Steel structure bridge panel detection method and system based on image recognition

Abstract

The present application provides a steel structure bridge panel detection method and system based on image recognition, which uses sample steel structure bridge panel images and their corresponding prior labeled defects to train a defect diagnosis neural network, so that the defect diagnosis neural network can learn the characteristics and laws of steel structure bridge panel defects. By continuously optimizing the network structure, the recognition accuracy and diagnostic accuracy of defects in candidate steel structure bridge panel images are improved. By automatically acquiring candidate steel structure bridge panel images and using the trained defect diagnosis neural network to perform defect diagnosis and generate defect diagnosis results, automated detection of steel structure bridge panels is achieved, greatly improving detection efficiency. In addition, the defect diagnosis neural network includes multiple defect diagnosis neural sub-networks, each prior labeled defect corresponds to a defect diagnosis neural sub-network, which can flexibly respond to different types of defects and has good scalability.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a method and system for detecting steel structure bridge panels based on image recognition. Background Art

[0002] With the acceleration of urbanization and the continuous development of transportation, steel structure bridges have been widely used in engineering construction. However, since steel structure bridges are exposed to harsh natural environments for a long time and bear increasing traffic loads, their panels are prone to various defects, such as corrosion, fatigue cracks, deformation, etc. If these defects are not discovered and repaired in time, they will seriously threaten the safe operation of the bridge.

[0003] At present, the traditional steel structure bridge panel inspection methods mainly rely on manual visual inspection and the use of some portable inspection equipment. However, these methods have problems such as low detection efficiency, low accuracy, and susceptibility to human factors. Especially in the inspection of large-scale and highly complex steel structure bridges, the limitations of traditional methods are more obvious. Summary of the invention

[0004] In view of this, the purpose of the present application is to provide a steel structure bridge panel detection method and system based on image recognition.

[0005] According to a first aspect of the present application, a method for detecting a steel structure bridge panel based on image recognition is provided, the method comprising:

[0006] Obtain a sample steel structure bridge panel image and at least one prior labeled defect corresponding to the sample steel structure bridge panel image, and train a defect diagnosis neural network based on the sample steel structure bridge panel image and the prior labeled defect corresponding to the sample steel structure bridge panel image; wherein the defect diagnosis neural network includes at least one defect diagnosis neural subnetwork, and each of the prior labeled defects corresponds to one of the defect diagnosis neural subnetworks;

[0007] Based on the requirements of this round of image recognition, corresponding network optimization instances are obtained, and the network structure of the defect diagnosis neural network is optimized;

[0008] A candidate steel structure bridge panel image is obtained, and defect diagnosis is performed on the candidate steel structure bridge panel image based on the defect diagnosis neural network after network structure optimization to generate a defect diagnosis result.

[0009] In a possible implementation of the first aspect, the defect diagnosis neural network training based on the sample steel structure bridge panel image and the prior labeled defects corresponding to the sample steel structure bridge panel image includes:

[0010] Generate corresponding defect representation image blocks from the sample steel structure bridge panel image based on the prior labeled defects, the prior labeled defects at least include fracture defects, corrosion defects and deformation defects, and the defect representation image blocks at least include fracture image blocks corresponding to the fracture defects, corrosion image blocks corresponding to the corrosion defects, and deformation image blocks corresponding to the deformation defects;

[0011] Defect diagnosis learning is performed based on at least one of the fracture image blocks, the corrosion image blocks and the deformation image blocks to generate the defect diagnosis neural network.

[0012] In a possible implementation of the first aspect, after obtaining a candidate steel structure bridge panel image, performing defect diagnosis on the candidate steel structure bridge panel image based on the defect diagnosis neural network after network structure optimization, and generating a defect diagnosis result, the method further includes:

[0013] Obtain defect diagnosis result retrieval instructions;

[0014] The construction service terminal that generates the defect diagnosis result retrieval instruction is verified, and for the defect diagnosis result retrieval instruction generated by the verified construction service terminal, the target steel structure bridge panel image and the target defect diagnosis result corresponding to the defect diagnosis result retrieval instruction are determined, and the target steel structure bridge panel image and the target defect diagnosis result are fed back to the construction service terminal.

[0015] In a possible implementation manner of the first aspect, the defect diagnosis neural network includes at least a damage state diagnosis network, and defect diagnosis learning is performed based on at least one of the fracture image blocks, the corrosion image blocks, and the deformation image blocks to generate the defect diagnosis neural network, including:

[0016] The self-attention unit based on the damage state diagnosis network generates a self-attention image block corresponding to the fracture image block from the sample steel structure bridge panel image;

[0017] Based on the fracture site classification unit of the damage state diagnosis network, the self-attention image block is classified to generate a first actual fracture site corresponding to the fracture image block;

[0018] The correlation analysis unit based on the damage state diagnosis network performs correlation analysis on the fracture image block and the first actual fracture site to generate a correlation loss parameter corresponding to the fracture image block;

[0019] The damage state classification unit based on the damage state diagnosis network judges the correlation loss parameter and generates a damage state classification result;

[0020] Based on the damage state classification result and the loss state label data corresponding to the sample steel structure bridge panel image, parameter learning is performed on the damage state diagnosis network to generate the damage state diagnosis network after parameter learning.

[0021] In a possible implementation of the first aspect, the correlation analysis unit based on the damage state diagnosis network performs a correlation analysis on the fracture image block and the first actual fracture site to generate a correlation loss parameter corresponding to the fracture image block, including:

[0022] Based on the feature representation branch in the correlation analysis unit, feature representation is performed on the fracture image block to generate a first defect representation vector corresponding to the fracture image block, and feature representation is performed on the first actual fracture site to generate a second defect representation vector corresponding to the first actual fracture site;

[0023] Retrieving the penetration feature vectors of the first defect characterization vector and the second defect characterization vector based on the penetration feature extraction branch in the correlation analysis unit, generating a first penetration parameter value corresponding to the first defect characterization vector and a second penetration parameter value corresponding to the second defect characterization vector, and then determining the fracture weight coefficient corresponding to the fracture image block and the first actual fracture site based on the first penetration parameter value and the second penetration parameter value;

[0024] Reconstructing the first actual break site based on the break weight coefficient to generate a second actual break site corresponding to the first actual break site;

[0025] Based on the fracture feature parameter unit in the correlation analysis unit, the fracture feature parameter of the fracture image block is calculated to generate a first fracture feature parameter corresponding to the fracture image block, the fracture feature parameter of the second actual fracture site is calculated to generate a second fracture feature parameter corresponding to the second actual fracture site, and the fracture feature parameter of the fracture image block and the second actual fracture site is jointly calculated to generate a corresponding third fracture feature parameter;

[0026] Based on the correlation loss determination branch in the correlation analysis unit, correlation loss is determined for the first fracture feature parameter, the second fracture feature parameter and the third fracture feature parameter to generate the correlation loss parameter corresponding to the fracture image block.

[0027] In a possible implementation of the first aspect, the defect diagnosis neural network further includes a corrosion diagnosis network, and defect diagnosis learning is performed based on at least one of the fracture image blocks, the corrosion image blocks, and the deformation image blocks to generate the defect diagnosis neural network, including:

[0028] The corrosion feature vector positioning unit based on the corrosion diagnosis network performs corrosion feature vector positioning on the corrosion image block to generate the corrosion feature vector corresponding to the corrosion image block;

[0029] A corrosion effect extraction unit based on the corrosion diagnosis network extracts corrosion effects from the corrosion feature vectors to generate corrosion effect vectors corresponding to the corrosion feature vectors;

[0030] Generate a predicted corrosion feature category corresponding to the corrosion image block based on the corrosion feature vector and the corrosion action vector;

[0031] The error unit based on the corrosion diagnosis network performs error calculation on the predicted corrosion feature category and the actual corrosion feature category corresponding to the sample steel structure bridge panel image, and performs parameter learning on the corrosion diagnosis network based on the calculated corrosion diagnosis error to generate the corrosion diagnosis network after parameter learning.

[0032] In a possible implementation of the first aspect, the defect diagnosis neural network further includes a deformation defect diagnosis network, and defect diagnosis learning is performed based on at least one of the fracture image blocks, the corrosion image blocks, and the deformation image blocks to generate the defect diagnosis neural network, including:

[0033] Determining a deformation defect diagnosis label corresponding to the deformation defect diagnosis network, and generating a reference image block corresponding to the deformation defect diagnosis label;

[0034] A deformation defect feature extraction unit based on the deformation defect diagnosis network extracts deformation defect features from the deformation image blocks to generate target deformation defect feature vectors corresponding to the deformation image blocks;

[0035] Calculating, based on a first comparison unit of the deformation defect diagnosis network, a first comparison parameter value corresponding to the target deformation defect feature vector in the reference image block, wherein the first comparison parameter value is an intra-class metric value;

[0036] Calculating, based on a second comparison unit of the deformation defect diagnosis network, a second comparison parameter value corresponding to the target deformation defect feature vector in the reference image block, wherein the second comparison parameter value is an out-of-class metric value;

[0037] The comparison fusion unit based on the deformation defect diagnosis network calculates the comparison fusion parameter value corresponding to the deformation defect diagnosis label of the deformation image block by using the first comparison parameter value and the second comparison parameter value;

[0038] Obtaining a maximum comparison fusion parameter value based on the deformation defect classification unit of the deformation defect diagnosis network, and determining the degree of deformation defect corresponding to the deformation image block based on the maximum comparison fusion parameter value;

[0039] The error unit based on the deformation defect diagnosis network performs error calculation on the deformation defect degree and the actual deformation defect degree corresponding to the sample steel structure bridge panel image, and performs parameter learning on the deformation defect diagnosis network based on the calculated deformation defect mask diagnosis error to generate the deformation defect diagnosis network after parameter learning.

[0040] In a possible implementation manner of the first aspect, the first comparison unit based on the deformation defect diagnosis network calculates a first comparison parameter value corresponding to the target deformation defect feature vector in the reference image block, including:

[0041] Using the first comparison unit to generate the number of blocks corresponding to the reference image block;

[0042] Calculate, based on the first comparison unit, a first weight parameter corresponding to the deformation defect diagnostic label of the target deformation defect feature vector;

[0043] The first comparison parameter value corresponding to the target deformation defect feature vector is calculated based on the first comparison unit using the number of blocks and the first weight parameter.

[0044] In a possible implementation manner of the first aspect, the second comparison unit based on the deformation defect diagnosis network calculates the second comparison parameter value corresponding to the target deformation defect feature vector in the reference image block, including:

[0045] Generating the number of labels corresponding to the deformation defect diagnosis label corresponding to the deformation defect diagnosis network based on the second comparison unit;

[0046] Calculate, based on the second comparison unit, a second weight parameter corresponding to the deformation defect diagnostic label of the target deformation defect feature vector;

[0047] The second comparison parameter value corresponding to the target deformation defect feature vector is calculated based on the second comparison unit using the number of labels and the second weight parameter.

[0048] According to the second aspect of the present application, a steel structure bridge panel detection system based on image recognition is provided, and the steel structure bridge panel detection system based on image recognition includes a machine-readable storage medium and a processor, and the machine-readable storage medium stores machine-executable instructions. When the processor executes the machine-executable instructions, the steel structure bridge panel detection system based on image recognition implements the aforementioned steel structure bridge panel detection method based on image recognition.

[0049] According to a third aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed, the aforementioned steel structure bridge panel detection method based on image recognition is implemented.

[0050] According to any one of the above aspects, the technical effect of the present application is:

[0051] The defect diagnosis neural network is trained using sample steel structure bridge panel images and their corresponding prior labeled defects, so that the defect diagnosis neural network can learn the characteristics and laws of steel structure bridge panel defects. By continuously optimizing the network structure, the recognition accuracy and diagnosis accuracy of defects in candidate steel structure bridge panel images are improved. By automatically acquiring candidate steel structure bridge panel images and using the trained defect diagnosis neural network to perform defect diagnosis and generate defect diagnosis results, the automatic detection of steel structure bridge panels is realized, which greatly improves the detection efficiency. In addition, the defect diagnosis neural network includes multiple defect diagnosis neural sub-networks, each prior labeled defect corresponds to a defect diagnosis neural sub-network, which can flexibly respond to different types of defects. Only by adding the corresponding defect diagnosis neural sub-network can the detection of new types of defects be realized, and it has good scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0053] Figure 1 A schematic flow chart of a steel structure bridge panel detection method based on image recognition provided in an embodiment of the present application.

[0054] Figure 2 A schematic diagram of the component structure of a steel structure bridge panel detection system based on image recognition for implementing the above-mentioned steel structure bridge panel detection method based on image recognition provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical scheme and advantages of the embodiment of the present application clearer, the technical scheme in the embodiment of the present application will be clearly and completely described below according to the drawings in the embodiment of the present application. It should be understood that the drawings in the present application only serve the purpose of explanation and description, and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn in real proportion. The flowchart used in this application shows the operations implemented according to some embodiments of the embodiment of the present application. It should be understood that the operations of the flowchart can be implemented out of sequence, and the steps without logical context can be reversed in order or implemented simultaneously. In addition, those skilled in the art can add multiple other operations to the flowchart under the guidance of the content of the present application, and can also destroy multiple operations from the flowchart.

[0056] Figure 1 The flowchart of the steel structure bridge panel detection method and system based on image recognition provided by the embodiment of the present application is shown. It should be understood that in other embodiments, the order of some steps in the steel structure bridge panel detection method based on image recognition of this embodiment can be shared with each other according to actual needs, or some steps can be omitted or maintained. The detailed steps of the steel structure bridge panel detection method based on image recognition include:

[0057] Step S110, obtaining a sample steel structure bridge panel image and at least one prior labeled defect corresponding to the sample steel structure bridge panel image, and training a defect diagnosis neural network based on the sample steel structure bridge panel image and the prior labeled defect corresponding to the sample steel structure bridge panel image. The defect diagnosis neural network includes at least one defect diagnosis neural subnetwork, and each of the prior labeled defects corresponds to one of the defect diagnosis neural subnetworks.

[0058] In this embodiment, the server first performs an operation of acquiring sample steel structure bridge panel images, which can be retrieved from a historical database or acquired in real time through a dedicated image acquisition device. For example, the server can be connected to a bridge detection system that regularly captures images of bridge panels and stores them on the server.

[0059] Next, the server obtains the prior labeled defects corresponding to these sample steel structure bridge panel images. These prior labeled defects may be previously identified and labeled on the sample steel structure bridge panel images by experts or automatic detection systems, indicating the specific defect types and locations of the bridge panels in the sample steel structure bridge panel images. The labeled defects may include fractures, corrosion, deformation, and other types.

[0060] The server then uses these sample steel bridge panel images with annotated defects to train a defect diagnosis neural network designed to recognize and classify different types of bridge panel defects. During the training process, the defect diagnosis neural network learns to extract features from the sample steel bridge panel images and judge the type and severity of the defects based on these features.

[0061] The structure of the neural network includes at least one defect diagnosis neural sub-network, each of which is responsible for identifying a specific defect type. For example, there can be a defect diagnosis neural sub-network dedicated to identifying fracture defects, another defect diagnosis neural sub-network for identifying rust defects, and so on. This modular design can improve the recognition accuracy of the neural network for specific defect types.

[0062] Step S120, obtaining a corresponding network optimization instance based on the current round of image recognition requirements, and optimizing the network structure of the defect diagnosis neural network.

[0063] In this embodiment, the server obtains network optimization instances according to the specific requirements of this round of image recognition, which may include recognition speed requirements, accuracy requirements, specific defect type concerns, etc. For example, if this round of recognition focuses on the fracture defect of the bridge panel, the server can obtain some network optimization instances specifically for fracture defect recognition.

[0064] After obtaining the network optimization instance, the server can optimize the network structure of the previously trained defect diagnosis neural network, for example, it can include adjusting the number of layers, number of nodes, connection mode, etc. of the neural network to improve the performance of the network on specific tasks. The optimization process may involve the use of various algorithms and techniques, such as pruning, quantization, distillation, etc., to reduce the computational complexity of the network and improve its real-time performance and accuracy.

[0065] Step S130, obtaining a candidate steel structure bridge panel image, performing defect diagnosis on the candidate steel structure bridge panel image based on the defect diagnosis neural network after network structure optimization, and generating a defect diagnosis result.

[0066] In this embodiment, the server starts to obtain candidate steel structure bridge panel images, which can be obtained from new bridge inspection tasks or images to be reviewed selected from the historical database. Similar to the sample steel structure bridge panel images, these candidate steel structure bridge panel images are also obtained through special image acquisition equipment.

[0067] After obtaining the candidate steel structure bridge panel images, the server can use the previously optimized defect diagnosis neural network to perform defect diagnosis on these candidate steel structure bridge panel images. The defect diagnosis neural network will first extract the features in the candidate steel structure bridge panel images, and then determine whether there are defects in the candidate steel structure bridge panel images, as well as the type and severity of the defects based on these features. This process is automated and can process a large amount of image data in a short time.

[0068] Finally, the server can generate defect diagnosis results, which can be a detailed report that lists the defect type, location, severity, and other information in each candidate steel structure bridge panel image. This result can be used in subsequent bridge maintenance and management decisions to help staff promptly detect and deal with potential safety hazards.

[0069] Based on the above steps, the defect diagnosis neural network is trained using sample steel structure bridge panel images and their corresponding prior labeled defects, so that the defect diagnosis neural network can learn the characteristics and laws of steel structure bridge panel defects. By continuously optimizing the network structure, the recognition accuracy and diagnostic accuracy of defects in candidate steel structure bridge panel images are improved. By automatically acquiring candidate steel structure bridge panel images and using the trained defect diagnosis neural network to perform defect diagnosis and generate defect diagnosis results, the automatic detection of steel structure bridge panels is realized, which greatly improves the detection efficiency. In addition, the defect diagnosis neural network includes multiple defect diagnosis neural sub-networks, each prior labeled defect corresponds to a defect diagnosis neural sub-network, which can flexibly respond to different types of defects. Only by adding the corresponding defect diagnosis neural sub-network can the detection of new types of defects be realized, and it has good scalability.

[0070] In a possible implementation, step S110 may include:

[0071] Step S111, based on the prior labeled defects, corresponding defect representation image blocks are generated from the sample steel structure bridge panel image, the prior labeled defects include at least fracture defects, corrosion defects and deformation defects, and the defect representation image blocks include at least fracture image blocks corresponding to the fracture defects, corrosion image blocks corresponding to the corrosion defects, and deformation image blocks corresponding to the deformation defects.

[0072] In this embodiment, the server identifies the location and range of the fracture defects in each sample steel structure bridge surface image, and based on these prior labeled defect information, uses image processing technology to segment the fractured part from the original image to generate fracture image blocks. These fracture image blocks clearly show the fracture conditions on the bridge panel, providing accurate material for subsequent defect diagnosis.

[0073] Next, the server performs similar processing on rust defects. By identifying the location and severity of the rust annotations, the server extracts the rusted areas from the original image and generates rust image blocks, which not only show the appearance of the rust, but also reflect the degree and distribution of the rust.

[0074] Finally, for deformation defects, the server also locates the deformed areas based on the prior labeled defect information, and uses image segmentation technology to extract these areas into deformed image blocks. These deformed image blocks can clearly reflect the deformation degree and shape change of the bridge deck.

[0075] Step S112, performing defect diagnosis learning based on at least one of the fracture image blocks, the corrosion image blocks and the deformation image blocks to generate the defect diagnosis neural network.

[0076] After completing the generation of fracture, corrosion and deformation image blocks, the server begins to use these fracture, corrosion and deformation image blocks to learn the defect diagnosis neural network.

[0077] The server first divides the fracture image into blocks as input and trains the neural network model through a deep learning algorithm so that it can accurately identify fracture defects. In this process, the neural network will learn the characteristics of fracture defects, such as the direction, width and depth of the crack, thereby establishing the ability to identify fracture defects.

[0078] Next, the server divides the rust image into blocks and inputs it into the neural network for learning. The neural network gradually builds a recognition model for rust defects by analyzing the color, texture, and shape of the rusted area. This model can accurately determine the degree and scope of rust, providing an important basis for subsequent maintenance and repair.

[0079] Finally, the server uses the deformed image blocks to further train the neural network. By learning the shape changes and displacements of the deformed area, the neural network can accurately identify the deformation defects and perform qualitative and quantitative analysis on them.

[0080] After the training and learning of the above steps, the server finally generated a defect diagnosis neural network with strong defect recognition capabilities. This neural network can automatically and accurately identify various defects on steel structure bridge panels, providing strong technical support for the safe maintenance and repair of bridges.

[0081] In a possible implementation manner, after step S130, the method further includes:

[0082] Step S140, obtaining a defect diagnosis result retrieval instruction.

[0083] Step S150, verifying the construction service terminal that generates the defect diagnosis result retrieval instruction, and determining the target steel structure bridge panel image and target defect diagnosis result corresponding to the defect diagnosis result retrieval instruction generated by the verified construction service terminal, and feeding back the target steel structure bridge panel image and the target defect diagnosis result to the construction service terminal.

[0084] In this embodiment, the server receives defect diagnosis result retrieval instructions from the construction service terminal. These defect diagnosis result retrieval instructions are usually issued by construction personnel or bridge maintenance personnel who need to evaluate the current condition of the bridge or plan maintenance work by viewing specific defect diagnosis results.

[0085] Before providing defect diagnosis results, the server first verifies the identity and authority of the construction service terminal that sent the retrieval instruction, which usually involves checking the terminal's unique identifier, login credentials, or access permission list, etc. Only verified terminals can continue to access sensitive data on the server.

[0086] Once the verification is passed, the server will search the database for the corresponding target steel structure bridge panel image and related target defect diagnosis results based on the specific parameters in the retrieval instruction (such as bridge ID, inspection date or defect type, etc.). These parameters help the server accurately locate the specific data required by the user.

[0087] Finally, the server sends the found target steel structure bridge panel image and the corresponding target defect diagnosis results to the construction service terminal that issued the retrieval instruction. Construction workers or bridge maintenance personnel can view these images and results on the terminal for further analysis, evaluation or maintenance plan.

[0088] In a possible implementation, the defect diagnosis neural network includes at least a damage status diagnosis network, and step S112 may include:

[0089] Step A110, based on the self-attention unit of the damage state diagnosis network, generates a self-attention image block corresponding to the fracture image block from the sample steel structure bridge panel image.

[0090] In this embodiment, the server first starts the self-attention unit in the damage state diagnosis network, which is designed to automatically focus on key areas in the sample steel structure bridge panel image, especially those areas that may contain fracture defects. The server inputs the fracture image block into the self-attention unit, which generates a self-attention image block by calculating the correlation between different positions in the image. This self-attention image block highlights the part of the original image that is most relevant to the fracture defect, providing important information for subsequent classification and correlation analysis.

[0091] Step A120, classifying the self-attention image block based on the fracture site classification unit of the damage status diagnosis network to generate a first actual fracture site corresponding to the fracture image block.

[0092] Next, the server uses the fracture site classification unit in the damage state diagnosis network to classify the self-attention image blocks. This classification unit has been trained to identify specific fracture sites in the image. The server inputs the self-attention image blocks into the classification unit, which generates the first actual fracture sites corresponding to the fracture image blocks by analyzing the features in the image, such as the direction, length and width of the cracks. These fracture sites are the exact locations of the fracture defects that actually exist in the image.

[0093] Step A130: performing a correlation analysis on the fracture image block and the first actual fracture site based on the correlation analysis unit of the damage status diagnosis network, and generating a correlation loss parameter corresponding to the fracture image block.

[0094] After determining the first actual fracture site, the server uses the correlation analysis unit in the damage state diagnosis network to perform correlation analysis on the fracture image blocks and these fracture sites. This analysis process aims to evaluate the degree of correlation between the fracture sites generated by the classification unit and the fracture defects in the original image. The server measures this correlation by calculating the correlation loss parameter between the fracture image blocks and the first actual fracture site. If the fracture site generated by the classification unit is highly consistent with the fracture defect in the original image, the correlation loss parameter will be low; conversely, if there is a large difference between the two, the correlation loss parameter will be high.

[0095] Step A140, judging the correlation loss parameter based on the damage state classification unit of the damage state diagnosis network, and generating a damage state classification result.

[0096] Based on the above correlation loss parameters, the server uses the damage state classification unit in the damage state diagnosis network to classify the damage state of the fracture defect. This classification unit can determine the severity or state of the fracture defect, such as slight fracture, moderate fracture or severe fracture, according to the size of the correlation loss parameter. The server inputs the correlation loss parameters into the classification unit, which generates a damage state classification result by analyzing these parameters. This result not only reflects the presence or absence of the fracture defect, but also provides detailed information about the severity of the defect.

[0097] Step A150, performing parameter learning on the damage state diagnosis network based on the damage state classification result and the loss state label data corresponding to the sample steel structure bridge panel image, to generate the damage state diagnosis network after parameter learning.

[0098] Finally, the server uses the above damage state classification results and the loss state label data corresponding to the sample steel structure bridge panel image to learn the parameters of the damage state diagnosis network. In this process, the server adjusts the parameters of the network to minimize the difference by comparing the difference between the classification results and the label data. This learning method enables the damage state diagnosis network to gradually adapt and accurately identify different types of fracture defects and their damage states. After multiple rounds of iterative learning, the server generates a damage state diagnosis network after parameter learning. This network will be able to more accurately diagnose fracture defects and their damage states on steel structure bridge panels in subsequent applications.

[0099] In a possible implementation, step A130 may include:

[0100] Step A131, based on the feature representation branch in the correlation analysis unit, feature representation is performed on the fracture image block to generate a first defect representation vector corresponding to the fracture image block, and feature representation is performed on the first actual fracture site to generate a second defect representation vector corresponding to the first actual fracture site.

[0101] In this embodiment, the server starts the correlation analysis unit in the damage state diagnosis network. First, the feature representation branch of the correlation analysis unit starts working. This branch receives the fracture image block and the first actual fracture site as input, and generates a first defect representation vector corresponding to the fracture image block and a second defect representation vector corresponding to the first actual fracture site by performing deep feature extraction on the image and the site. These two representation vectors are abstract representations of the input data and contain rich fracture defect information.

[0102] Step A132, based on the penetration feature extraction branch in the correlation analysis unit, the penetration feature vectors of the first defect characterization vector and the second defect characterization vector are retrieved to generate a first penetration parameter value corresponding to the first defect characterization vector and a second penetration parameter value corresponding to the second defect characterization vector, and then based on the first penetration parameter value and the second penetration parameter value, the fracture weight coefficient corresponding to the fracture image block and the first actual fracture site is determined.

[0103] Next, the penetration feature extraction branch in the correlation analysis unit processes the first defect characterization vector and the second defect characterization vector. This branch can retrieve the penetration feature vectors in the two characterization vectors, which reflect the penetration relationship between the image and the site of the fracture defect. Based on the retrieved penetration feature vectors, the branch calculates the first penetration parameter value and the second penetration parameter value, which respectively represent the penetration degree of the image and the site on the fracture defect. Subsequently, the server uses these two penetration parameter values ​​to determine the fracture weight coefficient corresponding to the fracture image block and the first actual fracture site. This weight coefficient reflects the importance of the image and site on the fracture defect.

[0104] Step A133: reconstruct the first actual break site based on the break weight coefficient to generate a second actual break site corresponding to the first actual break site.

[0105] With the fracture weight coefficient, the server begins to reconstruct the first actual fracture site. By applying the weight coefficient, the server adjusts and optimizes the position, shape and other properties of the first actual fracture site to generate a more accurate and detailed second actual fracture site. This reconstruction process aims to improve the correlation between the fracture site and the original image, and provide more reliable data for subsequent feature parameter calculation and loss determination.

[0106] Step A134, based on the fracture feature parameter unit in the correlation analysis unit, fracture feature parameter calculation is performed on the fracture image block to generate the first fracture feature parameter corresponding to the fracture image block, fracture feature parameter calculation is performed on the second actual fracture site to generate the second fracture feature parameter corresponding to the second actual fracture site, and fracture feature parameter calculation is performed jointly on the fracture image block and the second actual fracture site to generate the corresponding third fracture feature parameter.

[0107] The server continues to use the fracture characteristic parameter unit in the correlation analysis unit for calculation. First, the unit calculates the fracture characteristic parameters of the fracture image blocks to generate the first fracture characteristic parameters; then the same calculation is performed on the second actual fracture site to generate the second fracture characteristic parameters; finally, the unit performs a joint calculation on the fracture image blocks and the second actual fracture site to generate the third fracture characteristic parameters. These fracture characteristic parameters quantitatively describe various attributes of the fracture defect, such as length, width, direction, etc.

[0108] Step A135, performing correlation loss determination on the first fracture feature parameter, the second fracture feature parameter and the third fracture feature parameter based on the correlation loss determination branch in the correlation analysis unit, and generating the correlation loss parameter corresponding to the fracture image block.

[0109] Finally, the server determines the correlation loss parameter based on the correlation loss determination branch in the correlation analysis unit. This branch receives the first fracture feature parameter, the second fracture feature parameter, and the third fracture feature parameter calculated previously as input, and calculates the correlation loss parameter by analyzing the correlation and difference between these parameters. This loss parameter reflects the degree of correlation between the fracture image block and the actual fracture site and the degree of matching between them in fracture features. If the correlation between the two is strong and the matching degree is high, the loss parameter value will be low; otherwise, it will be high. The server uses this correlation loss parameter as an important basis for subsequent optimization and adjustment of the network model.

[0110] In a possible implementation, the defect diagnosis neural network further includes a corrosion diagnosis network, and step S112 may include:

[0111] Step B110: Based on the corrosion feature vector positioning unit of the corrosion diagnosis network, the corrosion feature vector of the corrosion image block is positioned to generate the corrosion feature vector corresponding to the corrosion image block.

[0112] The server starts the rust feature vector positioning unit in the rust diagnosis network. The rust feature vector positioning unit is responsible for processing rust image blocks and locating and extracting rust features in the image through a deep learning algorithm. Specifically, when the server receives a batch of rust image blocks, the rust feature vector positioning unit analyzes these images one by one, identifies the rust areas in the image, and generates corresponding rust feature vectors. These vectors not only contain the location information of the rust, but also encode features such as the shape, size, and severity of the rust.

[0113] Step B120: extracting corrosion effects from the corrosion feature vectors based on the corrosion effect extraction unit of the corrosion diagnosis network, and generating corrosion effect vectors corresponding to the corrosion feature vectors.

[0114] After obtaining the corrosion feature vector, the server further processes it using the corrosion effect extraction unit in the corrosion diagnosis network. The task of this unit is to deeply analyze the corrosion feature vector and extract key information about the corrosion effect. Through complex mathematical operations and pattern recognition technology, the corrosion effect extraction unit can reveal the interaction between different corrosion features and generate corrosion effect vectors. These vectors reflect the propagation mode and impact range of corrosion on the steel structure bridge panel, which is crucial for subsequent corrosion diagnosis and prediction.

[0115] Step B130: generating a predicted corrosion feature category corresponding to the corrosion image block based on the corrosion feature vector and the corrosion action vector.

[0116] Based on the extracted corrosion feature vectors and corrosion action vectors, the server begins to generate predicted corrosion feature categories corresponding to the corrosion image blocks. This step is achieved through the classification layer in the corrosion diagnosis network, which can automatically determine the type and level of corrosion based on the input vector data. For example, the server can classify corrosion into different categories such as mild, moderate and severe, and assign a unique identifier to each category. In this way, the server can quickly identify the categories to which different corrosion image blocks belong, providing decision support for subsequent bridge maintenance and repair work.

[0117] Step B140, based on the error unit of the corrosion diagnosis network, the error calculation is performed on the predicted corrosion feature category and the actual corrosion feature category corresponding to the sample steel structure bridge panel image, and the corrosion diagnosis network is parameter learned based on the calculated corrosion diagnosis error to generate the corrosion diagnosis network after parameter learning.

[0118] Finally, the server uses the error unit in the corrosion diagnosis network to calculate the error between the predicted corrosion feature category and the actual corrosion feature category. Specifically, the server can obtain a batch of sample steel structure bridge panel images with actual corrosion feature category labels and input these images into the corrosion diagnosis network for prediction. The error unit then compares the difference between the predicted results and the actual labels and calculates a corrosion diagnosis error value, which reflects the accuracy and reliability of the corrosion diagnosis network in the prediction process.

[0119] In order to continuously improve the performance of the corrosion diagnosis network, the server can learn the parameters of the network based on the calculated corrosion diagnosis error. By adjusting parameters such as weights and biases in the network, the server can gradually reduce the difference between the predicted results and the actual labels, thereby generating a corrosion diagnosis network after parameter learning. This process is achieved through the back propagation algorithm and gradient descent optimization technology, which can make the corrosion diagnosis network more adaptable to complex corrosion patterns and diverse bridge panel images.

[0120] In a possible implementation, the defect diagnosis neural network further includes a deformation defect diagnosis network, and step S112 may include:

[0121] Step C110, determining the deformation defect diagnosis label corresponding to the deformation defect diagnosis network, and generating a reference image block corresponding to the deformation defect diagnosis label.

[0122] The server first determines the deformation defect diagnosis labels corresponding to the deformation defect diagnosis network. These labels are pre-defined and used to describe various types of deformation defects that may occur in steel structure bridge panels, such as bulges, depressions, distortions, etc. Then, the server generates corresponding reference image blocks based on these labels. These reference image blocks are images without deformation defects under ideal conditions and are used as the benchmark for subsequent comparative analysis.

[0123] Step C120 , extracting deformation defect features from the deformation image blocks based on the deformation defect feature extraction unit of the deformation defect diagnosis network, and generating target deformation defect feature vectors corresponding to the deformation image blocks.

[0124] Next, the server starts the deformation defect feature extraction unit in the deformation defect diagnosis network. This unit receives the deformation image block as input and extracts the deformation defect features in the image through a deep learning algorithm. These features include the shape, size, direction of the deformation, and the relative relationship with other image contents. After the extraction is completed, the server generates the target deformation defect feature vector corresponding to the deformation image block.

[0125] Step C130, calculating a first comparison parameter value corresponding to the target deformation defect feature vector in the reference image block based on a first comparison unit of the deformation defect diagnosis network, wherein the first comparison parameter value is an intra-class metric value.

[0126] The server continues to use the first comparison unit of the deformation defect diagnosis network to calculate the first comparison parameter value corresponding to the target deformation defect feature vector in the reference image block. This parameter value is an intra-class metric value used to measure the similarity between the target deformation defect feature and other features in the same category (i.e., the same diagnostic label). By calculating this value, the server can determine whether the target deformation defect belongs to a known category.

[0127] Step C140, based on the second comparison unit of the deformation defect diagnosis network, calculates a second comparison parameter value corresponding to the target deformation defect feature vector in the reference image block, where the second comparison parameter value is an out-of-class metric value.

[0128] In addition to intra-class comparison, the server also needs to perform out-of-class comparison. Therefore, the server calls the second comparison unit of the deformation defect diagnosis network to calculate the second comparison parameter value corresponding to the target deformation defect feature vector in the reference image block. This parameter value is an out-of-class metric value used to measure the difference between the target deformation defect feature and the features of different categories (i.e., different diagnostic labels). By calculating this value, the server can further distinguish the boundary between the target deformation defect and other types of defects.

[0129] Step C150: The comparison fusion unit based on the deformation defect diagnosis network calculates the comparison fusion parameter value corresponding to the deformation defect diagnosis label of the deformation image block by using the first comparison parameter value and the second comparison parameter value.

[0130] After obtaining the first comparison parameter value and the second comparison parameter value, the server inputs them into the comparison fusion unit of the deformation defect diagnosis network. This unit fuses the two parameter values ​​into a comparison fusion parameter value through a specific algorithm (such as weighted average, voting mechanism, etc.). This fusion value integrates the information of intra-class similarity and extra-class difference, providing a more comprehensive basis for subsequent deformation defect classification.

[0131] Step C160: obtaining a maximum comparison fusion parameter value based on the deformation defect classification unit of the deformation defect diagnosis network, and determining the degree of deformation defect corresponding to the deformation image block based on the maximum comparison fusion parameter value.

[0132] Next, the server uses the deformation defect classification unit of the deformation defect diagnosis network to obtain the maximum comparison fusion parameter value, and determines the degree of deformation defect corresponding to the deformed image block based on this maximum value. For example, the server can classify the degree of deformation defect into three levels: slight, medium, and severe, and set a threshold range for each level. By comparing the relationship between the maximum comparison fusion parameter value and these thresholds, the server can accurately determine the degree of defect of the deformed image block.

[0133] Step C170, based on the error unit of the deformation defect diagnosis network, an error calculation is performed on the deformation defect degree and the actual deformation defect degree corresponding to the sample steel structure bridge panel image, and based on the calculated deformation defect mask diagnosis error, parameter learning is performed on the deformation defect diagnosis network to generate the deformation defect diagnosis network after parameter learning.

[0134] Finally, in order to verify the accuracy of the deformation defect diagnosis network and further improve its performance, the server needs to perform error calculation and parameter learning. Specifically, the server first obtains a batch of sample steel structure bridge panel images with actual deformation defect degree labels, and inputs these images into the deformation defect diagnosis network for prediction. Then, the server uses the error unit to calculate the error value between the prediction result and the actual label (i.e., the deformation defect diagnosis error). Finally, based on this error value, the deformation defect diagnosis network is parameter learned (such as weight adjustment, increase in the number of layers, etc.) to generate a deformation defect diagnosis network after parameter learning. This process is achieved through back propagation algorithms and other optimization techniques, aiming to make the network more adaptable to complex deformation defect patterns and diverse bridge panel images.

[0135] In a possible implementation, step C130 may include:

[0136] Step C131: Utilize the first comparison unit to generate the number of blocks corresponding to the reference image block.

[0137] The server first starts the first comparison unit in the deformation defect diagnosis network, which first processes the reference image blocks, counts and generates the number of reference image blocks. For example, if a reference image is divided into 100 small blocks, then the number of blocks is 100. This number information is crucial for the subsequent calculation of the relative position and influence of the target deformation defect feature vector in the reference image.

[0138] Step C132: calculating, based on the first comparison unit, a first weight parameter of the target deformation defect feature vector corresponding to the deformation defect diagnosis label.

[0139] Next, the first comparison unit assigns a first weight parameter to the target deformation defect feature vector according to the deformation defect diagnosis label. This weight parameter reflects the importance of the feature vector in diagnosing a specific deformation defect. For example, if a deformation defect feature vector is particularly critical for identifying the deformation defect of "bump", it will be assigned a higher weight value. The determination of the weight parameter is usually based on a large amount of training data and expert knowledge to ensure the accuracy and reliability of the diagnosis.

[0140] Step C133: Calculate the first comparison parameter value corresponding to the target deformation defect feature vector based on the first comparison unit using the number of blocks and the first weight parameter.

[0141] After obtaining the number of blocks and the first weight parameter, the first comparison unit starts to calculate the first comparison parameter value of the target deformation defect feature vector. This calculation process comprehensively considers the distribution of the feature vector in the reference image (reflected by the number of blocks) and its importance in diagnosing specific deformation defects (reflected by the first weight parameter). Specifically, the server can use weighted average, weighted summation or other complex mathematical operations to combine the number of blocks and the first weight parameter, and finally obtain a numerical first comparison parameter value, which will serve as one of the important bases for subsequent deformation defect classification and degree judgment.

[0142] In a possible implementation, step C140 may include:

[0143] Step C141: generating, based on the second comparison unit, the number of labels corresponding to the deformation defect diagnosis label corresponding to the deformation defect diagnosis network.

[0144] In this embodiment, the server starts the second comparison unit in the deformation defect diagnosis network, which first processes the deformation defect diagnosis labels corresponding to the deformation defect diagnosis network, and counts and generates the number of these labels. For example, if the deformation defect diagnosis network can identify five different types of deformation defects, then the number of labels is five, and this label number information is crucial for the subsequent calculation of the correlation and difference between the target deformation defect feature vector and different labels.

[0145] Step C142: calculating, based on the second comparison unit, a second weight parameter of the target deformation defect feature vector corresponding to the deformation defect diagnosis label.

[0146] Next, the second comparison unit assigns a second weight parameter to the target deformation defect feature vector according to the deformation defect diagnosis label. Similar to the first weight parameter, this second weight parameter also reflects the importance of the feature vector in the diagnosis process, but focuses more on measuring the degree of association between the feature vector and different labels. For example, if a deformation defect feature vector plays a key role in distinguishing the two deformation defects of "convex" and "concave", then its second weight parameter under these two labels may be higher. The determination of the second weight parameter is also based on training data and expert knowledge to ensure the accuracy of the diagnosis.

[0147] Step C143: Calculate the second comparison parameter value corresponding to the target deformation defect feature vector based on the second comparison unit using the number of labels and the second weight parameter.

[0148] After obtaining the number of labels and the second weight parameter, the second comparison unit begins to calculate the second comparison parameter value of the target deformation defect feature vector. This calculation process is intended to evaluate the similarity or difference between the feature vector and each deformation defect diagnostic label. Specifically, the server can use weighted distance calculation, similarity matrix operation and other methods to combine the number of labels and the second weight parameter, and finally obtain one or more numerical second comparison parameter values. These values ​​will serve as one of the important bases for subsequent deformation defect classification, degree judgment and new defect identification. In this way, the server can more accurately identify deformation defects in steel structure bridge panel images and assign the most appropriate diagnostic labels to them.

[0149] Figure 2 The steel structure bridge panel detection system 100 based on image recognition shown includes: a processor 1001 and a memory 1003. The processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the steel structure bridge panel detection system 100 based on image recognition may also include a transceiver 1004, which may be used for data interaction between the server and other servers, such as data transmission and / or data reception. It should be noted that the transceiver 1004 is not limited to one in actual scheduling, and the structure of the steel structure bridge panel detection system 100 based on image recognition does not constitute a limitation on the embodiments of the present application.

[0150] Processor 1001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of this application. Processor 1001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0151] The bus 1002 may include a path to transmit information between the above components. The bus 1002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. The bus 1002 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 2 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0152] The memory 1003 may be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), magnetic disk storage medium, other magnetic storage devices, or any other medium that can be used to carry or store program code and can be read by a computer, without limitation herein.

[0153] The memory 1003 is used to store program codes for executing the embodiments of the present application, and the execution is controlled by the processor 1001. The processor 1001 is used to execute the program codes stored in the memory 1003 to implement the steps shown in the above method embodiments.

[0154] An embodiment of the present application provides a computer-readable storage medium having program code stored thereon. When the program code is executed by a processor, the steps and corresponding contents of the aforementioned method embodiment can be implemented.

[0155] It should be understood that, although each operation step is indicated by arrows in the flowchart of the embodiment of the present application, the implementation order of these steps is not limited to the order indicated by the arrows. Unless there is clear explanation in this article, in some implementation scenarios of the embodiment of the present application, the implementation steps in each flowchart can be performed in other orders based on demand. In addition, some or all of the steps in each flowchart may include multiple sub-steps or multiple stages according to the actual implementation scenario, and some or all of these sub-steps or stages may be executed at the same time, and each sub-step or stage in these sub-steps or stages may also be executed at different times respectively. In different scenarios at the execution time, the execution order of these sub-steps or stages can be flexibly configured based on demand, and the embodiment of the present application does not limit this.

[0156] The above is only an optional implementation method for some implementation scenarios of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present application, other similar implementation methods based on the technical ideas of the present application are also within the protection scope of the embodiments of the present application.

Claims

1. A steel structure bridge panel detection method based on image recognition, characterized in that: The method comprises: Obtain a sample steel structure bridge panel image and at least one prior labeled defect corresponding to the sample steel structure bridge panel image, and train a defect diagnosis neural network based on the sample steel structure bridge panel image and the prior labeled defect corresponding to the sample steel structure bridge panel image; wherein the defect diagnosis neural network includes at least one defect diagnosis neural subnetwork, and each of the prior labeled defects corresponds to one of the defect diagnosis neural subnetworks; Based on the requirements of this round of image recognition, corresponding network optimization instances are obtained, and the network structure of the defect diagnosis neural network is optimized; Acquire a candidate steel structure bridge panel image, perform defect diagnosis on the candidate steel structure bridge panel image based on the defect diagnosis neural network after network structure optimization, and generate a defect diagnosis result; The defect diagnosis neural network training based on the sample steel structure bridge panel image and the priori labeled defects corresponding to the sample steel structure bridge panel image includes: Generate corresponding defect representation image blocks from the sample steel structure bridge panel image based on the prior labeled defects, the prior labeled defects at least include fracture defects, corrosion defects and deformation defects, and the defect representation image blocks at least include fracture image blocks corresponding to the fracture defects, corrosion image blocks corresponding to the corrosion defects, and deformation image blocks corresponding to the deformation defects; Perform defect diagnosis learning based on at least one of the fracture image blocks, the corrosion image blocks, and the deformation image blocks to generate the defect diagnosis neural network; The defect diagnosis neural network at least includes a damage state diagnosis network, and defect diagnosis learning is performed based on at least one of the fracture image blocks, the corrosion image blocks, and the deformation image blocks to generate the defect diagnosis neural network, including: The self-attention unit based on the damage state diagnosis network generates a self-attention image block corresponding to the fracture image block from the sample steel structure bridge panel image; Based on the fracture site classification unit of the damage state diagnosis network, the self-attention image block is classified to generate a first actual fracture site corresponding to the fracture image block; The correlation analysis unit based on the damage state diagnosis network performs correlation analysis on the fracture image block and the first actual fracture site to generate a correlation loss parameter corresponding to the fracture image block; The damage state classification unit based on the damage state diagnosis network judges the correlation loss parameter and generates a damage state classification result; Performing parameter learning on the damage state diagnosis network based on the damage state classification result and the loss state label data corresponding to the sample steel structure bridge panel image, and generating the damage state diagnosis network after parameter learning; The correlation analysis unit based on the damage state diagnosis network performs correlation analysis on the fracture image block and the first actual fracture site to generate a correlation loss parameter corresponding to the fracture image block, including: Based on the feature representation branch in the correlation analysis unit, feature representation is performed on the fracture image block to generate a first defect representation vector corresponding to the fracture image block, and feature representation is performed on the first actual fracture site to generate a second defect representation vector corresponding to the first actual fracture site; Retrieving the penetration feature vectors of the first defect characterization vector and the second defect characterization vector based on the penetration feature extraction branch in the correlation analysis unit, generating a first penetration parameter value corresponding to the first defect characterization vector and a second penetration parameter value corresponding to the second defect characterization vector, and then determining the fracture weight coefficient corresponding to the fracture image block and the first actual fracture site based on the first penetration parameter value and the second penetration parameter value; Reconstructing the first actual break site based on the break weight coefficient to generate a second actual break site corresponding to the first actual break site; Based on the fracture feature parameter unit in the correlation analysis unit, the fracture feature parameter of the fracture image block is calculated to generate a first fracture feature parameter corresponding to the fracture image block, the fracture feature parameter of the second actual fracture site is calculated to generate a second fracture feature parameter corresponding to the second actual fracture site, and the fracture feature parameter of the fracture image block and the second actual fracture site is jointly calculated to generate a corresponding third fracture feature parameter; Based on the correlation loss determination branch in the correlation analysis unit, correlation loss is determined for the first fracture feature parameter, the second fracture feature parameter and the third fracture feature parameter to generate the correlation loss parameter corresponding to the fracture image block.

2. The steel structure bridge panel detection method based on image recognition according to claim 1 is characterized in that: Acquiring a candidate steel structure bridge panel image, and performing defect diagnosis on the candidate steel structure bridge panel image based on the defect diagnosis neural network after network structure optimization, and generating a defect diagnosis result, further comprising: Obtain defect diagnosis result retrieval instructions; The construction service terminal that generates the defect diagnosis result retrieval instruction is verified, and for the defect diagnosis result retrieval instruction generated by the verified construction service terminal, the target steel structure bridge panel image and the target defect diagnosis result corresponding to the defect diagnosis result retrieval instruction are determined, and the target steel structure bridge panel image and the target defect diagnosis result are fed back to the construction service terminal.

3. The steel structure bridge panel detection method based on image recognition according to claim 1 is characterized in that: The defect diagnosis neural network further includes a corrosion diagnosis network, which performs defect diagnosis learning based on at least one of the fracture image blocks, the corrosion image blocks, and the deformation image blocks to generate the defect diagnosis neural network, including: The corrosion feature vector positioning unit based on the corrosion diagnosis network performs corrosion feature vector positioning on the corrosion image block to generate the corrosion feature vector corresponding to the corrosion image block; A corrosion effect extraction unit based on the corrosion diagnosis network extracts corrosion effects from the corrosion feature vectors to generate corrosion effect vectors corresponding to the corrosion feature vectors; Generate a predicted corrosion feature category corresponding to the corrosion image block based on the corrosion feature vector and the corrosion action vector; The error unit based on the corrosion diagnosis network performs error calculation on the predicted corrosion feature category and the actual corrosion feature category corresponding to the sample steel structure bridge panel image, and performs parameter learning on the corrosion diagnosis network based on the calculated corrosion diagnosis error to generate the corrosion diagnosis network after parameter learning.

4. The steel structure bridge panel detection method based on image recognition according to claim 1 is characterized in that: The defect diagnosis neural network further includes a deformation defect diagnosis network, which performs defect diagnosis learning based on at least one of the fracture image blocks, the corrosion image blocks, and the deformation image blocks to generate the defect diagnosis neural network, including: Determining a deformation defect diagnosis label corresponding to the deformation defect diagnosis network, and generating a reference image block corresponding to the deformation defect diagnosis label; A deformation defect feature extraction unit based on the deformation defect diagnosis network extracts deformation defect features from the deformation image blocks to generate target deformation defect feature vectors corresponding to the deformation image blocks; Calculating, based on a first comparison unit of the deformation defect diagnosis network, a first comparison parameter value corresponding to the target deformation defect feature vector in the reference image block, wherein the first comparison parameter value is an intra-class metric value; Calculating, based on a second comparison unit of the deformation defect diagnosis network, a second comparison parameter value corresponding to the target deformation defect feature vector in the reference image block, wherein the second comparison parameter value is an out-of-class metric value; The comparison fusion unit based on the deformation defect diagnosis network calculates the comparison fusion parameter value corresponding to the deformation defect diagnosis label of the deformation image block by using the first comparison parameter value and the second comparison parameter value; Obtaining a maximum comparison fusion parameter value based on the deformation defect classification unit of the deformation defect diagnosis network, and determining the degree of deformation defect corresponding to the deformation image block based on the maximum comparison fusion parameter value; The error unit based on the deformation defect diagnosis network performs error calculation on the deformation defect degree and the actual deformation defect degree corresponding to the sample steel structure bridge panel image, and performs parameter learning on the deformation defect diagnosis network based on the calculated deformation defect mask diagnosis error to generate the deformation defect diagnosis network after parameter learning.

5. The steel structure bridge panel detection method based on image recognition according to claim 4 is characterized in that: The first comparison unit based on the deformation defect diagnosis network calculates the first comparison parameter value corresponding to the target deformation defect feature vector in the reference image block, including: Using the first comparison unit to generate the number of blocks corresponding to the reference image block; Calculate, based on the first comparison unit, a first weight parameter corresponding to the deformation defect diagnostic label of the target deformation defect feature vector; The first comparison parameter value corresponding to the target deformation defect feature vector is calculated based on the first comparison unit using the number of blocks and the first weight parameter.

6. The steel structure bridge panel detection method based on image recognition according to claim 4 is characterized in that: The second comparison unit based on the deformation defect diagnosis network calculates the second comparison parameter value corresponding to the target deformation defect feature vector in the reference image block, including: Generating the number of labels corresponding to the deformation defect diagnosis label corresponding to the deformation defect diagnosis network based on the second comparison unit; Calculate, based on the second comparison unit, a second weight parameter corresponding to the deformation defect diagnostic label of the target deformation defect feature vector; The second comparison parameter value corresponding to the target deformation defect feature vector is calculated based on the second comparison unit using the number of labels and the second weight parameter.

7. A steel structure bridge panel detection system based on image recognition, characterized in that: It includes a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by the processor, the steel structure bridge panel detection method based on image recognition as described in any one of claims 1 to 6 is implemented.

Images