Large-span suspension bridge support appearance defect image detection method and system

By combining the improved YOLO1 algorithm with specific modules, a defect image segmentation network was constructed, which solved the problems of low accuracy and insufficient efficiency in the detection of appearance defects in the bearings of long-span suspension bridges. It achieved accurate identification and localization of multi-scale defects in complex backgrounds, improving the accuracy and reliability of detection.

CN121121391APending Publication Date: 2025-12-12HUBEI COMM PLANNING & DESIGN INST CO LTD +2
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202511019204.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-23
Publication Date
2025-12-12

AI Technical Summary

Technical Problem

Existing methods for detecting surface defects in bearings of long-span suspension bridges struggle to effectively extract multi-scale defect features in complex environments, resulting in low detection accuracy and insufficient efficiency, failing to meet the scientific and timely requirements of bridge maintenance.

Method used

An improved YOLO1 algorithm is adopted, and the C3k2_REPVGGOREPA and C3k2_EMSCP modules are introduced. Combined with the loss function of support defect features, a defect image segmentation network is constructed to perform multi-stage feature extraction, fusion and dimensionality reduction processing, thereby improving the ability to identify multi-scale defects in complex backgrounds.

Benefits of technology

It enables accurate identification and location of surface defects in the bearings of long-span suspension bridges, improving the accuracy and reliability of inspection and meeting the needs of bridge safety assessment and maintenance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121121391A_ABST
    Figure CN121121391A_ABST
Patent Text Reader

Abstract

The invention discloses a large-span suspension bridge support appearance defect image detection method and system, and the method comprises the steps: building a defect image segmentation network: introducing a module for multi-scale feature extraction and structure re-parameterization and a module for multi-scale cavity convolution and adaptive feature fusion into a backbone network, a loss function combined with appearance defect characteristics of the large-span suspension bridge support is added; multi-stage feature extraction is carried out on an input image through the two modules, and fusion and dimension reduction processing is carried out on the extracted features. Inputting a processing result into a head network, and performing up-sampling, feature splicing, multi-scale feature extraction and structure re-parameterization processing for multiple times; constructing a sample data set and completing model training; and obtaining defect information of the to-be-detected defect image based on the trained model. And the accuracy and the reliability of support appearance defect detection are effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of bridge inspection and image processing technology, and more specifically, relates to a method and system for detecting appearance defects in the bearings of long-span suspension bridges. Background Technology

[0002] As a crucial component of transportation infrastructure, long-span suspension bridges rely heavily on their bearings, which are the core parts connecting the superstructure and substructure. These bearings play a vital role in transferring loads and accommodating structural deformation. During long-term service, bearings are susceptible to various factors, including vehicle loads, environmental erosion, and material aging, leading to surface defects such as cracks, corrosion, wear, and deformation. If these defects are not detected and addressed promptly, they can gradually expand, affecting the bearing's load-bearing capacity and stability, ultimately threatening the overall structural safety of the bridge and even causing serious traffic accidents.

[0003] Traditional inspection of bearing defects relies primarily on manual patrols. Inspectors must use elevated platforms or bridge inspection vehicles to approach the bearings and record defect information through visual observation and ruler measurements. This method is not only inefficient and unable to meet the large-scale inspection needs of long-span suspension bridges, but also suffers from high labor intensity and long inspection cycles. Furthermore, manual inspection results are easily influenced by subjective factors such as the inspector's experience, sense of responsibility, and visual fatigue, leading to missed defects, misjudgments, or measurement errors, making it difficult to guarantee the accuracy and consistency of the inspection.

[0004] With the development of computer vision technology, image-based automated detection methods are gradually being applied to the field of bridge defect detection. Existing methods mostly employ traditional machine learning or general deep learning algorithms to process and analyze acquired bearing images to identify and locate defects. However, the bearing structures of long-span suspension bridges are complex, with diverse surface textures, and are often located in environments with uneven lighting and complex backgrounds (such as interference from surrounding steel structures, dust, and moisture). This results in diverse defect morphologies and large scale variations, with low contrast between small-scale defects (such as microcracks) and the background.

[0005] Existing technologies have significant shortcomings when dealing with such scenarios: on the one hand, they struggle to effectively extract multi-scale defect features, resulting in low detection accuracy for small-scale defects; on the other hand, they are poorly able to distinguish defects from the background under complex interference, leading to false detections. These problems prevent existing automated detection methods from meeting the stringent requirements of actual engineering projects for bearing defect detection, thus hindering the scientific rigor and timeliness of bridge maintenance decisions. Therefore, developing an image detection method for the appearance defects of long-span suspension bridge bearings that can adapt to complex environments and accurately identify multi-scale defects has become an urgent need in the field of bridge inspection. Summary of the Invention

[0006] This invention aims to address the problems of low accuracy and insufficient efficiency in existing methods for detecting visual defects in the bearings of long-span suspension bridges. By introducing the C3k2_REPVGGOREPA and C3k2_EMSCP modules based on an improved YOLO11 algorithm to enhance feature extraction capabilities, and incorporating a loss function that integrates bearing defect features, a highly efficient defect image segmentation network is constructed. This network enables accurate identification and localization of multi-scale defects in complex backgrounds, providing reliable support for bridge safety assessment and maintenance decisions.

[0007] To address the aforementioned deficiencies or improvement needs of existing technologies, as a first aspect of this invention, the present invention provides a method for detecting visual defects in the bearings of long-span suspension bridges using images, comprising:

[0008] S1. Establish a defect image segmentation network, and introduce the C3k2_REPVGGOREPA module configured for multi-scale feature extraction and structural reparameterization and the C3k2_EMSCP module configured for multi-scale dilated convolution and adaptive feature fusion into its backbone network; and add a loss function that combines the appearance defect features of long-span suspension bridge supports into the defect image segmentation network.

[0009] S2. Multi-stage feature extraction is performed on the input image through the C3k2_REPVGGOREPA and C3k2_EMSCP modules in the backbone network, and the extracted features are input to the SPPF module to complete the fusion and dimensionality reduction of the input features;

[0010] S3. Input the fusion and dimensionality reduction results into the head network for upsampling, feature concatenation, and processing by the C3k2_REPVGGOREPA module;

[0011] S4. Construct a sample dataset and train the defect image segmentation network to obtain a trained defect image segmentation network; input the defect image to be detected into the trained defect image segmentation network to obtain the defect information of the defect image to be detected.

[0012] Furthermore, the process of constructing the loss function in S1 is as follows:

[0013] Calculate the area S of the intersection of the predicted defect region bounding box and the actual defect region bounding box. intersection Calculate the area S of the union of the predicted defect region bounding box and the actual defect region bounding box. union The calculation method is as follows:

[0014] S union =S prediction +S ground-truth -S intersection

[0015] Among them, Sprediction To predict the area of ​​the defect region's bounding box, S ground-truth This represents the area of ​​the bounding box of the actual defect region.

[0016] Calculate the intersection-union ratio (IoU):

[0017]

[0018] Construct the loss function L:

[0019] L=1-IoU+αL ce +βL dice

[0020] Among them, L ce Cross-entropy loss is used to measure the difference between the predicted class and the true class; L dice α is the Dice loss, used to enhance the detection capability of small target defects; α and β are the balancing weight coefficients.

[0021] Furthermore, the feature extraction process of the C3k2_REPVGGOREPA module in S2 is as follows:

[0022] For the input feature map, the Conv layer in the C3k2_REPVGGOREPA module performs preliminary feature transformation through convolution operations, outputting a feature map, which is calculated as follows:

[0023]

[0024] Among them, W i The weights of the i-th convolutional kernel are represented by *, where * represents the convolution operation, X represents the input feature map, and b represents the weights of the i-th convolutional kernel. i represents the bias corresponding to the i-th element, and N represents the number of convolutional kernels;

[0025] Subsequently, using the RepVGGBlock_OREPA structure, the multi-branch structure is reparameterized into a single-path convolution during the inference stage, further improving feature extraction efficiency and expressive power, and outputting a feature map Y. REPVGGOREPA .

[0026] Furthermore, the specific process of reparameterizing the multi-branch structure into a single-path convolution during the inference phase is as follows:

[0027] Let the outputs of each branch during training be:

[0028] Y i =W i *X+b i

[0029] Where i = 1, 2, ..., n, W i Here, * represents the kernel weights of the i-th branch, * represents the convolution operation, X represents the input feature map, and b represents the weights of the convolution kernels of the i-th branch.i The bias for the i-th branch

[0030] Then the output of the single-path convolution after reparameterization is:

[0031] Y REPVGGOREPA = W eq * X + b eq

[0032] where W[[ID=十六]] eq [[ID=十七]]is the equivalent convolution kernel obtained by accumulating the weights W of each branch after an adaptation transformation (such as size padding, etc.), and b i is the equivalent bias obtained by accumulating the biases b of each branch after corresponding processing, thus achieving the conversion from multi-branch to single-path convolution. eq is the equivalent bias obtained by accumulating the biases b of each branch after corresponding processing, thus achieving the conversion from multi-branch to single-path convolution. i is the equivalent bias obtained by accumulating the biases b of each branch after corresponding processing, thus achieving the conversion from multi-branch to single-path convolution.

[0033] Furthermore, the process of feature extraction in the C3k2_EMSCP module in S2 is as follows:

[0034] The input feature map X is first divided into multiple sub-feature maps. For each sub-feature map, convolution operations are performed separately using convolution kernels of different scales, and the calculation method is:

[0035]

[0036] where represents the weight of the j-th convolution kernel with scale ks, X sub represents the sub-feature map,[[ID=四十]] represents the corresponding bias, and M represents the number of sub-feature maps;

[0037] Then, the convolution results of each sub-feature map are fused, and channel adjustment is performed through 1×1 convolution to obtain the final output feature map Y EMSCP .

[0038] Furthermore, the specific implementation process of the fusion and dimensionality reduction processing in S2 is as follows:

[0039] For the input feature map X, after performing a max-pooling operation with a pooling kernel size of s, the feature map Y is obtained. For the pixel value at the position (x, y) in the feature map Y, the calculation method is:

[0040]

[0041] where X(x + i, y + j) represents the pixel value at the position (x + i, y + j) in the input feature map, the value ranges of x and y are 0 ≤ x < W, 0 ≤ y < H, the size of the input feature map is W×H×C, W is the width, H is the height, C is the number of channels, and Y(x, y) represents the pixel value at the position (x, y) in the feature map Y.

[0042] Furthermore, the specific method of upsampling in S3 is as follows:

[0043] For the feature map X before upsampling pre X is obtained after a nearest neighbor upsampling operation with an upsampling factor of r. post For feature map X post The pixel value at position (x, y) is calculated as follows:

[0044]

[0045] Among them, X post (x,y) represents the feature map X. post The pixel value at position (x, y) in the middle. This indicates a round-down operation. This indicates that for feature map X pre After a nearest neighbor upsampling operation with an upsampling factor of r.

[0046] Furthermore, the specific method of feature concatenation in S3 is as follows: for one input feature in the input feature map with a dimension of C1×H1×W1 and another input feature with a dimension of C2×H2×W2, under the condition that H1=H2 and W1=W2, the dimension of the output feature after feature concatenation is (C1+C2)×H1×W1.

[0047] As a second aspect of the present invention, the present invention provides an image detection system for appearance defects of bearings of long-span suspension bridges, comprising:

[0048] A defect image segmentation network construction unit is used to build a defect image segmentation network. The backbone network introduces a C3k2_REPVGGOREPA module configured for multi-scale feature extraction and structural reparameterization and a C3k2_EMSCP module configured for multi-scale dilated convolution and adaptive feature fusion. Furthermore, a loss function combining the appearance defect features of long-span suspension bridge supports is added to the defect image segmentation network.

[0049] The multi-stage feature extraction unit is used to perform multi-stage feature extraction on the input image through the C3k2_REPVGGOREPA module and the C3k2_EMSCP module in the backbone network, and input the extracted features to the SPPF module to complete the fusion and dimensionality reduction of the input features;

[0050] The head network processing unit is used to input the results of fusion and dimensionality reduction into the head network for upsampling, feature concatenation, and processing by the C3k2_REPVGGOREPA module;

[0051] The training and detection unit is used to construct a sample dataset and train the defect image segmentation network to obtain a trained defect image segmentation network; input the defect image to be detected into the trained defect image segmentation network to obtain the defect information of the defect image to be detected.

[0052] As a second aspect of the invention, the invention provides a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor of any step of the method for detecting visual defects in the bearings of long-span suspension bridges.

[0053] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0054] 1. The present invention provides a method for detecting appearance defects in bearings of long-span suspension bridges. This method establishes a defect image segmentation network, introduces the C3k2_REPVGGOREPA and C3k2_EMSCP modules into the backbone network, and incorporates a loss function that combines the appearance defect features of long-span suspension bridge bearings, forming a targeted detection architecture. The C3k2_REPVGGOREPA module enhances the feature capture capability for defects at different scales through multi-scale feature extraction and structural reparameterization; the C3k2_EMSCP module improves the extraction effect of defect details in complex backgrounds through multi-scale dilated convolution and adaptive feature fusion. The loss function that combines bearing defect features better adapts to the morphological characteristics of defects, enabling the network to be targeted at bearing defects from the feature extraction stage, thus solving the problem of insufficient adaptability of traditional methods to multi-scale defects and complex backgrounds.

[0055] 2. The image detection method for appearance defects of long-span suspension bridge bearings of the present invention performs multi-stage feature extraction on the input image through the C3k2_REPVGGOREPA and C3k2_EMSCP modules in the backbone network, and then inputs the extracted features into the SPPF module for fusion and dimensionality reduction processing. During the multi-stage extraction, the C3k2_REPVGGOREPA module continuously optimizes the feature capture of defects of different sizes; the C3k2_EMSCP module continuously enhances the distinction between defects and the background. The SPPF module's fusion and dimensionality reduction of these features integrates the effective information extracted in the multi-stage extraction, reduces redundant data, makes subsequent processing more efficient, improves the overall integrity and accuracy of feature processing, and lays a good foundation for subsequent detection.

[0056] 3. The image detection method for appearance defects of bearings in long-span suspension bridges of the present invention achieves defect detection by inputting the fused and dimensionality-reduced results into a head network, performing multiple upsampling, feature stitching, and processing by the C3k2_REPVGGOREPA module. The multiple upsampling and feature stitching of the head network integrate feature information at different levels; the further processing by the C3k2_REPVGGOREPA module further enhances the expression of defect features. This processing method allows the network to more accurately locate and identify bearing defects, and the output defect information better meets actual detection needs, effectively improving the accuracy and reliability of bearing appearance defect detection. Attached Figure Description

[0057] Figure 1 This is a flowchart of the image detection method for the appearance defects of the bearings of long-span suspension bridges according to an embodiment of the present invention;

[0058] Figure 2 This is a schematic diagram of the defect image segmentation network according to an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the structure of the C3k2_REPVGGOREPA module according to an embodiment of the present invention;

[0060] Figure 4 This is a schematic diagram of the structure of the C3k2_EMSCP module according to an embodiment of the present invention;

[0061] Figure 5 This is a schematic diagram illustrating the construction process of the image sample library of appearance defects of long-span suspension bridge supports according to an embodiment of the present invention.

[0062] Figure 6 This is a system unit diagram of an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0064] Example 1

[0065] Please refer to Figure 1 This embodiment 1 provides a method for detecting visual defects in the bearings of long-span suspension bridges using images, including:

[0066] S1. Establish a defect image segmentation network, and introduce the C3k2_REPVGGOREPA module configured for multi-scale feature extraction and structural reparameterization and the C3k2_EMSCP module configured for multi-scale dilated convolution and adaptive feature fusion into its backbone network; and add a loss function that combines the appearance defect features of long-span suspension bridge supports into the defect image segmentation network.

[0067] S2. Multi-stage feature extraction is performed on the input image through the C3k2_REPVGGOREPA and C3k2_EMSCP modules in the backbone network, and the extracted features are input to the SPPF module to complete the fusion and dimensionality reduction of the input features;

[0068] S3. Input the fusion and dimensionality reduction results into the head network for upsampling, feature concatenation, and processing by the C3k2_REPVGGOREPA module;

[0069] S4. Construct a sample dataset and train the defect image segmentation network to obtain a trained defect image segmentation network; input the defect image to be detected into the trained defect image segmentation network to obtain the defect information of the defect image to be detected.

[0070] This embodiment 1 further elaborates on the above steps.

[0071] (1) Construction of Defect Image Segmentation Network

[0072] Please refer to Figure 2 When establishing the defect image segmentation network, the C3k2_REPVGGOREPA and C3k2_EMSCP modules are introduced into the backbone network. The former enhances the ability to capture defect features of different sizes through multi-scale feature extraction and structural reparameterization, while the latter improves the extraction accuracy of defect edge features in complex backgrounds through multi-scale dilated convolution and adaptive feature fusion. Simultaneously, a loss function incorporating the appearance defect features of long-span suspension bridge supports is added to the network. This is to optimize the network training direction based on the feature distribution characteristics of different defect types such as corrosion and deformation, making the model more focused on the key features of support defects. In a preferred embodiment, the defect image segmentation network is established based on the improved YOLO11 algorithm.

[0073] The workflow of this defect image segmentation network is as follows: The input image first undergoes multi-stage feature extraction through the C3k2_REPVGGOREPA and C3k2_EMSCP modules in the backbone network. The extracted features are then input to the SPPF module. The SPPF module fuses and reduces the dimensionality of the input features through pooling operations at different scales, reducing redundant information and retaining key features. The processed results are then input to the head network, which performs multiple upsampling and feature concatenation to restore spatial resolution. Finally, the images are processed again by the C3k2_REPVGGOREPA module to further enhance the expression of defect features.

[0074] In a preferred embodiment, the loss function is constructed as follows:

[0075] Calculate the area S of the intersection of the predicted defect region bounding box and the actual defect region bounding box. intersection Calculate the area S of the union of the predicted defect region bounding box and the actual defect region bounding box. union The calculation method is as follows:

[0076] S union =S prediction +S ground-truth -S intersection

[0077] Among them, S prediction To predict the area of ​​the defect region's bounding box, S ground-truth This represents the area of ​​the bounding box of the actual defect region.

[0078] Calculate the intersection-union ratio (IoU):

[0079]

[0080] Construct the loss function L:

[0081] L=1-IoU+αL ce +βL dice

[0082] Among them, L ce Cross-entropy loss is used to measure the difference between the predicted class and the true class; L dice α is the Dice loss, used to enhance the detection capability of small target defects; α and β are the balancing weight coefficients.

[0083] (2) Multi-stage feature extraction

[0084] Please refer to Figure 3The C3k2_REPVGGOREPA module first performs preliminary feature transformation through Conv layers to generate feature maps at multiple scales, capturing defect features from different dimensions and enhancing the model's sensitivity to defects of different sizes. Subsequently, the RepVGGBlock_OREPA structure is adopted, reparameterizing the multi-branch structure into a single-path convolution during the inference phase. This reduces computation and improves inference speed while preserving the rich feature representation capabilities of the multi-branch structure during training, enabling the model to maintain accurate defect feature extraction while operating efficiently.

[0085] In a preferred embodiment, the feature extraction process of the C3k2_REPVGGOREPA module is as follows:

[0086] For the input feature map, the Conv layer in the C3k2_REPVGGOREPA module performs preliminary feature transformation through convolution operations, outputting a feature map, which is calculated as follows:

[0087]

[0088] Among them, W i The weights of the i-th convolutional kernel are represented by *, where * represents the convolution operation, X represents the input feature map, and b represents the weights of the i-th convolutional kernel. i represents the bias corresponding to the i-th element, and N represents the number of convolutional kernels;

[0089] Subsequently, using the RepVGGBlock_OREPA structure, the multi-branch structure is reparameterized into a single-path convolution during the inference stage, further improving feature extraction efficiency and expressive power, and outputting a feature map Y. REPVGGOREPA .

[0090] In a preferred embodiment, the specific process of reparameterizing the multi-branch structure into a single-path convolution during the inference phase is as follows:

[0091] Let the outputs of each branch during training be:

[0092] Y i =W i *X+b i

[0093] Where i = 1, 2, ..., n, W i Here, * represents the kernel weights of the i-th branch, * represents the convolution operation, X represents the input feature map, and b represents the weights of the convolution kernels of the i-th branch. i The bias of the i-th branch

[0094] The output of a single convolution after reparameterization is:

[0095] Y REPVGGOREPA =W eq *X+b eq

[0096] Among them, W eq The weights W of each branch i The equivalent convolution kernel obtained by accumulating the kernels after adaptive transformation (such as size padding), b eq The bias b of each branch i The equivalent bias obtained by accumulating the corresponding processing is used to realize the conversion from multi-branch to single-path convolution.

[0097] Please refer to Figure 4 The C3k2_EMSCP module divides the input feature map into multiple sub-feature maps, which are then processed using convolutional kernels of different scales. This allows the model to focus on feature details of different sizes, enhancing its adaptability to complex defect morphologies. Subsequently, the convolution results of each sub-feature map are fused and the channels are adjusted using a 1×1 convolution, which integrates multi-scale feature information, reduces feature redundancy, and improves the compactness of feature representation.

[0098] In a preferred embodiment, the feature extraction process of the C3k2_EMSCP module is as follows:

[0099] The input feature map X is first divided into multiple sub-feature maps. For each sub-feature map, a convolution operation is performed using kernels of different scales. The calculation method is as follows:

[0100]

[0101] in, Let X represent the weights of the convolution kernel with scale ks at the j-th scale. sub Represents sub-feature maps, This represents the corresponding bias, and M represents the number of sub-feature maps;

[0102] Then, the convolution results of each sub-feature map are fused together, and the channels are adjusted using a 1×1 convolution to obtain the final output feature map Y. EMSCP .

[0103] When the SPPF module fuses and reduces the dimensionality of input features, it uses max pooling operations at different scales. This operation can extract key information from the feature map, enhance the robustness of the features, and at the same time reduce the amount of data by reducing dimensionality, thereby reducing the computational complexity of subsequent calculations. This allows the model to process and analyze features more efficiently and improve the overall detection performance.

[0104] In a preferred embodiment, the specific implementation process of fusion and dimensionality reduction is as follows:

[0105] After the input feature map X undergoes a max pooling operation with a kernel size of s to obtain the feature map Y, the pixel value at position (x, y) in the feature map Y is calculated as follows:

[0106]

[0107] Among them, X(x+i, y+j) represents the pixel value at the position (x+i, y+j) in the input feature map. The value ranges of x and y are 0 ≤ x < W and 0 ≤ y < H. The size of the input feature map is W×H×C, where W is the width, H is the height, and C is the number of channels. Y(x, y) represents the pixel value at the position (x, y) in the feature map Y.

[0108] (3) Head network processing

[0109] After inputting the result of fusion and dimensionality reduction processing into the head network, upsampling, feature concatenation, and processing of the C3k2_REPVGGOREPA module are performed in sequence.

[0110] The upsampling adopts the nearest neighbor method to magnify the size of the feature map by a specified multiple. This is done to restore the spatial resolution of the feature map, enable the features to match with higher-resolution feature maps, provide a basis for subsequent accurate defect location, and at the same time, this method has simple calculation and can ensure processing efficiency.

[0111] In a preferred embodiment, the specific method of upsampling is as follows:

[0112] For the feature map X before upsampling pre , after performing the nearest neighbor upsampling operation with an upsampling multiple of r, X post is obtained. For the pixel value at the position (x, y) in the feature map X post , the calculation method is:

[0113]

[0114] Among them, X post (x, y) represents the pixel value at the position (x, y) in the feature map X post , represents the floor operation, represents performing the nearest neighbor upsampling operation with an upsampling multiple of r on the feature map X pre .

[0115] Feature concatenation needs to be performed on the premise that the sizes of the two input feature maps are the same, and the channel dimensions of the two are merged. This can integrate feature information at different levels, enable the model to utilize both deep semantic features and shallow detail features simultaneously, enhance the comprehensive description ability of defects, and avoid information limitations brought by single features.

[0116] In a preferred embodiment, the specific method of performing feature concatenation is: for an input feature in the input feature map with a dimension of C1×H1×W1 and another input feature with a dimension of C2×H2×W2, under the condition that H1 = H2 and W1 = W2 are satisfied, the dimension of the output feature after feature concatenation is (C1 + C2)×H1×W1.

[0117] After upsampling and feature concatenation, the model is further processed by the C3k2_REPVGGOREPA module. This module enhances the fused feature representation through multi-scale feature extraction and structural reparameterization, reduces redundant information, and improves the discriminative ability of the features, enabling the model to more accurately identify and segment defect regions, thus meeting the accuracy requirements of actual detection.

[0118] (4) Training and Testing

[0119] The process of constructing a sample dataset and training a defect image segmentation network, and then inputting the image to be detected into the trained network to obtain defect information, revolves around data quality and model practicality.

[0120] When constructing the sample dataset, original images were acquired from multiple sources, including on-site collection and historical data. Clear and fully labeled images were selected and retained as the foundation to ensure data reliability and prevent low-quality data from interfering with model learning. Expanding the dataset through rotation, mirroring, and adjustments to brightness and contrast enriches the diversity of samples, enabling the model to adapt to detection scenarios in different environments and reducing insufficient generalization ability caused by limited data. Labeling all images with information such as defect type, location, and size provides accurate learning basis for network training, helping the model establish the correspondence between defect features and labels.

[0121] Please refer to Figure 5 In a preferred embodiment, the process of constructing a sample library of appearance defects of bridge bearings for long-span suspension bridges specifically includes: acquiring appearance defect images of bridge bearings from various data sources such as on-site collection and historical inspection data of long-span suspension bridges to establish an original defect image set; filtering the appearance defect images of bridge bearings in the original defect image set according to preset requirements such as image clarity and completeness of defect annotations to obtain bridge bearing appearance defect images that meet the preset requirements to construct an original defect image sample library; performing rotation, mirroring, brightness adjustment, and contrast adjustment on the appearance defect images of bridge bearings in the original defect image sample library to obtain an extended defect image sample library; and annotating the appearance defect images of bridge bearings in the original defect image sample library and the extended defect image sample library with information such as defect type, location, and size to establish a sample library of appearance defects of bridge bearings for long-span suspension bridges.

[0122] The network is trained using a pre-constructed sample dataset. By iteratively optimizing the parameters, the network gradually learns the patterns for identifying support defects, thus improving detection accuracy. After training, the network can quickly output defect information when an image to be detected is input, achieving automated detection.

[0123] For the detection results, clustering multiple defect locations can merge duplicate or adjacent detection results, avoid the same defect being identified multiple times, and ensure the accuracy of the results.

[0124] Clustering and calculating the area of ​​defects can provide quantitative data for assessing the severity of defects, meet the needs of graded treatment of defects in engineering, and make the detection results more practical.

[0125] Example 2

[0126] Please refer to Figure 6 This embodiment 2 provides an image detection system for appearance defects of bearings in long-span suspension bridges, including:

[0127] A defect image segmentation network construction unit is used to build a defect image segmentation network. The backbone network introduces a C3k2_REPVGGOREPA module configured for multi-scale feature extraction and structural reparameterization and a C3k2_EMSCP module configured for multi-scale dilated convolution and adaptive feature fusion. Furthermore, a loss function combining the appearance defect features of long-span suspension bridge supports is added to the defect image segmentation network.

[0128] The multi-stage feature extraction unit is used to perform multi-stage feature extraction on the input image through the C3k2_REPVGGOREPA module and the C3k2_EMSCP module in the backbone network, and input the extracted features to the SPPF module to complete the fusion and dimensionality reduction of the input features;

[0129] The head network processing unit is used to input the results of fusion and dimensionality reduction into the head network for upsampling, feature concatenation, and processing by the C3k2_REPVGGOREPA module;

[0130] The training and detection unit is used to construct a sample dataset and train the defect image segmentation network to obtain a trained defect image segmentation network; input the defect image to be detected into the trained defect image segmentation network to obtain the defect information of the defect image to be detected.

[0131] Example 3

[0132] This embodiment 3 also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement any step of a method for detecting visual defects in the bearings of long-span suspension bridges.

[0133] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0134] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.

[0135] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting visual defects in the bearings of long-span suspension bridges, characterized in that, Including: S1. Establish a defect image segmentation network, introduce the C3k2_REPVGGOREPA module configured for multi-scale feature extraction and structural reparameterization and the C3k2_EMSCP module configured for multi-scale dilated convolution and adaptive feature fusion into its backbone network; and add a loss function that combines the appearance defect features of the large-span suspension bridge bearing to the defect image segmentation network; S2. Perform multi-stage feature extraction on the input image through the C3k2_REPVGGOREPA module and the C3k2_EMSCP module in the backbone network, and input the extracted features into the SPPF module to complete the fusion and dimensionality reduction processing of the input features; S3. Input the result after fusion and dimensionality reduction processing into the head network for upsampling, feature splicing, and processing by the C3k2_REPVGGOREPA module; S4. Construct a sample data set and train the defect image segmentation network to obtain a trained defect image segmentation network; input the defect image to be detected into the trained defect image segmentation network to obtain the defect information of the defect image to be detected.

2. The method for detecting visual defects in the bearings of long-span suspension bridges according to claim 1, characterized in that, The construction process of the loss function in S1 is as follows: Calculate the area S of the intersection of the predicted defect region bounding box and the actual defect region bounding box. intersection Calculate the area S of the union of the predicted defect region bounding box and the actual defect region bounding box. union The calculation method is as follows: S union =S prediction +S ground-truth -S intersection Among them, S prediction To predict the area of ​​the defect region's bounding box, S ground-truth This represents the area of ​​the bounding box of the actual defect region. Calculate the intersection over union IoU: Construct the loss function L: L=1-IoU+αL ce +βL dice Among them, L ce Cross-entropy loss is used to measure the difference between the predicted class and the true class; L dice α is the Dice loss, used to enhance the detection capability of small target defects; α and β are the balancing weight coefficients.

3. The method for detecting visual defects in the bearings of long-span suspension bridges according to claim 1, characterized in that, The feature extraction implementation process of the C3k2_REPVGGOREPA module in S2 is as follows: For the input feature map, the Conv layer in the C3k2_REPVGGOREPA module performs preliminary feature transformation through convolution operations and outputs a feature map, and its calculation method is: Among them, W i The weights of the i-th convolutional kernel are represented by *, where * represents the convolution operation, X represents the input feature map, and b represents the weights of the i-th convolutional kernel. i represents the bias corresponding to the i-th element, and N represents the number of convolutional kernels; Subsequently, using the RepVGGBlock_OREPA structure, the multi-branch structure is reparameterized into a single-path convolution during the inference stage, further improving feature extraction efficiency and expressive power, and outputting a feature map Y. REPVGGOREPA .

4. The method for detecting visual defects in the bearings of long-span suspension bridges according to claim 3, characterized in that, The specific process of reparameterizing the multi-branch structure into a single-path convolution in the inference stage is as follows: Let the outputs of each branch during training be: Y i =W i *X+b i Where i = 1, 2, ..., n, W i Here, * represents the kernel weights of the i-th branch, * represents the convolution operation, X represents the input feature map, and b represents the weights of the convolution kernels of the i-th branch. i The bias of the i-th branch Then the output of the reparameterized single-path convolution is: Y REPVGGOREPA =W eq *X+b eq Among them, W eq The weights W of each branch i The equivalent convolution kernel obtained by accumulating the kernels after adaptive transformation (such as size padding), b eq The bias b of each branch i The equivalent bias obtained by accumulating the corresponding processing is used to realize the conversion from multi-branch to single-path convolution.

5. The method for detecting visual defects in the bearings of long-span suspension bridges according to claim 1, characterized in that, The feature extraction implementation process of the C3k2_EMSCP module in S2 is as follows: The input feature map X is first divided into multiple sub-feature maps. For each sub-feature map, convolution operations are performed respectively through convolution kernels of different scales, and its calculation method is: in, Let X represent the weights of the convolution kernel with scale ks at the j-th scale. sub Represents sub-feature maps, This represents the corresponding bias, and M represents the number of sub-feature maps; Then, the convolution results of each sub-feature map are fused together, and the channels are adjusted using a 1×1 convolution to obtain the final output feature map Y. EMSCP .

6. The method for detecting visual defects in the bearings of long-span suspension bridges according to claim 1, characterized in that, The specific implementation process of the fusion and dimensionality reduction processing in S2 is as follows: For the input feature map X, after performing a max-pooling operation with a pooling kernel size of s, the feature map Y is obtained. For the pixel value at the position (x, y) in the feature map Y, the calculation method is: Where, X(x + i, y + j) represents the pixel value at the position (x + i, y + j) in the input feature map, the value ranges of x and y are 0 ≤ x < W, 0 ≤ y < H, the size of the input feature map is W × H × C, W is the width, H is the height, C is the number of channels, and Y(x, y) represents the pixel value at the position (x, y) in the feature map Y.

7. The method for detecting visual defects in the bearings of long-span suspension bridges according to claim 1, characterized in that, The specific method of upsampling in S3 is: For the feature map X before upsampling pre X is obtained after a nearest neighbor upsampling operation with an upsampling factor of r. post For feature map X post The pixel value at position (x, y) is calculated as follows: Among them, X post (x,y) represents the feature map X. post The pixel value at position (x, y) in the middle. This indicates a round-down operation. This indicates that for feature map X pre After a nearest neighbor upsampling operation with an upsampling factor of r.

8. The method for detecting visual defects in the bearings of long-span suspension bridges according to claim 1, characterized in that, The specific method of feature splicing in S3 is: For an input feature in the input feature map with a dimension of C1 × H1 × W1 and another input feature with a dimension of C2 × H2 × W2, under the condition that H1 = H2 and W1 = W2 are satisfied, the dimension of the output feature after feature splicing is (C1 + C2) × H1 × W1.

9. A system for detecting visual defects in the bearings of long-span suspension bridges using images, characterized in that, Including: A defect image segmentation network construction unit is used to build a defect image segmentation network. The backbone network introduces a C3k2_REPVGGOREPA module configured for multi-scale feature extraction and structural reparameterization and a C3k2_EMSCP module configured for multi-scale dilated convolution and adaptive feature fusion. Furthermore, a loss function combining the appearance defect features of long-span suspension bridge supports is added to the defect image segmentation network. The multi-stage feature extraction unit is used to perform multi-stage feature extraction on the input image through the C3k2_REPVGGOREPA module and the C3k2_EMSCP module in the backbone network, and input the extracted features to the SPPF module to complete the fusion and dimensionality reduction of the input features; The head network processing unit is used to input the results of fusion and dimensionality reduction into the head network for upsampling, feature concatenation, and processing by the C3k2_REPVGGOREPA module; The training and detection unit is used to construct a sample dataset and train the defect image segmentation network to obtain a trained defect image segmentation network; input the defect image to be detected into the trained defect image segmentation network to obtain the defect information of the defect image to be detected.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor according to any one of claims 1-8, a method for detecting visual defects in the bearings of long-span suspension bridges.

Citation Information

Cited By

  • Automatic detection method and system for shear deformation of bridge rubber support

    CN121883429A