Through-the-core unmanned aerial vehicle and method for detecting bridge cable members

Abstract

The application discloses a through-core unmanned aerial vehicle and a bridge cable component detection method, which comprises an unmanned aerial vehicle body and a limiting device. The body of the unmanned aerial vehicle body is of a through-core structure and is composed of two symmetrical detachable parts. The limiting device is arranged around the inner wall of the through-core structure of the body and is retractable. The application solves the problem of efficient detection of surface defects of large-span bridge cables and has the advantages of high mobility and high-precision identification.

Description

Technical Field

[0001] This invention relates to the field of cable-stayed bridge cable defect detection. More specifically, this invention relates to a through-hole unmanned aerial vehicle (UAV) and a method for detecting cable-stayed bridge components. Background Technology

[0002] A cable-stayed bridge (also known as a cable-stayed bridge) is a modern bridge type that transfers the load of the main girder to the towers via cable stays. It features large spans, lightweight structures, and aesthetically pleasing designs. The cable stays are the main load-bearing components of a bridge, composed of high-strength steel wire bundles or strands. Through an anchoring system, they transfer the load to the towers, converting the bending moment of the main girder into axial tension, significantly reducing the beam's bending moment and improving span capacity. However, cable stays are subject to long-term impacts from vehicles, wind vibration, and corrosion, leading to fatigue cracks, rust, and other defects on their surface. These defects can ultimately cause a decrease in the overall strength of the cable stays or even breakage. Accidents have already occurred in bridges due to cable breakage. Timely detection and repair of these defects during regular inspections can effectively prevent cable breakage accidents. Since manual inspection is risky and inefficient, developing equipment for detecting surface defects in cables is crucial to solving this problem. Currently, there are many devices available for detecting defects on the surface of cables, but most of them are bulky, slow to move, and cumbersome to install, making them unsuitable for inspecting bridges with a large number of cables and long spans. Therefore, designing a device that can quickly scan the surface of cables and is lightweight has become an urgent need. Summary of the Invention

[0003] The purpose of this invention is to provide a method for detecting through-hole drones and cable-stayed bridge components, solving the problem of rapid scanning and detection of surface defects in cables of long-span, multi-cable bridges.

[0004] The technical solution adopted by the present invention to solve this technical problem is: a through-hole type unmanned aerial vehicle (UAV), comprising: the UAV body and a limiting device; The drone's fuselage has a through-hole structure and consists of two symmetrical, detachable parts. The limiting device is located around the inner wall of the body's core structure and is retractable.

[0005] As a further aspect of the present invention, the drone body includes: fuselage, feet, drive motor, drone arms, propellers, scanning camera, antenna, battery box, and flight tower.

[0006] As a further aspect of the present invention, the fuselage has a frame-like structure, and the drone arms extend outwards at intervals around the frame. The limiting device includes a track pulley block, a pulley seat plate, a shock-absorbing spring, and a spring seat. The shock-absorbing spring is installed on the inner side of the frame through the spring seat, and the pulley block is connected to the shock-absorbing spring through the pulley seat plate. The limiting device consists of four sets of equally spaced devices arranged around the inner side of the frame, enabling the aircraft to overcome external influences and fly stably along the cable curve.

[0007] The present invention also provides a detection method using the aforementioned through-hole drone, comprising the following steps: S1. Use drones to collect images of the bridge's stay cables from multiple angles to ensure that the images cover the entire length of the stay cables and key stress areas, and preprocess the images. S2. Defect identification model construction, including: S21. Construct a deep image segmentation model based on the improved DeepplabV3+; S22. The encoder in the deep image segmentation model introduces a channel attention module to adaptively enhance the response of key channels; at the same time, a spatial attention module is introduced to focus on local detail regions in the image, forming a dual attention mechanism module. S23. Introduce an image pyramid structure into the model to enable the feature extractor to recognize features at both coarse and fine granular levels. S3. Train and optimize the model of the component; S4. Defect location and identification accuracy assessment, including: S41. The mask image output by the model is used to mark cracks and rust areas in the image, and to further extract their position coordinates, shape features and area indicators. S42. Evaluate the image quality of the detection results; S43. If the recognition result reaches the set threshold, export the structured data, including: defect image, category, location coordinates and evaluation value.

[0008] As a further aspect of the present invention, the image preprocessing in S1 specifically includes: S11. Perform geometric correction on the original image to correct spatial distortion in the image and make the structure in the image present its true geometric shape. Enhance image contrast to make pixel values ​​more evenly distributed; Highlighting edges in the image enhances the model's ability to perceive defects; S12. Background suppression is implemented through graph filtering functions to suppress repetitive textures and unstructured interference.

[0009] As a further aspect of the present invention, step S22 specifically includes the following steps: S221. Input the preprocessed image into the convolutional network, extract the initial feature map, and then perform non-linear mapping on the convolution output through the ReLU activation function; The feature map is fed into the dual attention mechanism module to obtain a deeper level of feature representation; after each convolution operation, the dual attention mechanism module is accessed once to achieve collaborative extraction of local convolution and global attention; the ReLU activation function is calculated as follows: The spatial attention module consists of two pooling operations and a main structure. The two pooling layers are global max pooling and global average pooling, respectively. The outputs of the two pooling operations are concatenated along the channel dimension to form a joint spatial feature map. Subsequently, this feature map undergoes a convolution operation to integrate spatial information and generates a spatial attention map through a sigmoid activation function. This attention map is used to weight and enhance the main spatial features and is finally passed as input to the channel attention module. S222, the channel attention module includes a global average pooling layer, a multilayer perceptron, and a Softmax activation function. The channel feature vectors extracted by pooling are first input into the multilayer perceptron, and then activated by Softmax to generate the attention distribution of each channel. Finally, the channel attention module outputs a set of attention weights for the main path and a set of non-attention weights that retain residual information, in order to realize the importance adjustment and residual compensation of the feature channels.

[0010] As a further aspect of the present invention, S3 specifically includes: S31. Collect a dataset of bridge cable images with crack and corrosion annotations, and divide it into training set, validation set and test set; S32. Input the preprocessed image of the defective bridge cable into the model for training. Guide the optimization of the model parameters by setting an appropriate loss function, thereby constructing a neural network model with effective image defect recognition capability. S33. Employ transfer learning combined with data augmentation to expand the training samples; S34. The Adam optimizer is used during training to adaptively adjust the learning rate.

[0011] As a further aspect of the present invention, S42 specifically includes: The peak signal-to-noise ratio (PSNR) is used for evaluation. The distortion between the resulting image and the original image is measured by calculating the error between corresponding pixels. The calculation formula is as follows: in, In the formula: Indicates mean square error; The maximum pixel value in the entire image is represented by ; m and n represent the width and height of the image, respectively; I represents the resulting image; and K represents the original image, obtained by summing the squares of the differences between corresponding pixels in the two images. As a further aspect of the present invention, S42 further includes: Structural similarity assessment measures the similarity between the denoised image and the original image in terms of brightness, contrast, and structure. The calculation formula is as follows: Where x is the original image; y is the predicted result image; and These represent the mean gray values ​​of x and y, respectively. and This represents the standard deviation of the gray values ​​of x and y; Let C1 and C2 represent the covariance of x and y; C1 and C2 are both constants.

[0012] The present invention has at least the following beneficial effects: 1. Employing a symmetrical, detachable, through-type fuselage structure, combined with four sets of retractable limiting devices (including shock-absorbing springs and track pulleys), the UAV can fly stably along the cable curve and quickly pass through protruding structures, avoiding the risk of collision. High-definition scanning cameras are deployed in four directions, and the offset control capability of the limiting devices ensures full coverage image acquisition of the cable surface even under wind interference. A single flight can complete a full cable scan, significantly improving inspection efficiency.

[0013] 2. Dual attention mechanism: The DeepLabV3+ encoder integrates channel attention (adaptively enhances key feature channels) and spatial attention (focuses on local detail areas), which significantly improves the sensitivity to the recognition of small cracks and corrosion boundaries.

[0014] 3. Multi-scale fusion: The introduction of an image pyramid structure enables the model to simultaneously possess coarse-grained localization and fine-grained segmentation capabilities, thereby improving the detection rate of micro-defects in complex backgrounds.

[0015] 4. A weighted combination of cross-entropy loss and Dice loss is adopted to simultaneously optimize pixel classification accuracy and defect region overlap.

[0016] 5. In the preprocessing stage, geometric correction, contrast stretching, and background filtering are used to suppress repetitive texture interference from bridges, enhancing defect visibility. PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity) are used as dual indicators to quantify the image quality, ensuring the reliability of the defect data. A structured report (including defect image, category, coordinates, and evaluation value) is output, directly supporting maintenance decisions and reducing manual re-inspection costs.

[0017] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0018] Figure 1 Front view of a through-hole drone; Figure 2 A top-down view of a through-hole drone; Figure 3 Axonometric view of a through-hole type unmanned aerial vehicle (UAV); Figure 4 Axonometric view of the fuselage; Figure 5 : Axonometric view of a scanning camera; Figure 6 Axonometric view of the limiting device; Figure 7 Axonometric view of the combined motor and propeller structure of a drone device; Figure 8 Axonometric view of the antenna of a drone device; Figure 9 Axonometric view of the drone's battery box; Figure 10 Flowchart of a bridge cable defect detection system; Figure 11 Add an overall network structure diagram; Figure 12 Diagram of the improved DeepplabV3+ deep recognition network; Figure 13 Channel attention module structure diagram; Figure 14 Spatial attention module structure diagram; Figure 15 Comparison of the original image and the defect identification results.

[0019] Among them, 1-UAV fuselage, 2-foot, 3-limiting device, 4-drive motor, 5-UAV arm, 6-propeller, 7-scanning camera, 8-antenna, 9-battery box, 10-flying tower, 11-shock-absorbing spring, 12-track pulley group. Detailed Implementation

[0020] The present invention will now be described in detail and completely with reference to the accompanying drawings. Those skilled in the art will be able to implement the present invention based on these descriptions. Before describing the present invention with reference to the accompanying drawings, it should be particularly noted that the technical solutions and features provided in various parts of the present invention, including the following description, can be combined with each other without conflict.

[0021] Furthermore, the embodiments of the present invention described below are generally only some, not all, of the embodiments of the present invention. Therefore, all other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The specific implementation process is as follows: like Figures 1-9 As shown, the present invention provides a through-hole type unmanned aerial vehicle (UAV), comprising: a UAV body and a limiting device; The drone's fuselage has a through-hole structure and is composed of two symmetrical, detachable parts. High-definition cameras or scanning cameras are installed on the fuselage. The limiting device is located around the inner wall of the body's core structure and is retractable.

[0023] In this embodiment, the frame structure combined with a high-definition camera or scanning camera can obtain high-quality image information, thereby ensuring the detection level. It leverages the advantages of the drone's high maneuverability and lightweight structure, while a limiting device ensures the drone can fly safely on the cable. In terms of detection efficiency, the drone's high maneuverability combined with the limiting device allows the equipment to safely and quickly scan the entire cable.

[0024] This technical solution may also include the following technical details to better achieve the technical effect: The drone body includes: fuselage 1, feet 2, drive motor 4, drone arm 5, propeller 6, scanning camera 7, antenna 8, battery box 9 and flying tower 10.

[0025] In the above technical solution, the drone's feet 2 are bolted to the drone's body 1, the limiting device 3 is bolted to the drone's body 1, the propeller 6 is bolted to the drive motor 4, the drive motor 4 is bolted to the drone's arm 5, the drone's arm 5 is bolted to the drone's body 1, the antenna 8 is bolted to the drone's body 1, the scanning camera 7 is bolted to the drone's body 1, and the battery box 9 is bolted to the drone's body 1. The flight tower 10 is bolted to the battery box 9. The drone's body 1 itself consists of two symmetrical parts connected by bolts and hinges. When the power in the battery box 9 is turned on, the electronic speed controller in the flight tower 10 will control the motor to rotate according to the preset current distribution, thereby driving the propeller 6 to generate lift. At this time, the flight controller in the flight tower 10 will control the aircraft to fly along the predetermined route, and the limiting device 3 enables the aircraft to overcome external influences and fly stably along the cable curve. Thus, the scanning camera 7 can quickly and accurately record the surface condition of the cable.

[0026] Specifically, the four scanning cameras of the through-hole drone are distributed in four vertical directions. When flying along the spatial curve of the cable, the limiting device in front of the scanning camera ensures that the device's deviation is within a controllable range. The drone body 1 itself consists of two symmetrical parts, and the rotation-switching device allows the device to be quickly disassembled and installed on the cable. It has four sets of propellers. When the device flies on the cable, changes in the shape of the cable curve and the lateral effect of the wind will cause the camera to deviate relative to the pole surface. At this time, the limiting device will ensure that the aircraft will not collide with the cable, allowing the aircraft to fly safely at a high speed. This not only ensures the reliable quality of the full cable scanning image, but also significantly reduces the overall inspection time.

[0027] This technical solution may also include the following technical details to better achieve the technical effect: the fuselage has a frame-like frame, and the drone arms extend outward at intervals around the frame; the feet are located around the bottom of the frame. The limiting device includes a track pulley block, pulley seat plate, shock-absorbing spring 11, and spring seat. The shock-absorbing spring is mounted on the inner side of the frame through the spring seat, and the pulley block is connected to the shock-absorbing spring through the pulley seat plate. The limiting device consists of four sets of equally spaced devices arranged around the inner side of the frame, enabling the aircraft to overcome external influences and fly stably along the cable curve. If the UAV encounters a protruding structure during flight, the spring can shorten to allow the UAV to pass through, while avoiding a violent impact. In this embodiment, the four sets of limiting devices are located at the four corners of the inner side of the fuselage frame and face the cable, enabling the aircraft to overcome external influences and fly stably along the cable curve. Thus, the scanning camera 7 can quickly and accurately record the surface condition of the cable. Preferably, the scanning camera is mounted at the four corners of the fuselage.

[0028] In another embodiment, such as Figures 10-15 As shown, the method for inspecting cable-stayed bridge components using a through-hole UAV includes the following steps: S1. Use drones to collect images of the bridge's stay cables from multiple angles to ensure that the images cover the entire length of the stay cables and key stress areas, and preprocess the images. S2. Defect identification model construction, including: S21, such as Figure 12 As shown, a deep image segmentation model based on the improved DeepplabV3+ is constructed. The model includes structures such as encoder, decoder, and skip connections. S22, such as Figure 13 and 14 As shown, the encoder in the deep image segmentation model introduces a channel attention module (SEblock) to adaptively enhance the response of key channels; at the same time, it introduces a spatial attention module (CBAM) to focus on local detail areas in the image, such as small cracks and rusted boundary areas, forming a dual attention mechanism module. S23. To enhance the model's multi-scale perception capability, an image pyramid structure is introduced into the model, enabling the feature extractor to have recognition capabilities at both coarse and fine granular levels. S3. Train and optimize the model of the component; S4. Defect location and identification accuracy assessment, including: S41. The mask image output by the model is used to mark cracks and rust areas in the image, and to further extract their position coordinates, shape features and area indicators. S42. Evaluate the image quality of the detection results. Commonly used metrics include Peak Signal-to-Noise Ratio (PSNR, R0). PSNR ) and structural similarity (SSIM, f SSIM ); S43, such as Figure 15 As shown, if the recognition result reaches the set threshold, structured data is exported, including: defect image, category, location coordinates and evaluation value.

[0029] This technical solution may also include the following technical details to better achieve the technical effect: The image preprocessing in S1 specifically includes: S11. Perform geometric correction on the original image to correct spatial distortions and restore the true geometric shape of the structures in the image. The formula is as follows: Where (x,y) represents the position coordinates of a pixel in the original image, a, b, c, d represent rotation, scaling and shearing factors that control the linear changes of the image, and e and f represent translation amount and control the overall offset of the image; Enhancing image contrast and making pixel values ​​more evenly distributed highlights defects (cracks or rust) in previously dark areas, especially making textures in dark areas more apparent, which helps the model capture fine cracks or light rust; the formula is as follows: Where I represents the gray value of a pixel in the original image, CDF(I) represents the cumulative probability of that gray value in the gray-level distribution of the entire image, and L represents the gray level. This represents the new grayscale value after the transformation.

[0030] Highlighting edges in the image enhances the model's ability to perceive defects, making the edges of small cracks and the boundaries of rust blocks clearer; S12. To address common background interference in bridge structures, background suppression is implemented using a graph filtering function to suppress repetitive textures and unstructured interference. The formula is as follows: in, This represents the image value after background suppression. Indicates the original image at position grayscale or pixel values, The background template image is represented by λ, which represents the suppression coefficient. This technical solution may also include the following technical details to better achieve the technical effect: S22 specifically includes the following steps: S221. Input the preprocessed image into the convolutional network, perform convolution operation to extract the initial feature map, and then perform non-linear mapping on the convolution output through the ReLU activation function to enhance the feature representation capability. The feature map is fed into the dual attention mechanism module (DACNN) to obtain a deeper level of feature representation; after each convolution operation, the dual attention mechanism module is connected once to achieve collaborative extraction of local convolution and global attention; the ReLU activation function is calculated as follows: The spatial attention module consists of two pooling operations and a main structure. The two pooling layers are global max pooling and global average pooling, used to extract spatial information from different statistical perspectives. The outputs of the two pooling operations are concatenated along the channel dimension to form a joint spatial feature map. Subsequently, this feature map undergoes a convolution operation to integrate spatial information and is then activated by a sigmoid activation function to generate a spatial attention map. This attention map is used to weight and enhance the main spatial features and is finally passed as input to the channel attention module. The expression for the Sigmoid activation function is as follows: Spatial Attention Module F c The calculation formula is as follows: in, For the sigmoid activation function, f i is a convolution operation with kernel size i, and Cat is a concatenation operation. and These are the feature maps after average pooling and after max pooling, respectively.

[0031] S222, the channel attention module includes a global average pooling layer, a multilayer perceptron (MLP), and a softmax activation function. The MLP consists of two fully connected layers and is responsible for modeling and compressing global information between channels. The channel feature vectors extracted by pooling are first input into the MLP, then activated by softmax to generate the attention distribution for each channel. Finally, the channel attention module outputs a set of attention weights for the main path and a set of non-attention weights that retain residual information, used to adjust the importance of feature channels and compensate for residuals.

[0032] The formula for the Softmax activation function is as follows: Inter-channel attention module F s The calculation formula is as follows: In the formula, The function is Softmax activation, W1 and W2 are the weights of the multilayer perceptron, and b1 and b2 are the biases of the multilayer perceptron. This is the feature map after average pooling.

[0033] This technical solution may also include the following technical details to better achieve the technical effect: S3 specifically includes: S31. Collect a dataset of bridge cable images with crack and corrosion annotations, and divide it into training set, validation set and test set; S32. Input the preprocessed images of bridge cables containing defects into the model for training. Guide the optimization of model parameters by setting an appropriate loss function, thereby constructing a neural network model with effective image defect recognition capabilities. This stage mainly includes the following steps: The batch size, initial learning rate, number of iterations, learning rate decay strategy, and loss function of the denoising network model are set, and the loss function is optimized. The loss function is a joint optimization of cross-entropy loss and Dice loss. The mathematical expression for cross-entropy loss is: in, For real labels, p i This represents the probability that the model predicts this pixel as a defect. The predicted probabilities are logarithmed to emphasize the penalty for low-confidence predictions, and then the summation is performed on each pixel as a whole.

[0034] The mathematical expression for the Dice loss function is: Among them, y iIndicates the real label, p i Indicates the predicted probability. For the overlapping area of ​​prediction and labeling, This represents the union of the prediction and the label. The closer the Dice value is to 1, the more accurate the prediction.

[0035] The mathematical expression for the joint loss function is as follows: Here, α and β represent the weights controlling the two loss components. By combining the weighted cross-entropy loss and Dice loss, the model's ability to classify pixels and distinguish between overlapping regions is comprehensively optimized.

[0036] S33. To enhance the model's generalization ability, transfer learning is used in conjunction with data augmentation (rotation, scaling, brightness perturbation, random occlusion, etc.) to expand the training samples. S34. During training, the Adam optimizer is used to adaptively adjust the learning rate and monitor the validation set loss to prevent overfitting.

[0037] This technical solution may also include the following technical details to better achieve the technical effect: S42 specifically includes: The peak signal-to-noise ratio (PSNR) is used for evaluation. The distortion between the resulting image and the original image is measured by calculating the error between corresponding pixels. The calculation formula is as follows: in, In the formula: Indicates mean square error; The maximum pixel in the entire image is represented by ; m and n represent the width and height of the image, respectively; I represents the resulting image; and K represents the original image, which is obtained by summing the squares of the differences between corresponding pixels in the two images.

[0038] This technical solution may also include the following technical details to better achieve the technical effect: S42 further includes: Structural similarity assessment measures the similarity between the denoised image and the original image in terms of brightness, contrast, and structure. The calculation formula is as follows: Where x is the original image; y is the predicted result image; μ x and μ y These represent the mean gray values ​​of x and y, respectively. and This represents the standard deviation of the gray values ​​of x and y; Let C1 and C2 represent the covariance of x and y; C1 and C2 are both constants.

[0039] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and embodiments shown and described herein.

Claims

1. A method for inspecting cable-stayed bridge components using a through-hole type unmanned aerial vehicle (UAV), characterized in that, The through-hole type drone includes: the drone body and the limiting device; the drone body has a through-hole structure and is composed of two symmetrical, detachable parts; the limiting device is located around the inner wall of the through-hole structure and is retractable. The detection method includes the following steps: S1. Use drones to collect images of the bridge's stay cables from multiple angles to ensure that the images cover the entire length of the stay cables and key stress areas, and preprocess the images. S2. Defect identification model construction, including: S21. Construct a deep image segmentation model based on the improved DeeplabV3+; S22. The encoder in the deep image segmentation model introduces a channel attention module to adaptively enhance the response of key channels; at the same time, a spatial attention module is introduced to focus on local detail regions in the image, forming a dual attention mechanism module. S23. Introduce an image pyramid structure into the model to enable the feature extractor to recognize features at both coarse and fine granular levels. S3. Train and optimize the model of the components, including: S31. Collect a dataset of bridge cable images with crack and corrosion annotations, and divide it into training set, validation set and test set; S32. Input the preprocessed image of the defective bridge cable into the model for training. Guide the optimization of the model parameters by setting an appropriate loss function, thereby constructing a neural network model with effective image defect recognition capability. S33. Employ transfer learning combined with data augmentation to expand the training samples; S34. The Adam optimizer is used during training to adaptively adjust the learning rate; S4. Defect location and identification accuracy assessment, including: S41. The mask image output by the model is used to mark cracks and rust areas in the image, and to further extract their position coordinates, shape features and area indicators. S42. Evaluate the image quality of the detection results; S43. If the recognition result reaches the set threshold, export the structured data, including: defect image, category, location coordinates and evaluation value.

2. The bridge cable-stayed component testing method as described in claim 1, characterized in that, The image preprocessing in S1 specifically includes: S11. Perform geometric correction on the original image to correct spatial distortion in the image and make the structure in the image present its true geometric shape. Enhance image contrast to make pixel values ​​more evenly distributed; Highlighting edges in the image enhances the model's ability to perceive defects; S12. Background suppression is implemented through graph filtering functions to suppress repetitive textures and unstructured interference.

3. The bridge cable-stayed component testing method as described in claim 1, characterized in that, S22 specifically includes the following steps: S221. Input the preprocessed image into the convolutional network, extract the initial feature map, and then perform non-linear mapping on the convolution output through the ReLU activation function; The feature map is fed into the dual attention mechanism module to obtain a deeper level of feature representation; after each convolution operation, the dual attention mechanism module is accessed once to achieve collaborative extraction of local convolution and global attention; the ReLU activation function is calculated as follows: The spatial attention module consists of two pooling operations and a main structure. The two pooling layers are global max pooling and global average pooling, respectively. The outputs of the two pooling operations are concatenated along the channel dimension to form a joint spatial feature map. Subsequently, this feature map undergoes a convolution operation to integrate spatial information and generates a spatial attention map through a sigmoid activation function. This attention map is used to weight and enhance the main spatial features and is finally passed as input to the channel attention module. S222, the channel attention module includes a global average pooling layer, a multilayer perceptron, and a Softmax activation function. The channel feature vectors extracted by pooling are first input into the multilayer perceptron, and then activated by Softmax to generate the attention distribution of each channel. Finally, the channel attention module outputs a set of attention weights for the main path and a set of non-attention weights that retain residual information, in order to realize the importance adjustment and residual compensation of the feature channels.

4. The method for testing cable-stayed bridge components as described in claim 1, characterized in that, S42 specifically includes: The peak signal-to-noise ratio (PSNR) is used for evaluation. The distortion between the resulting image and the original image is measured by calculating the error between corresponding pixels. The calculation formula is as follows: in, In the formula: Indicates mean square error; The maximum pixel in the entire image is represented by ; m and n represent the width and height of the image, respectively; I represents the resulting image; and K represents the original image, which is obtained by summing the squares of the differences between corresponding pixels in the two images.

5. The bridge cable-stayed component testing method as described in claim 4, characterized in that, S42 further includes: Structural similarity assessment measures the similarity between the denoised image and the original image in terms of brightness, contrast, and structure. The calculation formula is as follows: Where x is the original image; y is the predicted result image; μ x and μ y σ represents the mean gray level of x and y, respectively; x and σ y σ represents the standard deviation of the gray values ​​of x and y. xy Let C1 and C2 represent the covariance of x and y; C1 and C2 are both constants.

Images