disease characteristics; image filtering is used to process the noisy areas in the apparent disease images of inclined cables, effectively suppressing high-frequency interference and random noise, and enhancing image smoothness and stability.

[0049] To address the problem of insufficient number of defect samples, after the above preprocessing, some apparent defect images of inclined cables were manually screened out, and image enhancement techniques such as rotation, scaling, and cropping were performed to enrich the defect morphology and expand the number of defect images. The defect areas were then finely manually annotated. Cracks and scratches are common defects that appear on the surface of inclined cables and are very harmful to them. Based on the dark black characteristics of the crack and scratch defect areas, the crack and scratch areas were finely manually annotated to construct a high-quality image dataset of apparent cracks and scratches on inclined cables, which can provide stable data support for subsequent model training and performance verification.

[0050] Specifically, the cable-stayed cable apparent defect segmentation model includes a high-precision semantic segmentation network and a lightweight semantic segmentation network, which work together through knowledge distillation and a generative adversarial mechanism.

[0051] In another specific embodiment, the high-precision lightweight cable apparent disease segmentation method used is as follows: Figure 3 As shown, it includes constructing a high-quality inclined-cable apparent disease dataset based on image preprocessing and image enhancement methods, a high-precision inclined-cable apparent disease segmentation network based on multi-scale semantic fusion, and a lightweight inclined-cable apparent disease segmentation network based on collaborative compression; the high-precision lightweight inclined-cable apparent disease segmentation model inputs the inclined-cable apparent disease verification set into the encoder of the pre-trained high-precision semantic segmentation network to extract feature maps of different scales, and fuses them through the semantic feature extraction network, and finally inputs them into the decoder to achieve accurate segmentation of inclined-cable apparent diseases; the high-precision semantic segmentation network is used as the teacher network, and the lightweight semantic segmentation network is used as the student network to learn the feature expression ability of the teacher network, so as to achieve high-precision and lightweight segmentation of inclined-cable apparent diseases.

[0052] Specifically, the high-precision semantic segmentation network includes a baseline semantic segmentation network, an SFF semantic feature fusion extraction network, and a SEA squeeze-enhanced axial attention network. The corresponding SFF semantic feature fusion extraction network is fused between the encoder and decoder of the baseline semantic segmentation network, and the SEA squeeze-enhanced axial attention network is introduced into the decoder of the baseline semantic segmentation network.

[0053] The SFF semantic feature fusion extraction network is used to fuse the feature maps of the cable-stayed cable apparent disease dataset at different scales, and the SEA squeeze-enhanced axial attention network is used to enhance the local detail perception capability.

[0054] The SFF semantic feature fusion extraction network first fuses the feature maps of different scales output by the encoder, and the fused feature maps are passed as input to the SEA squeeze enhanced axial attention network.

[0055] Further analysis of the structural features of a baseline semantic segmentation network revealed that its encoder outputs both low-level and high-level feature information. Low-level features primarily focus on local image details and edge information, capturing basic image information but lacking strong semantic information. High-level features focus on the global semantics of the entire image, capturing complex structures, object categories, scene descriptions, and other information within the image, but also lacking local details. However, fusing low-level and high-level features with significantly different semantic information during decoding—the fusion of different representations and semantic levels—can easily lead to information loss or redundancy, making it difficult for the decoder to effectively extract useful information, thus affecting segmentation accuracy. Therefore, a SFF semantic feature fusion module network was designed to fuse feature maps at different levels before inputting them into the decoder, improving the segmentation accuracy of the baseline semantic segmentation network for apparent cable defects.

[0056] In another specific embodiment, Figure 4-5 As shown in the figure, the input low-resolution feature map is first subjected to feature extraction through a convolution with a kernel size of 3×3, a stride of 1, and padding of 1. The spatial resolution of the feature map is then doubled through bilinear interpolation to align it with the spatial resolution of the high-resolution feature map. A 1×1 convolution is then performed on the feature map to align its channels with those of the high-resolution feature map. The feature map is then concatenated with the input high-resolution feature map along the channel dimension. Finally, feature extraction is performed through a convolution with a kernel size of 7×7, a stride of 1, and padding of 3, or a convolution with a kernel size of 3×3, a stride of 1, and padding of 1. The encoder output uses a 7×7 convolution for low-level features to capture large-scale features, while the output uses a 3×3 convolution for high-level features to avoid information loss. Integrating the corresponding SFF module between the encoder and decoder of the baseline model effectively combines low-level details with high-level semantic information, further enhancing the model's feature representation and ensuring clearer identified disease boundaries.

[0057] The SFF semantic feature fusion module network first performs 3×3 convolution on low-resolution features, followed by 2x upsampling and channel alignment. It then concatenates these features with high-resolution features and finally performs 7×7 or 3×3 convolution. The low-level features output by the encoder use 7×7 convolutions, denoted as SFF7, to capture large-scale features. The high-level features output use 3×3 convolutions, denoted as SFF3, to avoid information loss. SFF7 is used for semantic fusion between the 64×H / 4×W / 4 feature map and the 128×H / 8×W / 8 feature map output by the baseline model encoder. SFF7 is used for semantic fusion between the 128×H / 8×W / 8 feature map and the 320×H / 16×W / 16 feature map output by the baseline model encoder. SFF3 is used for semantic fusion between the 320×H / 16×W / 16 feature map and the 512×H / 32×W / 32 feature map output by the baseline model encoder. These three semantically fused feature maps and the 512 feature map are input into the decoder, which enhances the model feature expression capability and improves the accuracy of apparent cable defect segmentation.

[0058] Specifically, the SEA squeezed enhanced axial attention network includes SAA squeezed axial attention and DEK detail enhancement kernel. The global semantics is captured by SAA axial attention, and the local texture is restored by combining with the DEK detail enhancement kernel.

[0059] The SAA squeeze axial attention is used to squeeze the input features horizontally and vertically respectively, broadcasting the global semantics through multi-head attention, and the output feature map size remains unchanged;

[0060] The DEK detail enhancement kernel concatenates the original features with the SAA output, extracts local details through deep convolution, generates enhancement weights and multiplies them with the SAA features to restore edge local information.

[0061] In another specific embodiment, through the analysis of the baseline semantic segmentation network decoder, it was found that the decoder completely relies on the multi-layer perceptron for feature fusion, and the multi-layer perceptron is essentially a fully connected operation and cannot effectively capture local details (such as edge information and features of small targets). Therefore, the SEA squeeze-enhanced axial attention mechanism is added to the decoder to enhance local features. SEA consists of SAA squeeze-axial attention and DEK detail enhancement kernel. SAA squeeze-axial attention first squeezes the Q (focus position), K (attention weight) and V (feature aggregation) information of the initial feature map of the input horizontally, and then broadcasts the three horizontally squeezed information layers through the multi-head attention mechanism; similarly, the Q (focus position), K (attention weight) and V (feature aggregation) information of the initial feature map of the input vertically, and then broadcasts the three vertically squeezed information layers through the multi-head attention mechanism. The results of the horizontal and vertical squeezing after broadcasting through the multi-head attention mechanism are then element-wise added and then passed through a 1×1 convolution. The SAA operation is equivalent to propagating information only along two compressed axial features at each position in the feature map, retaining global information on a single axis, thereby extracting global semantic information but sacrificing local details. The DEK detail enhancement kernel first concatenates the Q (position of attention), K (attention weight), and V (feature aggregation) information of the initial input feature map, concatenates them in the channel dimension, and then passes them to a 3×3 depthwise convolution and batch normalization. Through the 3×3 convolution, auxiliary local details are gathered from Q (position of attention), K (attention weight), and V (feature aggregation). Finally, linear projection with an activation function and batch normalization are used to generate detail enhancement weights, thereby enhancing spatial details. Multiplying and fusing the features given by DEK and SAA allows the capture of long-range image dependencies while maintaining a high feature map resolution, ensuring the integrity and accuracy of the information. In summary, SAA compresses feature maps horizontally and vertically, preserving global semantic information. DEK compensates for the local details lost during the compression process of axial attention by extracting local features using 3×3 depthwise convolutions and fusing these details with the output of axial attention. This design captures effective local detail features and improves the detection accuracy of the model.

[0062] Specifically, the high-precision semantic segmentation network serves as the teacher network, transferring feature expression capabilities to the lightweight semantic segmentation network through knowledge distillation. The lightweight semantic segmentation network serves as the student network, learning the feature expression of the teacher network through knowledge distillation.

[0063] Specifically, the lightweight semantic segmentation network is trained through knowledge distillation to learn the high-precision semantic segmentation network feature extraction capability, and the adversarial training of the lightweight semantic segmentation network is achieved through generative adversarial methods.

[0064] Knowledge distillation training uses the KL divergence loss function to enable the student network to learn the probability distribution of the teacher network;

[0065] Adversarial training introduces a discriminator to distinguish the outputs of the teacher network and the student network, and optimizes the student network through adversarial training.

[0066] In another specific embodiment, the high-precision cable-stayed apparent defect segmentation network based on multi-scale semantic fusion has the problems of large spatial occupancy and parameter redundancy, which leads to its lack of sufficient real-time performance in practical applications. Therefore, a joint training framework of knowledge distillation and generative adversarial is designed, in which the high-precision cable-stayed apparent defect segmentation network based on multi-scale semantic fusion is used as the teacher network, and the lightweight semantic segmentation network is used as the student network. The student network is allowed to learn the feature expression in the teacher network. In order to improve the effect of knowledge distillation, the generative adversarial network strategy is introduced in the training process, and the lightweight semantic segmentation network as the student network is used as the generator. Based on past experience, a discriminator is built through 4 standard convolutions with a convolution kernel size of 4 and a stride of 2 and 1 global average pooling layer. Through repeated iterative training, it is ensured that the trained lightweight semantic segmentation network can segment the cable-stayed apparent defects with high precision and efficiency.

[0067] Furthermore, a discriminator was constructed using 4 standard convolutional layers and 1 global average pooling layer, where the convolution kernel size of each standard convolutional layer is 4 and the stride is 2, and each standard convolution is followed by a LeakyReLU activation function and a batch normalization layer, which can avoid "neuron death", accelerate model training, and improve model stability; the teacher network and the student network are both semantic segmentation networks, and both are composed of encoder-decoders. The encoder of the teacher network adopts global modeling and multi-scale feature extraction, and the decoder of the teacher network adopts multi-layer perceptron direct prediction; the encoder of the student network adopts local optimization and lightweight convolutional layers, and the decoder of the student network adopts attention-assisted fusion. The integrated structure of the student network and the generator is further designed to unify the knowledge distillation and generative adversarial tasks. The joint KL divergence loss, supervised training loss and generative adversarial loss function are adopted to implement model training. The knowledge distillation includes KL divergence loss and supervised training loss. The KL divergence loss is used to measure the difference between the predicted distribution of the student network and the output probability distribution of the teacher network. Knowledge distillation is achieved by minimizing the KL divergence of the two. The probability distribution is reshaped by the knowledge distillation temperature parameter, and the softmax output probability distribution of the teacher network and the student network is smoothly regulated to achieve the effect of knowledge distillation. The supervised training loss optimizes the model classification performance and focuses on improving the model segmentation performance. The generative adversarial loss is included in the generative adversarial process. It is necessary to fix one network first and then train the other network. Through iterative training, the output distribution of the generator (student network) gradually approaches the probability distribution of the teacher network, thereby achieving high-precision segmentation of apparent defects of lightweight inclined cables.

[0068] Specifically, a high-precision lightweight stay-cable apparent defect segmentation system based on deep learning includes:

[0069] Image acquisition module: obtains the apparent damage image of the inclined cable;

[0070] Preprocessing module: Perform image preprocessing and image enhancement methods on the apparent damage images of the inclined cables to obtain the apparent damage dataset of the inclined cables;

[0071] Model building module: build the apparent damage segmentation model of the cable;

[0072] Output module: Input the cable apparent disease dataset into the cable apparent disease segmentation model and output the disease segmentation results.

[0073] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0074] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A high-precision method for segmenting apparent defects of lightweight stay cables based on deep learning, characterized by: include: Step 1: Obtain the apparent damage image of the cable; Step 2: Perform image preprocessing and image enhancement methods on the cable-stayed apparent damage images to obtain the cable-stayed apparent damage dataset; Step 3: Construct the apparent damage segmentation model of the stay cable; Step 4: Input the cable-stayed cable apparent disease dataset into the cable-stayed cable apparent disease segmentation model and output the disease segmentation results.

2. The high-precision method for segmenting apparent defects of lightweight stay cables based on deep learning according to claim 1 is characterized in that: In step 2, the image is pre-processed by using histogram equalization and Gaussian filtering, and then the image enhancement method of rotation, scaling and cropping is used to expand the number of diseased images to obtain the apparent disease data set of the inclined cable.

3. The high-precision lightweight stay cable apparent defect segmentation method based on deep learning according to claim 1 is characterized in that: The cable apparent defect segmentation model includes a high-precision semantic segmentation network and a lightweight semantic segmentation network, which work together through knowledge distillation and a generative adversarial mechanism.

4. The high-precision method for segmenting apparent defects of lightweight stay cables based on deep learning according to claim 3 is characterized in that: The high-precision semantic segmentation network includes the baseline semantic segmentation network, the SFF semantic feature fusion extraction network, and the SEA squeeze-enhanced axial attention network; The corresponding SFF semantic feature fusion extraction network is fused between the encoder and decoder of the baseline semantic segmentation network, and the SEA squeeze-enhanced axial attention network is introduced into the decoder of the baseline semantic segmentation network; The SFF semantic feature fusion extraction network is used to fuse the feature maps of the cable-stayed cable apparent disease dataset at different scales, and the SEA squeeze-enhanced axial attention network is used to enhance the local detail perception capability. The SFF semantic feature fusion extraction network first fuses the feature maps of different scales output by the encoder, and the fused feature maps are passed as input to the SEA squeeze enhanced axial attention network.

5. The high-precision method for segmenting apparent defects of lightweight stay cables based on deep learning according to claim 4 is characterized in that: The SEA squeezed enhanced axial attention network includes SAA squeezed axial attention and DEK detail enhancement kernel, which captures global semantics through SAA axial attention and restores local texture in combination with DEK detail enhancement kernel; The SAA squeeze axial attention is used to squeeze the input features horizontally and vertically respectively, broadcasting the global semantics through multi-head attention, and the output feature map size remains unchanged; The DEK detail enhancement kernel concatenates the original features with the SAA output, extracts local details through deep convolution, generates enhancement weights and multiplies them with the SAA features to restore edge local information.

6. The high-precision method for segmenting apparent defects of lightweight stay cables based on deep learning according to claim 3 is characterized in that: The high-precision semantic segmentation network serves as the teacher network and transfers feature expression capabilities to the lightweight semantic segmentation network through knowledge distillation. The lightweight semantic segmentation network serves as the student network and learns the feature expression of the teacher network through knowledge distillation.

7. The high-precision method for segmenting apparent defects of lightweight stay cables based on deep learning according to claim 6 is characterized in that: Through knowledge distillation training, a lightweight semantic segmentation network is used to learn high-precision semantic segmentation network feature extraction capabilities, and adversarial training of lightweight semantic segmentation networks is achieved through generative adversarial methods. Knowledge distillation training uses the KL divergence loss function to enable the student network to learn the probability distribution of the teacher network; Adversarial training introduces a discriminator to distinguish the outputs of the teacher network and the student network, and optimizes the student network through adversarial training.

8. A high-precision lightweight cable apparent defect segmentation system based on deep learning, characterized by: include: Image acquisition module: obtains the apparent damage image of the inclined cable; Preprocessing module: Perform image preprocessing and image enhancement methods on the apparent damage images of the inclined cables to obtain the apparent damage dataset of the inclined cables; Model building module: build the apparent damage segmentation model of the cable; Output module: Input the cable apparent disease dataset into the cable apparent disease segmentation model and output the disease segmentation results.

Images