An image recognition-based prefabricated retaining wall defect identification method and system

Abstract

The application relates to the field of soil wall defect identification, and discloses an assembled retaining wall defect identification method and system based on image recognition, which comprises the following steps: acquiring multi-view image data of the assembled retaining wall, combining a boundary detail strengthening mechanism and a regional light compensation strategy, and constructing a defect identification image data set; performing edge-guided feature extraction on the defect identification image data set, constructing an image deep feature model based on a boundary context fusion network; judging whether the response intensity change of the image deep feature model in a crack region reaches a preset threshold value through an edge response enhancement mechanism; if yes, labeling that the crack region has potential defects; using a multi-scale morphological structure analysis method to perform contrast perception optimization on the corrected crack region; and based on the defect identification result, combining a component positioning mechanism and component historical operation and maintenance data to perform severity grading evaluation on the defects. The application has the advantage of improving the sensitivity to edge regions.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of soil wall defect identification, in particular to an assembled retaining wall defect identification method and system based on image recognition. BACKGROUND

[0002] In the existing assembled retaining wall defect identification technology, the image recognition algorithm focuses on identifying large-area cracks, obvious deformation and other significant defects, but there are obvious deficiencies in processing small displacement cracks caused by construction errors or material aging. When the defect is at the edge of the assembled retaining wall joint, the light contrast is small or the background texture is complex, the existing algorithm is easy to misjudge these small cracks as texture noise, resulting in missed detection. This problem is particularly critical in early structural anomaly detection, because small cracks are often a precursor to structural disease development, and if they cannot be accurately identified in a timely manner, it will pose a risk to subsequent operation and safety assessment. Therefore, it is necessary to design an assembled retaining wall defect identification method and system based on image recognition which improves the sensitivity to edge regions. SUMMARY

[0003] In view of the deficiencies of the prior art, the present application provides an assembled retaining wall defect identification method and system based on image recognition, which has the advantage of improving the sensitivity to edge regions and solving the problems in the above background art.

[0004] To achieve the above-mentioned purpose of improving the sensitivity to edge regions, the present application provides the following technical scheme: an assembled retaining wall defect identification method based on image recognition, comprising the following steps:

[0005] Obtain multi-view image data of the assembled retaining wall, and combine a boundary detail enhancement mechanism and a regional illumination compensation strategy to construct a defect identification image data set;

[0006] Perform edge-guided feature extraction on the defect identification image data set, construct an image deep feature model based on a boundary context fusion network, and introduce an asymmetric feature aggregation structure to improve the perception ability of the model to small edge defects;

[0007] Based on an edge response enhancement mechanism, determine whether the response intensity change of the image deep feature model in the crack region reaches a preset threshold value, if so, label that there is a potential defect in the edge region, and combine a crack and background contrast adjustment mechanism to dynamically correct the preliminary labeling result;

[0008] According to the corrected edge defect region, apply a multi-scale morphological structure analysis method to the corrected result for contrast perception optimization processing, extract the geometric parameters of the crack region, and dynamically generate a defect identification result;

[0009] Based on the defect identification result, combined with the component positioning mechanism and the component historical operation and maintenance data, the severity of the defect is classified and evaluated, and a defect analysis report is generated.

[0010] Preferably, the process of constructing the defect recognition image dataset is:

[0011] Collecting multi-view image samples under different time, angle of view and light conditions;

[0012] Performing brightness distribution analysis on the multi-view image samples, dynamically evaluating the local brightness features of the multi-view images for the uneven lighting problem, and adaptively determining the regional light compensation window;

[0013] On the basis of light compensation, a boundary detail enhancement mechanism is used to enhance the crack and edge details in the multi-view images, while a variety of data enhancement methods are fused;

[0014] The pre-processed and enhanced multi-view images are uniformly standardized, and a standardized defect recognition image dataset is constructed.

[0015] Preferably, the process of constructing the image deep feature model based on the boundary context fusion network is:

[0016] Input the defect recognition image dataset into the boundary context fusion network;

[0017] The boundary context fusion network uses multi-scale feature channels to perform layer-by-layer feature fusion on the crack edges and surrounding structures, generating a fusion feature map containing the context relationship of the defect;

[0018] An asymmetric feature aggregation structure is introduced to learn the differential features of the crack target and background area in the fusion feature map;

[0019] During the training phase of the boundary context fusion network, based on the artificially labeled crack edge area in the defect image;

[0020] After training, an image deep feature model with weak defect perception ability is output.

[0021] Preferably, the process of determining whether the response intensity change of the image deep feature model in the crack area reaches a preset threshold is:

[0022] Input the standardized defect recognition image into the trained image deep feature model to obtain a deep feature map containing edge structure information;

[0023] Perform edge response analysis based on the deep feature map, combine the edge gradient change and the boundary context feature, and extract the crack area where the crack exists;

[0024] In the crack area, the local response intensity of the crack area is calculated;

[0025] a preset threshold of weak crack response is set, and the local response intensity of the crack region is compared with the threshold;

[0026] If the local response intensity of the crack region is greater than or equal to the threshold, it is determined that the crack region responds;

[0027] If the local response intensity of the crack region is less than the threshold, it is determined that the crack region does not respond.

[0028] Preferably, the process of dynamically correcting the preliminary labeling result in combination with the contrast adjustment mechanism of the crack and the background is:

[0029] In the image deep feature model output, a crack region with a response intensity meeting the requirements is identified, and a local image block is generated with the crack region as the center;

[0030] In the local image block, the local contrast value between the crack region and the adjacent background region is calculated;

[0031] According to the change trend of the local contrast with the boundary position, the boundary definition of the preliminary labeled region is evaluated;

[0032] If the contrast value is lower than the set threshold, it indicates that the boundary is blurred, and then based on the pixel gradient direction information and the neighborhood gray scale statistical features in the local image block, the boundary trend is dynamically judged;

[0033] Based on the boundary trend, edge expansion is performed on the preliminary labeled boundary to correct the defective contour;

[0034] In the dynamic correction process, an image gradient consistency discrimination algorithm is introduced to evaluate the consistency between the corrected boundary and the main direction of the crack.

[0035] Preferably, the process of contrast perception optimization of the corrected crack region by using a multi-scale morphological structure analysis method is:

[0036] Based on the crack region after dynamic correction, a plurality of size structure elements are selected, and morphological dilation and erosion operations are performed on the crack region at each scale to extract the crack contour;

[0037] A local contrast perception mechanism is introduced, and the crack region after dilation and erosion processing is evaluated for local contrast at each scale, the brightness difference distribution between the crack contour and the background region is calculated, and the improvement degree of local contrast is evaluated;

[0038] The crack contours obtained at each scale are fused by weighting to remove low-contrast redundant boundary information, and a unified crack contour map after fusion optimization is obtained;

[0039] Based on the unified crack contour map, the geometric parameters of the crack region are extracted.

[0040] Preferably, the process of dynamically generating defect identification results is:

[0041] According to the optimized extraction of the geometric parameters of the crack area, the defect recognition rules and boundary constraints are set, and the defect type and characteristic boundary are determined;

[0042] The two-dimensional coordinates of the crack area boundary are mapped to the image coordinate system in the original image acquisition process, the position index relationship of the defect area on the original image is established, and the defect identification label containing the defect position information, type and geometric description is generated;

[0043] According to the attribute characteristics in the defect identification label, combined with the label type, structural complexity and crack density of the defect, the defect severity evaluation index is calculated, and the preliminary severity grade is given;

[0044] The defect identification label is mapped to generate a defect identification result layer, and the defect identification result layer is overlaid on the original image through the image superposition mechanism to generate a defect identification result.

[0045] Preferably, the process of generating a defect analysis report is:

[0046] The defect identification result is mapped to the spatial component of the assembled retaining wall through the component positioning mechanism, and the number and spatial position coordinates of the corresponding structural component are obtained;

[0047] Combined with the historical operation and maintenance database, it is searched whether there are similar defects and development trends in the same component in the past detection period;

[0048] Based on the defect geometric parameter change trend and historical evolution record, the current defect severity is comprehensively determined;

[0049] According to the defect grade, component function importance and operation and maintenance cycle requirements, a defect analysis report is generated.

[0050] An assembled retaining wall defect identification system based on image recognition, comprising:

[0051] Image acquisition module: acquires multi-angle image data of assembled retaining wall, and performs illumination compensation and boundary enhancement processing;

[0052] Image modeling module: constructs image deep feature model based on boundary context fusion network to realize defect area recognition;

[0053] Defect detection module: realizes preliminary labeling and dynamic correction of defects through edge response enhancement and contrast analysis mechanism;

[0054] Structure analysis module: combined with multi-scale morphological analysis method, extracts defect geometric features and generates identification results;

[0055] Analysis output module: comprehensive historical data and structural information, hierarchical evaluation of defects and output of defect analysis report.

[0056] Compared with the prior art, the present application provides an assembled retaining wall defect identification method and system based on image recognition, which has the following beneficial effects:

[0057] 1. By collecting image samples under different time, angle and light conditions, and combining regional light compensation and boundary detail enhancement mechanism, the brightness uniformity and edge definition of defect areas such as cracks in the image are significantly improved, thereby constructing a more high-quality and representative defect identification image dataset, laying a stable foundation for subsequent model training and feature extraction.

[0058] 2. By introducing edge guiding mechanism and boundary context fusion network, multi-scale fusion and extraction of defect area edge details and surrounding structure can be realized, and the model's response sensitivity to fine-grained defect edges is enhanced by using asymmetric feature aggregation structure, improving the model's ability to identify tiny cracks and fuzzy defects.

[0059] 3. The edge response enhancement mechanism effectively identifies potential crack areas with prominent feature responses in the image, and the contrast adjustment mechanism further corrects the initial labeling results of blurred edges, effectively improving the accuracy and integrity of crack boundaries, reducing false positives and omissions, and improving the credibility of the labeling results.

[0060] 4. The multi-scale morphological structure processing can enhance the saliency of crack contour boundaries, combined with the local contrast perception mechanism, effectively eliminating background interference and redundant information, improving the accuracy of crack geometric feature extraction, providing accurate support for defect positioning, identification and classification, and finally realizing the visualization of defects.

[0061] 5. By establishing the positioning relationship between the defect identification results and the specific retaining wall components, and integrating historical operation and maintenance records, quantitative evaluation and trend analysis of current defects are realized, which facilitates dynamic control of the structure's health status, assists subsequent maintenance decision-making and life cycle management, and improves the scientific nature and efficiency of the assembled retaining wall operation and maintenance management. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 The present application is a schematic diagram of the method;

[0063] Figure 2 The present application is a schematic diagram of the structure. DETAILED DESCRIPTION

[0064] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0065] Embodiment 1: refer to Figure 1 As shown in the drawings, the image recognition-based prefabricated retaining wall defect identification method provided by the embodiments of the present application comprises the following steps:

[0066] S1: Obtain multi-view image data of the prefabricated retaining wall, and combine a boundary detail enhancement mechanism and a regional light compensation strategy to construct a defect identification image data set.

[0067] The process of constructing the defect identification image data set in S1 is as follows:

[0068] Collect multi-view image samples under different times, angles and light conditions; select representative prefabricated retaining wall sample paragraphs, and set multiple shooting points; collect images at time periods such as early morning, noon and evening to obtain image samples under different natural light intensities; adjust the shooting angle, including front view, side view and oblique view, to construct multi-view coverage; use an image collection device with fixed resolution and exposure parameters to ensure uniform sampling quality; uniformly number the original collected image samples and store them in an original image database;

[0069] Perform brightness distribution analysis on the multi-view image samples, dynamically evaluate the local brightness features of the multi-view images for uneven light problems, and adaptively determine the regional light compensation window; perform local brightness histogram analysis on each collected image to calculate the gray scale distribution offset of the image; use a sliding window to evaluate the brightness mean value and standard deviation of the image region by region to determine whether there is uneven light; construct a regional light compensation model based on gray scale equalization and local contrast enhancement; according to the analysis result, adaptively select the compensation window size and enhancement parameters to realize regional-level light balance; output the light compensated image, and retain the original detail structure of the image;

[0070] On the basis of illumination compensation, a boundary detail enhancement mechanism is used to enhance the crack and edge details in multi-view images, and a variety of data enhancement methods are fused; on the compensated image, a guided filter or gradient enhancement algorithm is applied to highlight the fine cracks and edge contours in the image; Laplace edge enhancement, structure tensor analysis and other technologies are used to improve the boundary definition; in view of the problem of insufficient data set, random rotation, scaling, mirroring, noise disturbance and other data enhancement strategies are introduced; the enhancement intensity and quantity ratio are controlled to maintain the original image semantics without distortion; the enhanced images are merged into a unified processing flow to construct a rich image sample set;

[0071] The pre-processed and enhanced multi-view images are standardized and processed to construct a standardized defect recognition image data set.

[0072] S2: edge-guided feature extraction is performed on the defect recognition image data set, a deep image feature model is constructed based on a boundary context fusion network, and an asymmetric feature aggregation structure is introduced to improve the model's perception of micro-edge defects.

[0073] The process of constructing the image deep feature model based on the boundary context fusion network in S2 is as follows:

[0074] The defect recognition image data set is input into the boundary context fusion network; the standardized defect recognition image data set constructed in the early stage is input into the neural network training framework in batches; the input image format is unified to a fixed resolution and a unified channel; a multi-scale input layer is set at the front end of the network to perform scale normalization and preliminary feature extraction on the image; a data iterator is established to load the image and its corresponding crack edge label mask in batches during the network training phase;

[0075] The boundary context fusion network uses multi-scale feature channels to perform layer-by-layer feature fusion on the crack edges and surrounding structures to generate a fusion feature map containing the context relationship of the defects; the network main structure includes multiple convolutional encoders and decoder channels, which use a pyramid multi-scale structure to obtain features at different semantic levels; the edge detail features extracted at each scale are transmitted to the high-level decoding network through horizontal connection and up-sampling; the network performs layer-by-layer fusion of local edge features and global semantic information; multi-scale information is integrated through skip connection, attention mechanism or dynamic weighting mechanism to generate a fusion feature map with context perception ability;

[0076] An asymmetric feature aggregation structure is introduced to learn the difference between the crack targets and background regions in the fusion feature map; an asymmetric convolution module is added in the feature aggregation stage to enhance the boundary response difference between the cracks and the background; a region contrast loss or feature difference enhancement module is used to learn the contrast between the crack regions and adjacent background regions in the fusion feature map; the background interference features are effectively suppressed to highlight the local variation features and edge responses of the cracks;

[0077] In the boundary context fusion network training phase, based on the manually labeled crack edge region in the defect image; the crack edge region with manual annotation mask is called from the data set as the training label; a joint loss function is constructed, including edge supervision loss and feature consistency loss; the network training adopts the back propagation algorithm to optimize the model parameters and update the feature extraction and fusion weight; the model response accuracy and generalization ability to the crack region in the training process are evaluated with the validation set;

[0078] After training, an image deep feature model with weak defect perception ability is output.

[0079] S3: Based on the edge response enhancement mechanism, determine whether the response intensity change of the image deep feature model in the crack region reaches the preset threshold value, if yes, label the edge region as having potential defects, and dynamically correct the preliminary labeling result in combination with the contrast adjustment mechanism of the crack and the background.

[0080] The process of determining whether the response intensity change of the image deep feature model in the crack region reaches the preset threshold value in S3 is:

[0081] The normalized defect recognition image is input into the trained image deep feature model to obtain a deep feature map containing edge structure information;

[0082] Based on the deep feature map, edge response analysis is performed, and the crack region with cracks is extracted in combination with the edge gradient change and the boundary context feature.

[0083] In the crack region, the local response intensity of the crack region is calculated.

[0084] A preset threshold value of weak crack response is set, and the local response intensity of the crack region is compared with the threshold value.

[0085] If the crack region feature response intensity is greater than or equal to the threshold value, the crack region is determined to respond.

[0086] If the crack region feature response intensity is less than the threshold value, the crack region is determined not to respond.

[0087] The process of dynamically correcting the preliminary labeling result in combination with the contrast adjustment mechanism of the crack and the background in S3 is:

[0088] In the image deep feature model output, the crack region with a response intensity meeting the requirements is identified, and a local image block is generated with the crack region as the center.

[0089] In the local image block, the local contrast value between the crack region and the adjacent background region is calculated.

[0090] According to the trend of the local contrast with the change of the boundary position, the boundary definition of the preliminary labeled region is evaluated; the local contrast value of each image block is compared with a preset crack saliency contrast threshold, a low-contrast region is identified, the brightness and texture continuity near the labeled boundary are counted, and whether the boundary has a fuzzy or error region is judged; if the contrast change trend is gentle, it is considered that the current boundary definition is insufficient and there is a risk of false labeling; the boundary blur risk score of each region is output as the basis for whether to enter the dynamic correction process;

[0091] If the contrast value is lower than the set threshold, it indicates that the boundary is blurred, and the boundary trend is dynamically judged based on the pixel gradient direction information in the local image block and the neighborhood gray scale statistical characteristics; a neighborhood sliding window is constructed with the current labeled boundary as the center, and the pixel information inside and outside the boundary is extracted; the gradient direction is used to judge the expansion trend of the crack; if the gradient direction is consistent with the extension direction of the original labeled boundary, and the external neighborhood response value is obviously improved, the edge expansion operation is performed; if the boundary gradient disappears or the proportion of background features increases, the boundary contraction is implemented to remove the mislabeling region;

[0092] Based on the boundary trend, the edge expansion is performed on the preliminary labeled boundary to correct the defective contour;

[0093] In the dynamic correction process, the image gradient consistency discrimination algorithm is introduced to evaluate the consistency between the corrected boundary and the main direction of the crack; the gradient direction vector field analysis is performed on the boundary region after correction, and the included angle between the main direction of the crack and the boundary extension direction is calculated; the gradient consistency factor is used as an evaluation index, and the formula is:

[0094]

[0095] In the formula, is the main gradient direction of the crack, is the gradient direction of the corrected boundary;

[0096] If the consistency factor is lower than the set threshold, the correction operation is cancelled, and the original labeled boundary is maintained; for all regions that have completed correction and passed the consistency verification, the final high-precision crack labeling mask image is output.

[0097] It can be understood that the judgment of whether the response strength change of the image deep feature model in the crack region reaches the preset threshold is:

[0098] Effect 1: By quantitatively analyzing the edge response strength output by the image deep feature model, the real crack and the pseudo crack caused by background noise, texture interference, etc. can be effectively distinguished; setting the threshold as the judgment standard can help to remove low-response regions, reduce misjudgment, and improve the accuracy and stability of defect identification;

[0099] The second function is that the response intensity of the crack region can reflect the structural integrity and boundary significance, thereby providing a judgment condition for subsequent boundary dynamic correction and crack profile optimization; and the differences in crack characteristics corresponding to different response intensities can be used as an important reference basis for calculating the severity of the defect, thereby supporting the classification and visual identification of the defect level.

[0100] The technical solution of the embodiment is as follows: The edge response intensity output by the image deep feature model is used to determine the potential defect position of the crack region, and when the response intensity reaches a set threshold, the region is marked as a suspected crack edge region; subsequently, a contrast adjustment mechanism for the crack and the background is introduced to dynamically correct the preliminary marking result, and the position and profile of the crack boundary are optimized according to the local contrast variation trend and pixel gradient information, thereby enhancing the accuracy of defect marking and the clarity of the boundary. Through the dual judgment and dynamic correction of the edge response and the local contrast, the embodiment can effectively enhance the recognition ability of the model for weak or fuzzy crack regions, significantly improve the boundary accuracy and stability of the defect recognition result, reduce the false detection rate caused by light and texture interference, and provide a more reliable region basis for subsequent crack shape parameter extraction and severity evaluation, thereby overall improving the practicability and robustness of image recognition in the defect detection scene of the fabricated retaining wall.

[0101] Embodiment 2: As shown in the following table, a defect identification method for a fabricated retaining wall based on image recognition comprises the following steps: Figure 1

[0102] S4: According to the corrected edge defect region, a multi-scale morphological structure analysis method is applied to perform contrast perception optimization processing on the correction result, extract the geometric parameters of the crack region, and dynamically generate a defect marking result.

[0103] The process of using the multi-scale morphological structure analysis method to perform contrast perception optimization on the corrected crack region in S4 is as follows:

[0104] Based on the crack region that is dynamically corrected, a plurality of size structure elements are selected, morphological dilation and erosion operations are performed on the crack region at each scale, and the crack profile is extracted; the dynamically corrected crack marking region image is cropped to extract the image block of the region of interest; a plurality of structure elements with different sizes are set, and morphological operations are performed on the image:

[0105] Dilation: Enhance the crack boundary profile and connect the broken regions;

[0106] Erosion: Remove noise pixels and compress non-structural interference;

[0107] ​A local contrast perception mechanism is introduced to evaluate the local contrast of the crack region after dilation and erosion at each scale, and the brightness difference distribution between the crack contour and the background region is calculated to evaluate the improvement degree of local contrast;

[0108] The crack contours obtained at each scale are fused by weighting to remove low-contrast redundant boundary information, and a unified crack contour map is obtained after fusion optimization. The multiple structure maps retained in the optimal scale set are unified and integrated by using a pixel-level or contour-level fusion strategy:

[0109] Pixel-level fusion: weighted summation or maximum value superposition is performed on all structure maps;

[0110] Contour-level fusion: the edge contours extracted at multiple scales are aggregated in connected regions, and the parts with consistent direction and continuous structure are retained;

[0111] An edge consistency filtering algorithm is introduced to remove redundant edges and repeated responses caused by multi-scale synthesis. A high-contrast, clear boundary and complete structure crack contour map is output as the final optimized labeling result;

[0112] Based on the unified crack contour map, the geometric parameters of the crack region are extracted, including length, width, direction and distribution density, etc., to provide parameter basis for defect identification results. The optimized crack contour map is subjected to connected component analysis, and the boundary point set of each crack region is extracted;

[0113] The following geometric feature parameters are calculated respectively:

[0114] Length: the longest path is calculated by principal axis fitting or skeleton extraction algorithm;

[0115] Width: the maximum and minimum distances are measured in the direction perpendicular to the crack principal axis;

[0116] Direction: the crack strike angle is obtained by principal component analysis or least squares fitting;

[0117] Distribution density: the number or total length of cracks per unit area, reflecting the overall defect distribution;

[0118] The above geometric feature parameters are encoded into structured data for defect severity evaluation, image statistical analysis or subsequent report generation. The final output includes the complete defect identification results including the optimized crack contour map and the geometric feature table.

[0119] The process of dynamically generating defect identification results in S4 is as follows:

[0120] According to the optimized crack geometric parameters, set the defect recognition rules and boundary constraint conditions, determine the defect type and characteristic boundary; read the geometric parameter data output in the crack contour optimization stage, including: length, width, direction, density, etc.; according to the pre-defined defect recognition rule library, judge whether the crack meets the standard of a specific defect type; introduce boundary constraint conditions to identify and eliminate false defect responses in the image edge or overlapping area; combine parameter characteristics and spatial distribution to assign a defect type label to each suspected crack area and label its boundary coordinates;

[0121] Map the two-dimensional coordinates of the crack area boundary to the image coordinate system in the original image acquisition process, establish the location index relationship of the defect area on the original image, and generate a defect recognition label containing defect position information, type and geometric description; for each defect area, extract its absolute pixel coordinates in the image; structure store the attribute information such as defect type, geometric parameter, direction angle and confidence, and form a multi-field identification tuple; according to the image size and acquisition meta information, support mapping the defect label to the real space coordinates; all defect labels are summarized to form an image-level defect label set

[0122] According to the attribute characteristics in the defect recognition label, combined with the label type, structural complexity and crack density of the defect, calculate the defect severity evaluation index and give the preliminary severity level; define the defect severity evaluation index, consider the following factors:

[0123] Defect type (structural > surface type);

[0124] Geometric size (the longer and wider, the higher the risk);

[0125] Crack density (total length / number of cracks per unit area);

[0126] Direction consistency and expansion trend (such as cracks perpendicular to the load direction are more dangerous);

[0127] Establish a defect level discrimination model, such as using interval rules or machine learning classifiers, to grade the defects:

[0128] Level1: slight (no intervention needed)

[0129] Level2: moderate (recommended for review)

[0130] Level3: severe (immediate action required)

[0131] The defect identification label mapping generates a defect identification result layer, and the defect identification result layer is overlaid on the original image through an image superposition mechanism to generate a defect identification result. A layer canvas with the same size as the original image is created, and all defect boundary information is drawn in sequence. Information such as type, grade, number, length, etc. is labeled beside each defect area, and color coding or icons are used to enhance identification. For example, red for severe, orange for moderate, and green for slight. The layer and the original image are superimposed in transparency to generate an output image containing defect in-situ labeling.

[0132] S5: Based on the defect identification result, the severity of the defect is evaluated by combining the component positioning mechanism and the component historical operation data to generate a defect analysis report.

[0133] The process of generating a defect analysis report in S5 is as follows:

[0134] The defect identification result is mapped to the spatial components of the prefabricated retaining wall through the component positioning mechanism to obtain the number and spatial position coordinates of the corresponding structural components. According to the position information bound during image acquisition, the image coordinate system is mapped to the spatial coordinate system of the retaining wall. The spatial position parameters in the defect identification result are read and matched to the BIM model or structural component database of the retaining wall. It is determined whether the defect boundary falls within the geometric range of a certain structural component to determine the defect component number. The spatial positioning information of the component is extracted and bound with the defect label for saving;

[0135] In combination with the historical operation database, it is checked whether there are similar defects and development trends in the same component in the past detection period. The component number is used as the main index to enter the historical operation database or defect monitoring file system to retrieve the past detection records of the component. Through spatial position comparison and crack pattern matching, it is determined whether there are defects in the same or adjacent areas in the historical records. The crack center position deviation, direction consistency, and image similarity are compared. If there are historical records, the length, width, grade, and other parameters of the defect in multiple detection periods are extracted to form time series evolution data. The evolution trend of the defect is labeled and saved as a structured analysis intermediate result.

[0136] Based on the defect geometric parameter variation trend and historical evolution record, the severity of the current defect is comprehensively determined.

[0137] According to the defect grade, component function importance, and operation cycle requirements, a defect analysis report is generated. The defect basic information, crack parameters, historical evolution trend, etc. are summarized. According to the importance of the component function, the priority information is extracted from the operation database. Combined with the structure health assessment specification and the actual operation strategy, the recommended measures are proposed. A picture-text analysis report is automatically generated.

[0138] Embodiment 3: Please refer to Figure 2As shown, an image recognition-based prefabricated retaining wall defect identification system comprises:

[0139] An image acquisition module: acquires multi-angle image data of the prefabricated retaining wall, and performs illumination compensation and boundary enhancement processing;

[0140] An image modeling module: constructs an image deep feature model based on a boundary context fusion network to realize defect area identification;

[0141] A defect detection module: realizes preliminary labeling and dynamic correction of defects through an edge response enhancement and contrast analysis mechanism;

[0142] A structure analysis module: extracts defect geometric features and generates identification results in combination with a multi-scale morphological analysis method;

[0143] An analysis output module: comprehensively analyzes historical data and structure information, grades and evaluates defects, and outputs a defect analysis report.

[0144] It should be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the term "includes", "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed or inherent to such a process, method, article or device.

[0145] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An image recognition-based prefabricated retaining wall defect identification method, characterized by The method comprises the following steps: Obtain multi-view image data of the assembled retaining wall, and combine a boundary detail enhancement mechanism and a regional light compensation strategy to construct a defect recognition image dataset; Perform edge-guided feature extraction on the defect recognition image dataset, construct an image deep feature model based on a boundary context fusion network, and introduce an asymmetric feature aggregation structure to improve the perception ability of the image deep feature model to micro edge defects; Determine whether the response intensity of the image deep feature model in the crack region reaches a preset threshold value through an edge response enhancement mechanism, and if so, label the crack region as having potential defects, and dynamically correct the crack region label result by combining a crack and background contrast adjustment mechanism; Use a multi-scale morphological structure analysis method to optimize the contrast perception of the corrected crack region, extract the geometric parameters of the crack region, and dynamically generate a defect label result; Based on the defect label result, combine a component positioning mechanism and historical operation data of the component to perform severity grading evaluation on the defects, and generate a defect analysis report; The process of determining whether the response intensity of the image deep feature model in the crack region reaches a preset threshold value comprises the following steps: Input the standardized defect recognition image into the trained image deep feature model to obtain a deep feature map containing edge structure information; Perform edge response analysis based on the deep feature map, combine edge gradient changes and boundary context features, and extract the crack region where the crack exists; Calculate the local response intensity of the crack region in the crack region; Set a preset threshold value of weak crack response, and compare the local response intensity of the crack region with the threshold value; If the local response intensity of the crack region is greater than or equal to the threshold value, it is determined that the crack region responds; If the local response intensity of the crack region is less than the threshold value, it is determined that the crack region does not respond; The process of dynamically correcting the preliminary label result by combining the crack and background contrast adjustment mechanism comprises the following steps: Identify the crack region with a response intensity meeting the requirements in the image deep feature model output, and generate a local image block centered on the crack region; Calculate the local contrast value between the crack region and the adjacent background region in the local image block; According to the change trend of the local contrast with the boundary position, evaluate the boundary sharpness of the preliminary labeled region; If the contrast value is lower than the set threshold value, it indicates that the boundary is blurred, and the boundary trend is dynamically judged based on the pixel gradient direction information and the neighborhood gray scale statistical features in the local image block; Based on the boundary trend, perform edge expansion on the preliminary labeled boundary to correct the defect contour; An image gradient consistency discrimination algorithm is introduced in the dynamic correction process to evaluate the consistency between the corrected boundary and the crack main direction.

2. The image recognition-based prefabricated retaining wall defect identification method according to claim 1, characterized in that, The process of constructing the defect recognition image dataset comprises the following steps: Collect multi-view image samples under different time, view angle and light conditions; Perform brightness distribution analysis on the multi-view image samples, dynamically evaluate the local brightness features of the multi-view images for uneven light, and adaptively determine the regional light compensation window; On the basis of light compensation, use the boundary detail enhancement mechanism to enhance the crack and edge details in the multi-view images, and simultaneously fuse multiple data enhancement methods; The pre-processed and enhanced multi-view images are uniformly normalized to construct a standardized defect recognition image dataset. 3.The image recognition-based prefabricated retaining wall defect identification method according to claim 2, characterized in that, The process of constructing an image deep feature model based on a boundary context fusion network is as follows: inputting the defect recognition image dataset into the boundary context fusion network; the boundary context fusion network uses multi-scale feature channels to perform layer-by-layer feature fusion on the crack edge and the surrounding structure to generate a fusion feature map containing the context relationship of the defect; an asymmetric feature aggregation structure is introduced to learn the differential features of the crack target and the background region in the fusion feature map; during the training phase of the boundary context fusion network, the crack edge region labeled by artificial labeling in the defect image is used for training; after training, an image deep feature model with weak defect perception capability is output.

4. The image recognition-based prefabricated retaining wall defect identification method according to claim 3, characterized in that, The process of using a multi-scale morphological structure analysis method to perform contrast perception optimization on the corrected crack region is as follows: based on the dynamically corrected crack region, select multiple size structure elements, respectively perform morphological dilation and erosion operations on the crack region at each scale, and extract the crack contour; a local contrast perception mechanism is introduced, which evaluates the local contrast of the crack region after dilation and erosion processing at each scale, calculates the brightness difference distribution between the crack contour and the background region, and evaluates the improvement degree of local contrast; the crack contours obtained at each scale are fused by weighting, and the low-contrast redundant boundary information is removed to obtain a fused and optimized unified crack contour map; based on the unified crack contour map, the geometric parameters of the crack region are extracted.

5. The image recognition-based prefabricated retaining wall defect identification method according to claim 4, characterized in that, The process of dynamically generating a defect identification result is as follows: according to the geometric parameters of the optimized extracted crack region, set the defect recognition rules and boundary constraint conditions, determine the defect type and feature boundary; map the two-dimensional coordinates of the crack region boundary to the image coordinate system in the original image acquisition process, establish the position index relationship of the defect region on the original image, and generate a defect identification label containing defect position information, type and geometric description; according to the attribute features in the defect identification label, combined with the label type, structural complexity and crack distribution density of the defect, calculate the defect severity evaluation index and give the preliminary severity level; map the defect identification label to generate a defect identification result layer, and overlay the defect identification result layer to the original image through the image superposition mechanism to generate a defect identification result.

6. The image recognition-based prefabricated retaining wall defect identification method according to claim 5, characterized in that, The process of generating a defect analysis report is as follows: map the defect identification result to the spatial components of the fabricated retaining wall through the component positioning mechanism, obtain the number and spatial position coordinates of the corresponding spatial components; combined with the historical operation and maintenance database, search whether the same component has defects and development trend in the past detection period; based on the defect geometric parameter change trend and historical evolution record, comprehensively determine the current defect severity; generate a defect analysis report according to the defect level, component function importance and operation and maintenance cycle requirements.

7. An image recognition based prefabricated retaining wall defect identification system, implementing the method of any one of claims 1-6, characterized by, It includes: image acquisition module: acquire multi-angle image data of the fabricated retaining wall, and perform illumination compensation and boundary enhancement processing; image modeling module: construct an image deep feature model based on a boundary context fusion network to realize defect region recognition; Defect detection module: Through edge response enhancement and contrast analysis mechanism, realize the preliminary labeling and dynamic correction of defects; Structure analysis module: Combined with multi-scale morphological analysis method, extract the geometric features of defects and generate identification results; Analysis output module: Comprehensive historical data and structure information, grade evaluation and output defect analysis report.

Images