Steel box girder bottom damage identification and synchronous positioning method based on distributed camera network

Abstract

The application discloses a steel box girder bottom damage identification and synchronous positioning method based on a distributed camera network. The method comprises the following steps: S1, a series of imaging device units are arranged in a spatially parallel structure, and the imaging device units are connected through Bluetooth and a wireless local area network to collect image data of any position on the steel box girder bottom; S2, the steel box girder bottom area is divided horizontally according to the shooting coverage range of each imaging device unit; and S3, two-stage synchronous identification and positioning based on deep learning driving are carried out, in the first stage, MobilenetV4 is used as a key feature extraction tool to perform lightweight reconstruction on panoramic images, and disease area identification and positioning are performed at a global level; and in the second stage, a YOLOv9 target detection framework is used to analyze the disease area and provide disease information at a local level. The design of the application can be widely applied to large-span bridge bottom inspection and efficient analysis of diseases.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of structural health detection in civil engineering, more specifically, it relates to a software and hardware system combining advanced equipment and intelligent algorithms, aiming to realize accurate identification and rapid positioning of beam bottom diseases. BACKGROUND

[0002] Bridges play a vital role in the modern infrastructure system, serving as key nodes in the transportation network and core support structures for economic development and resource scheduling. During service, they face multiple problems such as structural fatigue, material aging, environmental erosion, and overloading stress, which affect service life and safety performance. To effectively address these challenges, regular health condition detection of bridges has become an essential measure. However, with the advancement of technology and the continuous improvement of detection standards, traditional manual detection methods gradually fail to meet the efficiency and accuracy requirements of modern detection. Especially in the detection scenario of the bridge bottom, due to its hidden location and complex environment, the detection difficulty increases significantly. Therefore, the bridge bottom detection task urgently needs to introduce more advanced technology.

[0003] In recent years, many scholars have developed customized detection equipment to address the bottleneck problems of the above detection scenarios. These intelligent detection devices are favored for their portability and flexibility, and can quickly reach the detection area, providing convenience for bridge detection. However, a core challenge is that in complex beam bottom environments, most intelligent detection devices still rely on manual operation. This dependence not only limits the degree of automation of the detection process, but also greatly affects the ability and experience of the technicians, leading to the inability to ensure the comprehensiveness and accuracy of the results. In addition, the endurance of the equipment is limited, which directly reduces the efficiency of the detection task. Mobile vehicle-based detection systems achieve image acquisition of large areas of the bridge. Relying on such detection platforms can reduce the dependence on manual control and achieve stable and comprehensive detection of bridge bottom diseases. However, they lack portability and flexibility in practical applications, and have certain requirements for the working environment, such as occupying emergency lanes. In particular, in the detection of large-span bridges, there are blind areas that are difficult to cover.

[0004] Due to differences in personal experience, different technicians may draw different conclusions when analyzing the same detection data. This subjective judgment-based analysis method not only introduces uncertainty into the results, but also affects the efficiency and reliability of the detection. Deep learning technology, as an innovative driving force in various fields, has revolutionized traditional data processing methods. Deep learning models, through the construction of multi-layer neural networks, can automatically extract and learn complex features in data, simulating the structure of the human brain.

[0005] At present, the research on bridge bottom detection technology mainly focuses on small and medium-sized bridges, and the existing equipment mostly relies on manual operation, and has not realized automation. In the face of the detection needs of large-span bridges, especially the areas that are difficult to reach at the bottom, comprehensive and efficient data collection is particularly difficult, which constitutes a major technical challenge. SUMMARY

[0006] In view of the above technical problems, the present application provides a steel box girder bottom damage identification and synchronous positioning method based on a distributed camera network, which breaks through the obstacle of automatic analysis of massive data and realizes efficient processing of data. Through the innovative two-stage analysis method, efficient and intuitive guidance of the disease area is realized at the macro level, and the detailed development of the disease is accurately captured at the micro level. This analysis means optimally integrates between the global and local levels, providing a new solution for comprehensive detection of the bridge bottom.

[0007] In order to realize the above technical purpose, the present application adopts the following technical scheme:

[0008] A steel box girder bottom damage identification and synchronous positioning method based on a distributed camera network, comprising the following steps:

[0009] S1, a series of imaging device units are composed according to a spatially parallel structure, and the imaging device units are connected together through Bluetooth and wireless local area network for image data collection of any position of the steel box girder bottom;

[0010] S2, according to the shooting coverage range of each imaging device unit, the steel box girder bottom area is horizontally divided, the steel box girder bottom area is divided into a plurality of sub-areas, the number corresponds to the imaging device unit, and each sub-area is assigned a unique position code;

[0011] The steel box girder bottom area is divided into blocks in the longitudinal direction, each sub-area structure segment is defined as an independent block, the position code of the original sub-area is introduced to create the position code of the block, and the correlation between the images in each block range of the steel box girder bottom space is constructed according to the defined block concept, to realize the generation of block images; finally, the local level block images are integrated according to the original spatial order by using the position code of each block, to form a complete steel box girder bottom panoramic image;

[0012] S3, a two-stage synchronous identification and positioning method based on deep learning driving, wherein,

[0013] The first stage adopts a recognition tool with a MobilenetV4 architecture to quickly screen out key features related to diseases, and introduces a fusion method based on a diffusion model to improve the generalization ability of the recognition tool; then, according to the identified disease features, a lightweight reconstruction is performed on the panoramic image of the steel box girder bottom, and then a feature fine-grained parameter is introduced to adjust the information richness of the lightweight panoramic image; finally, a position-related digital code is marked on each region of the lightweight panoramic image to identify and locate the disease region of the steel box girder bottom.

[0014] In the second stage, the steel box girder bottom disease region identified in the first stage is analyzed in detail in terms of disease number, type and local position by using a target detection tool.

[0015] In step S2, the block image generation specifically includes the following sub-steps:

[0016] S21, divide the image into a feature extraction region and a reserved region along the longitudinal direction; the feature extraction region is further divided into left and right staple regions in the transverse direction, and in the matching process, the left and right staple regions of the adjacent time images are fixed by analogy to using a stapler, that is, the corresponding relationship of the feature points is found respectively;

[0017] S22, a homography matrix calculation is performed to obtain a stable conversion relationship matrix, and the image sequence contained in the block image is grouped, and each two adjacent images are taken as a group for splicing at the same time;

[0018] S23, repeat steps S21 to S22 until the generation of the block image is completed.

[0019] In step S21, the division ratio between the feature extraction region and the reserved region is 0.5.

[0020] It also includes step S24, which decomposes the image into multiple levels according to the frequency by using a Gaussian filter and sampling operation, then performs smoothing, weighting and superposition processing on the high and low frequency components of each level, finally, adds each frequency band component to realize effective elimination of edge discontinuity.

[0021] In the first stage of the two-stage synchronous recognition and positioning method based on deep learning, in the model training stage, a diffusion model is introduced to simulate and expand the real disease scene, thereby improving the generalization ability of the model.

[0022] In the fusion method based on the diffusion model, the generated disease image is rotated, scaled, color and brightness adjusted by using the Poisson fusion method before the fusion operation;

[0023] Next, the fused region in the target image is randomly selected, and the gradient information of the region is calculated to obtain the trend of image detail changes;

[0024] Finally, the boundary features of the source image and the gradient information of the target image are used as constraint conditions by using the Poisson equation.

[0025] Step S3 is based on the first stage of the two-stage synchronous recognition and positioning method driven by deep learning:

[0026] Using the macro information of the global level as a guiding opinion, the disease area in the lightweight panoramic image is traced back, first, the disease area is identified in the panoramic image through macro analysis, and then the specific position of these areas is accurately positioned;

[0027] Next, these areas are extracted from the lightweight panoramic image and restored to more detailed image-based block representations; thus, without losing the original details, the disease area is finely analyzed;

[0028] Finally, the extracted detailed features are mapped with the global information through the recorded position encoding mapping, ensuring the consistency and comprehensiveness of the disease analysis.

[0029] In step S3, the second stage of the two-stage synchronous recognition and positioning method driven by deep learning, the YOLOv9 target detection framework is used to accurately analyze the identified disease area of the steel box girder bottom.

[0030] The beneficial effects of the present application are:

[0031] In view of the complexity and challenge of the detection of the bottom area of the large-span bridge, the present application proposes and realizes a software and hardware system combining advanced equipment and intelligent algorithms. The proposed method has shown significant efficiency and accuracy in the detection of the bottom area of the large-span bridge. Through automatic data acquisition and accurate disease identification, the system effectively optimizes the detection process. Detailed disease information supports engineers to develop accurate targeted maintenance strategies, and improves the health management level of the bridge. The innovative solution of the present application provides an efficient global perspective for the detection of the bottom area of the large-span bridge, significantly improving the detection efficiency and accuracy.

[0032] 1. Dynamic visual perception device: The developed device successfully breaks through the limitations of traditional control, acquisition and transmission links, allowing technical experts to operate through a remote interactive interface, and automatically completing the comprehensive data acquisition of the girder bottom area. This design significantly improves the efficiency and comprehensiveness of data acquisition, reducing the need for on-site manual intervention.

[0033] 2. Panorama image generation and optimization: By introducing the concept of blocks, the stitching process of panorama images is simplified, and the accuracy of image processing is improved. Spatially ordered partitioning and extraction of overlapping area feature points effectively reduce the processing complexity and optimize the image alignment process. Although panorama images may have complexity problems in information presentation, this method provides the advantage of global perspective.

[0034] 3. Deep learning driven two-stage recognition and positioning: First stage: using the recognition tool with MobilenetV4 architecture to quickly filter out key features. In the model training stage, a diffusion model is introduced to simulate and expand the real disease scene, thereby improving the generalization ability of the model. By using the MobilenetV4 architecture for rapid screening of key features, combined with the diffusion model to simulate and expand the disease scene, the generalization ability of the model is improved. Through lightweight reconstruction and feature fine-grained adjustment, the recognition and positioning of disease areas in panorama images are significantly optimized. Second stage: using the YOLOv9 target detection framework to accurately analyze the recognized disease areas, realizing the detailed information recognition of disease areas. The fine processing of this stage further enhances the accuracy of disease analysis. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The hardware module structure diagram of the steel box girder bottom damage recognition and synchronous positioning method based on a distributed camera network according to the present application;

[0036] Figure 2 The schematic diagram of a series of imaging equipment units according to the present application in a spatially parallel structure;

[0037] Figure 3 The panorama generation process diagram based on the block concept according to the present application;

[0038] Figure 4 The diffusion model and fusion technology according to the present application generate real damage scenes;

[0039] Figure 5 The two-phase method schematic diagram of the steel box girder bottom damage synchronous recognition and positioning according to the present application;

[0040] Figure 6 The system structure schematic diagram of the feature description tool according to the present application;

[0041] Figure 7 The YOLOv9 model framework schematic diagram according to the present application;

[0042] Figure 8 The structure diagram of the generalized efficient layer aggregation network according to the present application;

[0043] Figure 9 The bridge specification and equipment arrangement diagram according to the present application;

[0044] Figure 10 application of dynamic perception in bottom panoramic imaging;

[0045] Figure 11 detailed information display of the proposed block image;

[0046] Figure 12 proposed lightweight reconstructed panoramic image;

[0047] Figure 13 proposed damage analysis result display. DETAILED DESCRIPTION

[0048] The present application will be further clarified by the following examples and drawings, which should not be taken as limiting the scope of the present application. The following examples are presented by way of illustration of the present application and not by way of limitation.

[0049] The present application will be further clarified by the following examples and drawings, which should not be taken as limiting the scope of the present application. The following examples are presented by way of illustration of the present application and not by way of limitation.

[0050] The method framework diagram is shown in Figure 1 From the system level, the method can be divided into three parts:

[0051] S1, dynamic visual perception device. This device breaks through the boundaries of control, collection and transmission, allowing technical experts to remotely operate through an interactive interface, thereby automatically completing the comprehensive data collection of the beam bottom without the need for on-site manual intervention.

[0052] S2, block-based panoramic generation strategy. Through ordered division, the relevance between images is constructed, and block images are independently and parallelly processed. In the feature extraction and matching stage, the principle of staples is used to simplify the image alignment process and reduce the processing complexity by concentrating on the extraction of feature points in the overlapping area.

[0053] S3, two-stage synchronous recognition and positioning method driven by deep learning. In the first stage, the recognition tool with MobilenetV4 architecture is used to quickly filter out key features. In the model training stage, a diffusion model is introduced to simulate and expand the real disease scene, thereby improving the generalization ability of the model. Then, according to the identified key features, the research team performs lightweight reconstruction on the panoramic image. At the same time, feature fine-grained adjustment parameters and position encoding technology are introduced to optimize the representation form of global information and improve the accuracy of disease area recognition and positioning. In the second stage, the focus shifts to in-depth analysis of the disease areas identified in the first stage. Through the position encoding technology, it can quickly trace back to the relevant original block image. Further, an advanced target detection framework YOLOv9 is introduced to accurately identify the detailed information of the disease area.

[0054] The main components of the dynamic visual perception device in S1 are:

[0055] The device is mainly composed of a perception module, a control transmission module and a motion module, as shown in Figure 2 The perception module uses a distributed camera to replace the visual system of a technician, and is connected together through Bluetooth and a wireless local area network.

[0056] The edge computer is applied to the control transmission module, and a technician can easily set and adjust various parameters in the collection stage, including time, collection mode, focal length and the like, and start and stop collection actions.

[0057] The beam bottom maintenance vehicle serves as the motion module, and gives the distributed camera the ability to collect data at any position on the beam bottom.

[0058] The specific content of the block-based panoramic generation strategy in S2 is:

[0059] Traditional methods have problems such as large resource consumption, high time cost and strong dependence on manual work when processing large-scale image data. In addition, these methods are limited to local analysis of a single image and cannot consider the correlation between images, making it difficult to effectively analyze the comprehensive state information. In order to overcome these challenges, the present application is based on the concept of panoramic image generation and correlates discrete image data. Then, the beam bottom panoramic image is reorganized by taking blocks as the construction unit.

[0060] The traditional panoramic image construction method, i.e. splicing discrete images into a complete image containing all the contents, has obvious limitations for the complex structure of large-span bridges, especially the bridge bottom. Therefore, according to the coverage range of each unit in the dynamic perception chain, the beam bottom area is divided horizontally. This process aims to decompose the beam bottom area into a number of sub-areas corresponding to the perception units. At the same time, a unique position code is assigned to each sub-area. Further, the beam bottom area is finely divided in the longitudinal direction. Based on the characteristics of the structure design, each structure segment within a sub-area is defined as an independent block, thereby realizing the ordered discretization of the entire beam bottom space. Based on the position code of the original sub-area, the structure segment information is introduced to create the position code of the block. According to the above defined block concept, the correlation between images in each block range of the beam bottom space is constructed, and the block image is generated. Finally, the local hierarchical block image is integrated according to the original spatial order using the position code of each block, forming a complete beam bottom panoramic image. The introduction of the block concept allows each block to be processed independently and simultaneously, as shown in Figure 3 At the same time, each block is assigned a unique code for quick positioning of any area in the global space.

[0061] The block image generation firstly divides the image into feature extraction region and reserved region along the longitudinal direction. Considering the maximum speed of the motion module and the overlap degree of the adjacent time images, the division ratio between the two regions is set to 0.5. The feature extraction region is further divided into left and right staple regions in the transverse direction, and parallel computing is used. In the matching process, the left and right staple regions of the adjacent time images are fixed, that is, the corresponding relationship of the feature points is found. Finally, the stable conversion relationship matrix is obtained by using the homography matrix calculation. Under the condition of sufficient computing resources, the image sequence contained in the block image can be grouped, and each two adjacent images are taken as a group for splicing at the same time. This grouping strategy enables multiple pairs of images to be spliced at the same time, thereby effectively improving the overall processing efficiency. Then, the cycle operation is carried out according to this strategy until the generation of the block image is completed.

[0062] The initial block image may present obvious boundaries or transition regions in the overlapping part of adjacent time. In order to solve this problem, the invention adopts the method of pyramid unfolding, which decomposes the image into multiple levels according to the frequency from high to low through the Gaussian filter and sampling operation, and then smoothes and weights each level of high and low frequency components and superimposes them. Finally, the frequency band components are added again to effectively eliminate the edge discontinuity.

[0063] The specific content of the depth learning driven two-stage synchronous identification and positioning method in S3 is as follows:

[0064] The invention adopts a diffusion model with high image generation quality and easy training as a disease synthesis tool for the main disease characteristics of the bottom of the steel box girder, i.e. coating peeling and corrosion. The invention does not directly train a disease in a large scene, but focuses on the local disease area. Resources are concentrated on the synthesis of disease data to ensure that the main content of the synthesized image is concentrated on the information of the coating deterioration area. Then, the synthesized disease information is mapped and fused into the real detection scene by methods such as Poisson fusion, as shown in Figure 4 .

[0065] The diffusion model mainly consists of two key stages: forward diffusion and reverse denoising, as shown in Figure 4 . Among them, the forward diffusion stage is a Markov chain process with typical characteristics. In this process, Gaussian noise plays a key role in destroying the input sampled image ( is the data distribution of a large number of real images) in a continuous manner. The orderliness of the input image gradually disappears, and gradually transforms into a series of noise distribution images , where each state transition represents a stage of advancement, from Z1 to Z2, and then to subsequent stages, until the final formation variance of the Gaussian noise distribution where denotes the mean coefficient. The state value at each time instant t is only related to the previous time instant . And as increases, the distribution state of tends to be closer to pure noise, i.e.

[0066] In the reverse denoising stage, the original image distribution is recovered The neural network architecture with parameters theta is used to carry out the prediction of the reverse distribution . In this way, with the powerful learning and fitting ability of the neural network, the effective prediction and estimation of the reverse distribution are approximated and realized as much as possible.

[0067] The neural network needs to predict the mean coefficient and variance of the reverse distribution . In the training process of the network, the variance value of is explicitly specified as only needs to be learned and predicted . There is a direct correlation between it and the noise distribution epsilon. Therefore, the neural network is a tool specially used for noise prediction, and its purpose is to make the output noise as close to the normal distribution state as possible. By training the noise predictor, the model tries to find an optimal way to approximate the normal distribution in continuous learning and adjustment. After training is completed, a noise sample can be obtained by randomly sampling from the standard normal distribution. Then, the neural network involved in the reverse process is used to restore and reshape the noise sample into a complete image based on the noise sample.

[0068] In further research, the present application adopts advanced Poisson fusion technology. Before the fusion operation, the generated disease image needs to be rotated, scaled, color and brightness adjusted. These enhanced samples have different perspectives, sizes and lighting conditions, providing a richer data basis for the simulation process. Next, the fusion area is randomly selected in the target image, and the gradient information of the area is calculated to obtain the trend of image detail change. Finally, a mathematical model is constructed based on the Poisson equation, and the boundary features of the source image and the gradient information of the target image are taken as constraint conditions.

[0069] Figure 5 The present application proposes an innovative two-stage disease synchronous identification and positioning method, such asThe first stage is to optimize the reconstruction of panoramic images through a deep learning-based feature representation method. The second stage further learns the detailed disease information based on the learned features of the disease area in the first stage, and provides a complete bridge service status: disease number, type and specific location.

[0070] The block image contains a large amount of non-key background information, which has no substantial contribution to operation and maintenance analysis. To solve this problem, the present application introduces the MobilenetV4-S network architecture to preliminarily analyze the block image of the beam bottom area. The superior performance of MobilenetV4-S is due to the special module design, i.e. the universal inverted bottleneck module design Universal Inverted Bottleneck (UIB), as shown in Figure 6 A simple digital coding system is used to represent these features. In this system, '0' coding indicates that no disease is detected in the block image, implying that the block is in a normal service state; on the contrary, '1' coding indicates that disease is found in the block image, implying that the block may be in an abnormal service state. The present application expands the visualization properties of the lightweight panoramic image by introducing color labels. Specifically, the 0 code is assigned a red label, and the 1 code is assigned a green label. In order to further highlight the disease area and provide spatial information, the present application also proposes a multi-dimensional feature representation method. This method is an extension of the original 0-1 coding. In the two-dimensional plane, '0' represents that no disease is detected in the block image, and '1' indicates that there is disease. In the three-dimensional space, this information is copied, and each block is assigned an additional attribute value in the third dimension perpendicular to the two-dimensional plane. For the disease area, the value of the third dimension can be set to 1, and for the normal area, the value of this dimension remains 0.

[0071] The present application introduces a feature granularity adjustment parameter to optimize the disease feature representation in the global reconstruction stage. Combined with the characteristics of most bridge engineering structures, the present application focuses on the scene with high length-width ratio, and plays a role in the longitudinal direction of the block image, which uniformly subdivides the image into multiple sub-blocks. Each sub-block is extracted more finely. In addition, the third section position coding is improved in combination with the fine-grained parameter to form a new coding system. In the block refinement process, each sub-block is coded with clear position information.

[0072] To solve the challenge of effectively combining global and local information, the present application uses macro information at the global level as a guiding opinion to trace the source of the disease area in the lightweight panoramic map. First, the disease area is identified in the panoramic map through macro analysis, and then the specific location of these areas is located. Next, these areas are extracted from the lightweight panoramic map and restored to more detailed image-based block representations. Without losing the original details, the disease area is analyzed in detail. Finally, the extracted detailed features are mapped with the global information through the recorded location encoding mapping.

[0073] The fine detection link uses an advanced target detection framework YOLOv9 to establish a mapping relationship between network output and input data, and the model structure is as shown in Figure 7 .

[0074] Embodiments

[0075] The Nansha Bridge project consists of 7 approach bridges, 3 interchanges and 2 large-span suspension bridges, each with a length of more than 1 km. The focus of this study is the Dashahai Strait Bridge, which has a double-tower, single-span suspension design spanning 360+1200+480m. The main span uses a steel box girder with a total width of 49.7m. In the longitudinal direction, the 3rd to 84th sections are selected for testing, each section being 12.8m long, for a total length of 1049.6m. The transverse inspection selects plates 5-11 as the target. Figure 9 The overall layout of the bridge is shown in

[0076] In the dynamic perception system of the present application, the arrangement strategy of the perception module is carefully designed, as shown in Figure 10 . The perception sub-module is installed on the outer side of the platform through a customized peripheral support. In addition, considering the influence of extreme weather conditions such as typhoons and long-term vibration on the stability of the system, the design of the peripheral support takes targeted measures to firmly fix it on the inspection vehicle using high-strength bolt fasteners. The arrangement position of each perception module is accurately calculated to be located directly below the center of the plate, with a field of view range of about 3.5 meters, which ensures that a single module can fully meet the detection needs of the plate, as shown in Figure 10 . In terms of computing control modules, a dual-module configuration is used, with one controlling 3 cameras and the other controlling 4 cameras. This configuration is used to evaluate the stability of the control module in multi-task processing. Before the brake module is started, the acquisition instruction is sent to each perception unit synchronously through the software interface in the edge calculator. After the acquisition task is completed, the control software is responsible for safely transmitting the data in the perception module to the central control system through wireless transmission. Subsequently, the data is stably transmitted to the remote server through the intranet, and the server's computing resources are used for in-depth data analysis and processing.

[0077] According to the panoramic generation strategy based on the block concept set forth in the present application, meticulous integration processing is performed on the collected discrete data of the beam bottom. The physical size of a single plate in a standard beam section is 3 meters wide and 12.8 meters long. Through innovative splicing technology and combined with the correction method of homography matrix, the resolution of the output image is 3000x12800 pixels, and each pixel point corresponds to an actual physical size of 1 millimeter. In order to verify the accuracy of the block image, the test steps are carefully designed. Specifically, first, a variety of red rectangular labels are artificially made, and the sizes of these labels are accurately measured and recorded as reference standards for subsequent verification. Subsequently, these labels are accurately attached to the artificially designed test points. The position information of these points is determined in advance through manual measurement. After the image acquisition task is completed, the research team analyzes the collected block images and compares the red rectangular labels in the images with the previously recorded size and position information. The conclusion shows that the size measurement accuracy of the entire system can be controlled within 2 mm, and the positioning accuracy error can be controlled within 3 mm, which proves the reliability and effectiveness of the image acquisition and processing method in practical application. The specific comparison results are shown in Figure 11 .

[0078] The panoramic generation stage generates a total of 574 independent block images, each of which corresponds to a part of the actual beam section, and the total area covers a beam section length of 1,049.6 meters. The present application adopts a positioning system based on block mode. These block images can be simply combined to reconstruct a complete beam bottom panoramic image according to the established position coding rules, as shown in Figure 11 (only part of the content is displayed). The strategy of simply combining block images into a panoramic image is not feasible in practice. To address this challenge, the present application adopts the proposed two-stage disease synchronous identification and positioning method to divide the problem into global and local two levels. In the global level, the representation of main features is focused on, avoiding the complexity of showing all detailed information in the global range. Then focus on important local disease areas and use high-precision processing means to analyze these details.

[0079] In the first stage, the proposed feature representation method is used to perform in-depth feature analysis on 574 plate images. The total pixels of the panoramic image after light reconstruction are reduced by 99.99% compared to the image-based representation form in terms of size. Due to the service life of the bridge and the timeliness of steel box bridge maintenance, there are certain limitations in verifying the effectiveness of the method proposed in the present application.

[0080] Figure 12The maintenance of the steel box girder bottom diseases is demonstrated. To overcome these limitations, the present application designs a pre-experiment of simulating diseases based on a synthetic database (covering real and synthetic disease patterns). First, a complete girder bottom dataset is established using real scene data with fine-grained processing. The dataset contains 7 transverse target areas, each of which is further divided into 492 sub-blocks, totaling 3444 sub-block images. To simulate the random distribution of real diseases, 40 digital encodings of disease areas are randomly generated in each sub-area, which correspond to the position encoding in the block coordinates. Then, disease pattern data are randomly extracted from the synthetic database, with the extraction frequency set to 280 times, and ensuring that the extracted data are not reused in subsequent operations. The extracted disease pattern data are fused into the sub-block images corresponding to the position encoding. The entire process adopts an end-to-end processing mode, with the input being 3444 sub-block images and the output also being 3444 sub-block images containing disease areas, and the processing process maintains the "black box" feature. The 3444 sub-block images containing disease areas are input into the feature characterization tool of the first stage for preliminary analysis, and a total of 277 disease areas are identified, but 3 areas are not identified. After analyzing these missed areas, it is found that their disease size is small and the pattern is not obvious. Therefore, in the future, it is necessary to further improve the ability of the disease feature characterization tool in small target identification. According to the disease feature description results (including the 3 missed areas added manually) and the position encoding, a lightweight panoramic map is generated, as shown in Figure 12 In the block coordinate system, each sub-block is not expanded along the longitudinal segment coordinate axis, but is added in the plate block coordinate axis direction to form a sub-block coordinate axis.

[0081] The second stage automatically extracts the position encoding of the disease area and maps the feature description back to the original sub-block image. Using the detection tool, the disease category, number and position of all sub-block images are identified in detail, and the specific results are shown in Figure 13 In the figure, the first column shows the position encoding of the sub-block image, the second column shows the disease detection effect of the corresponding area, and the results prove that the peeling and rust diseases can be accurately identified. The third column records the coordinate information of the disease in the block image in detail. In order to more accurately locate the position of the disease, a coordinate system is established with the upper left corner of each block image as the origin. Combining the local positioning coordinates in the sub-block with the S encoding information of the sub-block image, it can be converted into global block coordinates. The present application applies a heat map to the estimation of the disease affected area to provide results that are more consistent with the disease pattern. The last column in the figure shows the visualization results of the disease affected area estimation by heat map.

[0082] In summary, the specific embodiments verify the effectiveness and applicability of the scheme proposed by the present application to complex engineering.

[0083] The above disclosed is only a typical embodiment of the present application, but the present application is not limited to this, any homogenous modification to the present application by any person skilled in the art after reading the patent should fall into the protection scope of the present application.

Claims

1. A steel box girder bottom damage identification and simultaneous localization method based on a distributed camera network, characterized in that, It comprises the following steps: S1, a series of imaging device units are arranged in a spatially parallel structure, and the imaging device units are connected through Bluetooth and wireless local area network to collect image data at any position of the steel box girder bottom; S2, according to the shooting coverage of each imaging device unit, the steel box girder bottom area is divided horizontally into a plurality of sub-areas corresponding to the imaging device units, and each sub-area is assigned a unique position code; The steel box girder bottom area is divided into blocks in the longitudinal direction, and each sub-area is defined as an independent block. Based on the position code of the original sub-area, the position code of the block is created by introducing the structure segment information. According to the concept of the defined block, the correlation between images in each block range of the steel box girder bottom space is constructed to realize the generation of block images. Finally, the local hierarchical block images are integrated in the original spatial order according to the position code of each block to form a complete steel box girder bottom panoramic image; S3, based on a two-stage synchronous recognition and positioning method driven by deep learning, wherein In the first stage, the recognition tool with MobilenetV4 architecture is used to quickly filter out the key features related to diseases, and a fusion method based on diffusion model is introduced to improve the generalization ability of the recognition tool. Then, according to the identified disease features, the steel box girder bottom panoramic image is reconstructed in a lightweight manner, and then the feature fine-grained parameter adjustment is introduced to adjust the information richness of the lightweight panoramic image. Finally, the position-related digital code is marked on each region of the lightweight panoramic image to identify and locate the disease area of the steel box girder bottom. In the second stage, the target detection tool is used to analyze the number, type and local position of the diseases in the steel box girder bottom area identified in the first stage.

2. The steel box girder bottom damage identification and simultaneous localization method based on a distributed camera network according to claim 1, characterized in that, In step S2, the block image generation specifically includes the following sub-steps: S21, divide the image into a feature extraction area and a reserved area along the longitudinal direction; the feature extraction area is further divided into left and right staple areas in the horizontal direction, and in the matching process, the left and right staple areas of adjacent images are fixed by analogy to using a stapler, i.e. finding the corresponding relationship of feature points for them respectively, by using parallel computing; S22, use homography matrix calculation to obtain a stable conversion relationship matrix, group the image sequence contained in the block image, and splice each two adjacent images as a group; S23, repeat steps S21 to S22 until the generation of the block image is completed.

3. The steel box girder bottom damage identification and simultaneous localization method based on a distributed camera network according to claim 2, characterized in that, In step S21, the division ratio between the feature extraction area and the reserved area is 0.

5.

4. The steel box girder bottom damage identification and simultaneous localization method based on a distributed camera network according to claim 2, characterized in that, Step S24 further includes the steps of decomposing the image into multiple levels according to the frequency by using a Gaussian filter and sampling operation, then smoothing and weighting the high and low frequency components of each level and superimposing them, finally adding the frequency band components to realize effective elimination of edge discontinuity.

5. The method for identifying and synchronously locating bottom damage of a steel box girder based on a distributed camera network according to claim 1, characterized in that: In step S3, in the first stage of the two-stage synchronous identification and positioning method driven by deep learning, in the model training stage, a diffusion model is introduced to simulate and expand the real disease scene, thereby improving the generalization ability of the model.

6. The steel box girder bottom damage identification and simultaneous localization method based on a distributed camera network according to claim 2, characterized in that, In step S3, in the fusion method based on the diffusion model, before using the Poisson fusion method, the generated disease image is rotated, scaled, color and brightness adjusted; Next, select the fused region in the target image at will, and calculate the gradient information of the region to obtain the trend of image detail change; Finally, the boundary features of the source image and the gradient information of the target image are used as constraint conditions by using the Poisson equation.

7. The steel box girder bottom damage identification and simultaneous localization method based on a distributed camera network according to claim 1, characterized in that, In step S3, in the first stage of the two-stage synchronous identification and positioning method driven by deep learning: Use the macro information of the global level as a guiding opinion to trace the disease area in the lightweight panoramic image. First, identify the disease area in the panoramic image through macro analysis, and then accurately locate the specific position of these areas. Next, extract these areas from the lightweight panoramic image and restore them to more detailed image-based block representations; thereby performing fine analysis on the disease area without losing the original details; Finally, map the extracted detailed features to the global information through the recorded position encoding mapping to ensure the consistency and comprehensiveness of the disease analysis.

8. The steel box girder bottom damage identification and simultaneous localization method based on a distributed camera network according to claim 1, characterized in that, In step S3, in the second stage of the two-stage synchronous identification and positioning method driven by deep learning, the YOLOv9 target detection framework is used to accurately analyze the identified disease area of the steel box girder bottom.

Images