A road pavement crack image recognition processing method and system

Abstract

The present application relates to the technical field of image processing, in particular to a road pavement crack image recognition processing method and system, comprising the following steps: normalizing and denoising the collected image, extracting contrast and gradient features to screen crack candidate areas, optimizing crack recognition through pixel classification and morphology completion, evaluating defect safety in combination with evolution logs, ranking maintenance priority according to crack density and traffic flow, and obtaining road maintenance decision support dataset; in the present application, environmental light interference is eliminated through illumination normalization and texture enhancement, self-adaptive threshold segmentation is realized in combination with local contrast and gradient distribution, crack edge features are strengthened, recognition robustness is improved, small cracks are perceived, completed and analyzed for morphological continuity, irregular morphologies are accurately restored, and correlation analysis is carried out by fusing crack evolution logs and traffic flow data, crack density and aggregation characteristics of risk road sections are quantified, high-risk areas with expansion trend are screened, and maintenance priority and resource allocation efficiency are optimized.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for recognizing and processing road surface cracks. Background Technology

[0002] Image processing technology primarily involves the analysis and processing of image data using computer technology to extract useful information, improve image quality, or achieve specific functions. Core aspects of image processing include image enhancement, image segmentation, feature extraction, and pattern recognition. This field has wide applications, covering areas such as medical image processing, video surveillance, autonomous driving, and remote sensing image processing. The implementation of image processing technology relies on image acquisition, algorithm design, and model training. By identifying, tracking, and classifying targets in images, it enables effective analysis and application of specific scenes or objects.

[0003] Traditional road surface crack image recognition methods involve acquiring images of the road surface using image acquisition devices (such as cameras and drones) and then automatically detecting and analyzing cracks using image recognition technology. This method employs image preprocessing techniques, such as noise reduction and contrast enhancement, to improve the recognizability of crack features. Through image segmentation and feature extraction techniques, it identifies information such as the location, shape, and size of the cracks. Traditional image recognition methods rely on edge detection, threshold segmentation, and template matching techniques for preliminary crack detection; however, their recognition accuracy and robustness are affected by factors such as image quality and lighting changes, and their processing speed is relatively slow.

[0004] Existing traditional road surface crack image recognition and processing technologies rely solely on edge detection and global threshold segmentation logic. When faced with complex and variable lighting environments and road surface texture interference, fixed parameters are difficult to effectively adapt to local contrast differences, resulting in incomplete crack feature extraction or misjudging background noise as cracks. Furthermore, their template-matching-based operation mode lacks the ability to perceive and complete small cracks, making it difficult to construct a continuous description of crack morphology, resulting in poor robustness of recognition results. At the same time, simple morphology recognition cannot be linked to historical evolution logs and traffic flow data, leaving the analysis of crack spatial evolution trends and distribution density incomplete. This results in a lack of multi-dimensional data support for assessing the severity of crack damage and the urgency of maintenance, leading to a lack of scientific rigor and accuracy in the final road maintenance decisions. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a method for recognizing and processing road surface crack images, comprising the following steps: To achieve the above objectives, the present invention adopts the following technical solution: a method for recognizing and processing road surface crack images, comprising the following steps: S1: Based on the collected road surface image data, perform illumination normalization and noise suppression, extract texture enhancement image features, generate a standardized image dataset, store the standardized image data, update the global image processing configuration, detect the data synchronization status, and obtain the basic image dataset for crack recognition. S2: Based on the crack recognition basic image dataset, extract the local contrast index and gradient distribution features of the image, screen potential crack candidate regions, perform adaptive threshold segmentation on low contrast regions, perform edge enhancement on high gradient regions, record texture interference regions, and obtain crack candidate region dataset. S3: Based on the crack candidate region dataset, perform pixel-level classification, analyze the continuity of crack morphology, filter irregular morphological regions, perform perceptual completion on small crack regions, analyze crack direction and width distribution, and obtain an optimized road surface crack identification dataset. S4: Based on the optimized pavement crack identification dataset, detect the spatial location information of cracks, extract inspection records, detect the evolution log of cracks at the same location, and filter out areas of cracks that reappear or expand within a short period of time to obtain a pavement defect safety assessment dataset.

[0006] As a further aspect of the present invention, the basic image dataset for crack recognition includes illumination intensity information, noise level information, texture enhancement coefficient, multi-scale feature map, and synchronization status identifier; the crack candidate region dataset includes local contrast information, gradient magnitude information, edge enhancement parameters, interference region markers, segmentation threshold configuration, and candidate region coordinates; the optimized pavement crack recognition dataset includes crack pixel classification labels, morphological continuity score, context completion identifier, crack direction vector, width distribution histogram, and crack type index; and the pavement defect safety assessment dataset includes spatial location coordinates, original crack records, evolution trend identifier, recurrence count, and expansion rate parameter.

[0007] As a further aspect of the present invention, the steps of obtaining the basic image dataset for crack recognition are as follows: S101: Based on the collected road surface image data, including the original RGB image, illumination intensity distribution and environmental noise parameters, white balance correction and gamma correction are performed on the image, wavelet denoising is performed to generate a denoised image, multi-scale Gabor filtering is performed on the denoised image, texture responses with different directions and scales are extracted, and multi-scale feature maps are generated by fusion. S102: Based on the multi-scale feature map, calculate the local variance of the channel, apply contrast stretching to the low variance region, perform dynamic range compression on the high variance region, generate a standardized image dataset, store the standardized image dataset in the image feature database, and generate a unique data identifier. S103: Based on the unique data identifier, check the consistency of the data version between the local cache and the global image processing configuration. If the versions are inconsistent, trigger the incremental synchronization mechanism, upload the difference data block and update the version number in the global configuration. If they are consistent, clear the temporary cache and obtain the crack recognition basic image dataset.

[0008] As a further aspect of the present invention, the steps for obtaining the crack candidate region dataset are as follows: S201: Based on the crack recognition basic image dataset, calculate the contrast standard range and gradient magnitude benchmark range of the local region of the image, identify regions that exceed the preset lower limit of contrast but are lower than the upper limit of gradient, and mark them as low contrast candidate regions. Identify regions whose gradient magnitude exceeds the preset upper limit of gradient and mark them as high gradient candidate regions to obtain an initial candidate region set. S202: Based on the initial candidate region set, apply the Otsu adaptive thresholding algorithm to the low-contrast candidate regions to generate a binary segmentation mask, apply Canny edge detection to the high-gradient candidate regions and superimpose non-extreme suppression to generate an edge enhancement mask, and merge the two types of masks to form a unified candidate region. S203: Based on the unified candidate region, detect whether there are non-crack interference objects of the same type as the crack texture in the region. If so, generate interference region markers based on the texture spectrum features, and record the interference type and location to obtain the crack candidate region dataset.

[0009] As a further aspect of the present invention, the steps for obtaining the optimized pavement crack identification dataset are as follows: S301: Based on the crack candidate region dataset, crop the candidate region image patch, input it into the pre-trained U-Net++ semantic segmentation network, output the crack probability map of each pixel, perform morphological closing operation on the probability map and fill in the break points to generate a preliminary segmentation result; S302: Based on the preliminary segmentation results, calculate the connected components of the crack centerline, perform attention mechanism to complete the components whose length is less than a preset length threshold, predict the missing segments using the crack direction of the adjacent area, and generate a completed crack map. S303: Based on the completed crack map, extract the centerline direction vector and cross-sectional width distribution of each crack, count the frequency of each type of crack in the image, and sort them in descending order of frequency to obtain the optimized road surface crack identification dataset.

[0010] As a further aspect of the present invention, the calculation of the connected components of the crack centerline refers to extracting the crack centerline with a single pixel width based on the geometric center distribution of the crack region in the preliminary segmentation result, and marking the connectivity of the pixel set connected by eight neighborhoods in the crack centerline to form multiple connected components of the crack centerline.

[0011] As a further aspect of the present invention, the steps for obtaining the road surface defect safety assessment dataset are as follows: S401: Based on the optimized pavement crack identification dataset, extract the geographic coordinate information of each crack, match it with the road grid database, determine the road segment number and station number, and obtain a spatial positioning list; S402: Based on the spatial location list, query the original records in the crack defect management database, extract the crack area, length and type change data of the same station location in the past inspection cycle, and generate a crack evolution time series; S403: Based on the crack evolution time series, calculate the crack propagation rate per unit time, count the number of times the same location is identified in two consecutive inspection cycles, filter out areas where the propagation rate exceeds a preset rate threshold or the number of recurrences exceeds a preset number threshold, and obtain a road surface defect safety assessment dataset.

[0012] As a further aspect of the present invention, the method further includes step S5: S5: Based on the road surface defect safety assessment dataset, extract the crack distribution density of risky road sections, analyze the crack aggregation pattern, screen areas where the crack density exceeds the preset density threshold, and combine traffic flow data to sort the maintenance priorities to obtain a road maintenance decision support dataset. The road maintenance decision support dataset includes crack density distribution maps, clustering pattern categories, a list of risky road sections, maintenance priority scores, and traffic flow association identifiers.

[0013] As a further aspect of the present invention, the steps of the road maintenance decision support dataset are specifically as follows: S501: Based on the road surface defect safety assessment dataset, count the number and total area of ​​cracks within a unit length of road segment, calculate the crack density index, compare it with the preset density threshold, filter out road segments that exceed the standard, and generate a set of high-density areas. S502: Based on the high-density area set, the DBSCAN clustering algorithm is applied to identify the spatial clustering pattern of cracks, distinguishing between linear extension type, network crack type and random scattered type. Combined with the daily average traffic flow level of the road segment, maintenance priority weights are assigned to generate a road maintenance decision support dataset.

[0014] A road surface crack image recognition and processing system includes: Based on the collected road surface image data, the image processing module performs illumination normalization, noise suppression and multi-scale texture enhancement to generate a standardized image dataset, which is stored in the image feature database. If synchronization fails, it is cached in the local storage unit to obtain the basic image dataset for crack recognition. Based on the crack recognition base image dataset, the candidate region generation module calculates local contrast and gradient distribution, performs adaptive threshold segmentation and edge enhancement, marks texture interference regions, and obtains a crack candidate region dataset. The semantic segmentation optimization module performs pixel-level classification based on the crack candidate region dataset, performs morphological continuity analysis and context-aware completion, and counts the frequency of crack types to obtain the optimized road crack identification dataset. Based on the optimized pavement crack identification dataset, the defect assessment module matches spatial location information, queries original inspection records, calculates crack propagation rate and recurrence frequency, filters risk areas, and obtains a pavement defect safety assessment dataset. Based on the pavement defect safety assessment dataset, the maintenance decision module calculates crack density, identifies spatial clustering patterns, and allocates maintenance priorities by combining traffic flow data, thereby generating a road maintenance decision support dataset.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by performing illumination normalization and texture enhancement processing on the acquired images, interference caused by changes in ambient lighting is effectively eliminated. Adaptive threshold segmentation is performed using local contrast indicators and gradient distribution features to strengthen edge features in high-gradient areas, significantly improving the anti-interference capability and robustness of crack identification. Perceptual completion and morphological continuity analysis are performed on small crack areas to accurately restore irregular crack morphologies. Multi-dimensional correlation analysis is conducted by combining crack evolution logs and traffic flow data to quantify the distribution density and clustering patterns of cracks in risky road sections, accurately identifying high-risk areas with an expansion trend, and optimizing the maintenance priority ranking logic to ensure the scientific and efficient allocation of road maintenance resources. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0019] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0020] Please see Figure 1 This invention provides a method for recognizing and processing road surface crack images, comprising the following steps: S1: Based on the collected road surface image data, the image is processed by illumination normalization and noise suppression, the image features after texture enhancement are extracted, a standardized image dataset is generated, and the standardized image dataset is stored in the image feature database, synchronously updated to the global image processing configuration, the data synchronization status is detected, and the basic image dataset for crack recognition is obtained. S2: Based on the crack recognition basic image dataset, extract the local contrast index and gradient distribution features of the image, screen potential crack candidate regions, perform adaptive threshold segmentation on low contrast regions, perform edge enhancement operation on high gradient regions, and record and mark the interference regions if there is texture interference to obtain the crack candidate region dataset. S3: Based on the crack candidate region dataset, the candidate regions are classified at the pixel level, the continuity of crack morphology is analyzed, irregular morphological regions are filtered, perceptual completion is performed on small crack regions, the crack direction and width distribution are analyzed, and high-frequency crack types are filtered to obtain the optimized road crack identification dataset. S4: Based on the optimized pavement crack identification dataset, detect the spatial location information of crack areas, extract the original inspection records, detect the crack evolution log at the same location, and filter out crack areas that reappear or expand in a short period of time to obtain the pavement defect safety assessment dataset. S5: Based on the road surface defect safety assessment dataset, extract the crack distribution density of risky road sections, analyze the crack aggregation pattern, screen areas where the crack density exceeds the preset density threshold, and combine traffic flow data to rank maintenance priorities, thus obtaining a road maintenance decision support dataset.

[0021] The basic image dataset for crack recognition includes illumination intensity information, noise level information, texture enhancement coefficient, multi-scale feature map, and synchronization status identifier. The crack candidate region dataset includes local contrast information, gradient magnitude information, edge enhancement parameters, interference region markers, segmentation threshold configuration, and candidate region coordinates. The optimized pavement crack recognition dataset includes crack pixel classification labels, morphological continuity scores, context completion identifiers, crack direction vectors, width distribution histograms, and crack type indexes. The pavement defect safety assessment dataset includes spatial location coordinates, original crack records, evolution trend identifiers, recurrence counts, and expansion rate parameters. The road maintenance decision support dataset includes crack density distribution maps, clustering pattern categories, risky road segment lists, maintenance priority scores, and traffic flow association identifiers.

[0022] Please see Figure 2 The specific steps for using the basic image dataset for crack identification are as follows: S101: Based on the collected road surface image data, including the original RGB image, illumination intensity distribution and environmental noise parameters, white balance correction and gamma correction are performed on the image, wavelet denoising is performed to generate a denoised image, multi-scale Gabor filtering is performed on the denoised image, texture responses with different directions and scales are extracted, and multi-scale feature maps are generated by fusion. First, the vehicle-mounted high-resolution industrial data acquisition equipment and environmental sensing components are activated to acquire raw RGB road surface images with a resolution of 4096×2048 pixels, real-time illumination intensity distribution data measured in lux, and environmental noise parameters. After acquiring the raw data, the process immediately performs white balance correction on the RGB image based on the gray-world hypothesis. Specifically, the arithmetic mean of the pixel brightness in the red, green, and blue channels is calculated. The average value of the green channel is selected as the reference value. The ratio of the average value of the red channel to the average value of the green channel is calculated as the red gain coefficient, and the ratio of the average value of the blue channel to the average value of the green channel is calculated as the blue gain coefficient. Then, all pixel values ​​in the red and blue channels are multiplied by their respective gain coefficients to eliminate color cast caused by color temperature differences. Next, the process performs gamma nonlinear correction to address uneven illumination distribution. The gamma correction parameter is set to 0.45, and the normalized brightness value of each pixel in the image is raised to the power of 0.45. For example, if a pixel's normalized brightness is 0.2, the output brightness after the operation is approximately 0.48. This improves the grayscale levels in dark areas and suppresses overexposure in highlight areas. Subsequently, the process introduces the db4 wavelet basis function to perform a three-level wavelet decomposition on the corrected image, extracting high-frequency detail coefficients in the horizontal, vertical, and diagonal directions. A soft thresholding function is used to shrink these high-frequency coefficients to filter out Gaussian white noise. Then, the image is reconstructed using inverse wavelet transform to generate the denoised image. Based on this, the process constructs a multi-scale Gabor filter bank, setting directional parameters covering six directions (0°, 30°, 60°, 90°, 120°, and 150°) and scale parameters covering four scales (3°, 5°, 7°, and 9°), generating a total of 24 differentiated Gabor filter kernels. The process involves performing convolution operations between the denoised image and each of the 24 filter kernels to obtain the texture response amplitude at the corresponding direction and scale. For each pixel location, the maximum value among the 24 response amplitudes is selected as the feature value of that point, and finally, the features are fused to generate a multi-scale feature map.

[0023] S102: Based on multi-scale feature maps, calculate the local variance of the channels, apply contrast stretching to low-variance regions, perform dynamic range compression on high-variance regions, generate a standardized image dataset, store the standardized image dataset in the image feature database, and generate a unique data identifier. Local statistical characteristic analysis and data standardization are performed based on multi-scale feature maps. Specifically, a 15×15 pixel sliding window is used to traverse the feature map row by row and column by column, calculating the local variance of all pixel grayscale values ​​within each window. A pre-set variance threshold of 255 is used. When the calculated local variance value is less than this threshold, the window region is determined to be a low-contrast region, and a contrast stretching algorithm is applied to this region. The stretching operation logic is as follows: First, the maximum and minimum grayscale values ​​of the pixels within the window are obtained. The difference between the current pixel value and the minimum value is calculated, and the difference between the maximum and minimum values ​​is calculated again. The former is divided by the latter to obtain the normalization ratio. Finally, this ratio is multiplied by 255 to obtain the stretched pixel value. For example, if the minimum grayscale value within the window is 40, the maximum grayscale value is 60, and the current pixel value is 50, the calculated difference is 10, the range is 20, and the ratio is 0.5. Multiplying this by 255 yields a new pixel value of 127.5. When the calculated local variance value is greater than or equal to the variance baseline threshold of 255, the region is determined to be a high-variance region. This process performs dynamic range compression, employing a logarithmic transformation algorithm to calculate the natural logarithm of the pixel value plus 1, then multiplying it by a preset scaling factor to compress the dynamic range of the highlight region while preserving edge details. After the above partitioning process, a standardized image dataset is generated. This dataset is then written into an image feature database, and the binary data stream of each standardized image is hashed using the MD5 message digest algorithm to generate a unique 32-bit data identifier string, which serves as the fingerprint information for subsequent version verification.

[0024] S103: Based on the unique data identifier, check the consistency of the data version between the local cache and the global image processing configuration. If the versions are inconsistent, trigger the incremental synchronization mechanism, upload the difference data block and update the version number in the global configuration. If they are consistent, clear the temporary cache and obtain the crack recognition basic image dataset. A consistency check between the local cache and the global configuration center is performed based on a unique data identifier. Specifically, the unique data identifier of the latest generated image data in the local cache is extracted, and a comparison request is sent to the global image processing configuration center. The local identifier is matched character-level with the identifier string in the global record. If the comparison shows a difference between the two identifier strings, the data versions are determined to be inconsistent, and the incremental synchronization mechanism is immediately triggered. Incremental synchronization compares the local data stream and the global data stream byte-by-byte using a binary differential algorithm, selecting only byte blocks with different binary values ​​as differential data. A transmission channel is then established to upload only these differential data blocks, significantly reducing network transmission load. For example, if the local identifier is string A and the global record is string B, and the differential calculation shows a difference of 150 kilobytes, only these 150 kilobytes are transmitted. After uploading, the process obtains the current timestamp and operation sequence number, combines them to generate a new version number, such as 20251020001, and updates the version record in the global configuration. If the comparison results show that the two identifier strings are exactly the same, it indicates that the local cached data is the latest version. The process then performs a memory cleanup operation to release the temporary cache area used to store intermediate calculation results, and only retains the finally generated standardized image dataset to ensure that the data called by subsequent processing steps has complete consistency and timeliness.

[0025] Please see Figure 3 The specific steps for creating the crack candidate region dataset are as follows: S201: Based on the crack recognition basic image dataset, calculate the standard range of contrast and the benchmark range of gradient magnitude of local regions of the image, identify regions that exceed the preset lower limit of contrast but are lower than the upper limit of gradient, and mark them as low contrast candidate regions. Identify regions whose gradient magnitude exceeds the preset upper limit of gradient and mark them as high gradient candidate regions, thus obtaining the initial candidate region set. Based on the crack identification image dataset, local region grayscale statistical feature calculations are performed to screen potential crack regions. Specifically, the image is divided into 50×50 pixel grid cells, and the contrast value and average gradient magnitude of the pixels within each grid cell are calculated. This process pre-sets a lower contrast threshold of 30 and an upper gradient threshold of 85. For any grid cell, the calculated contrast value is compared with the lower contrast threshold, and the average gradient magnitude is compared with the upper gradient threshold. If the contrast value of a grid cell is greater than 30 and the average gradient magnitude is less than 85, the process determines that the region has a shallow surface texture but exhibits grayscale differences and marks it as a low-contrast candidate region. If the average gradient magnitude of a grid cell is greater than the upper gradient threshold of 85, the process determines that the region has significant edge features and marks it as a high-gradient candidate region. For example, if a certain grid region has a contrast ratio of 45 and an average gradient magnitude of 60, this region is accurately labeled as a low-contrast candidate region because 45 is greater than 30 and 60 is less than 85. Another grid region has an average gradient magnitude of 120, which is greater than 85, and is therefore labeled as a high-gradient candidate region. This process summarizes all labeled low-contrast and high-gradient candidate regions to form an initial candidate region set.

[0026] S202: Based on the initial candidate region set, apply the Otsu adaptive thresholding algorithm to the low-contrast candidate region to generate a binary segmentation mask, apply Canny edge detection to the high-gradient candidate region and superimpose non-extreme suppression to generate an edge enhancement mask, and merge the two types of masks to form a unified candidate region. Differential segmentation strategies are employed for different marked regions in the initial candidate region set. For image blocks marked as low-contrast candidate regions, the Otsu adaptive thresholding algorithm is applied. By traversing all gray levels from 0 to 255 as candidate thresholds, the inter-class variance of foreground and background pixel classes is calculated separately. The gray level corresponding to the maximum inter-class variance is selected as the optimal segmentation threshold, and the image block is binarized to generate a binary segmentation mask. For image blocks marked as high-gradient candidate regions, the Canny edge detection algorithm is applied. First, the magnitude and direction of the image gradient are calculated. Non-extreme suppression is performed along the gradient direction, retaining only the pixels with the largest local gradient. Then, a double-threshold concatenation method is used, setting the high threshold to 100 and the low threshold to 50, to concatenate strong edge pixels with weak edge pixels to generate an edge enhancement mask. Finally, the process performs a bitwise OR operation on the generated binary segmentation mask and the edge enhancement mask. That is, for each pixel position in the image, as long as the pixel is marked as foreground in either mask, it is retained as a candidate crack pixel in the final merging result, thereby merging the two types of masks to form a unified candidate region.

[0027] S203: Based on the unified candidate region, detect whether there are non-crack interference objects in the region that are similar to the crack texture. If so, generate interference region labels based on the texture spectrum features, and record the interference type and location to obtain the crack candidate region dataset. Interference removal and purification operations are performed on unified candidate regions. Specifically, all connected components within the unified candidate region are traversed. For each connected component image patch, a two-dimensional discrete Fourier transform is performed to convert it to the frequency domain. The energy distribution concentration index in the spectral image is calculated. This process pre-sets an energy concentration threshold of 0.75. If the calculated spectral energy concentration value of a connected component is greater than this threshold, and the energy is mainly distributed in the low-frequency component region, the process determines that the region is a blocky interference rather than a linear crack. Once interference is confirmed, the process generates interference region markers based on texture spectral features, resets the pixel values ​​of the region to background values, and records the specific type of interference, such as oil stains, patch blocks, and their geometric center position in the image coordinate system. For example, if the calculated spectral energy concentration of a candidate connected component is 0.82, which is greater than the threshold of 0.75, the process determines it to be oil stain interference and removes it. After traversing and filtering all connected components, the remaining regions constitute a high-confidence crack candidate region dataset.

[0028] Please see Figure 4 The specific steps for optimizing the road surface crack identification dataset are as follows: S301: Based on the crack candidate region dataset, crop the candidate region image patch, input it into the pre-trained U-Net++ semantic segmentation network, output the crack probability map of each pixel, perform morphological closing operation on the probability map and fill in the break points to generate preliminary segmentation results; High-precision semantic segmentation is performed based on a crack candidate region dataset. First, image patches of 256×256 pixels are cropped from the original image according to the bounding box coordinates of the candidate regions. These patches are normalized and then input into a pre-trained U-Net++ semantic segmentation network. This network structure includes an encoder path, a decoder path, and nested dense skip connections. The encoder uses ResNet34 as the feature extraction backbone, and the decoder upsamples layer by layer to restore resolution. Each layer's nodes are connected by dense convolutional blocks, and each node is configured with a ReLU activation function and a batch normalization layer. The network output layer uses a Sigmoid activation function to output the probability value of each pixel belonging to the crack category, ranging from 0 to 1. A probability threshold of 0.5 is set for this process; pixels with an output probability value greater than 0.5 are classified as crack pixels. Subsequently, morphological closing operations are performed on the probability map using a circular structuring element with a radius of 3 pixels. Dilation followed by erosion is performed to connect broken crack fragments and fill internal holes, generating preliminary segmentation results. The specific training parameters of the network are shown in Table 1. The model is trained by strictly setting the hyperparameters to ensure the segmentation accuracy of the network when dealing with minor cracks. For example, when using the combined loss function of Dice_Loss and BCE_Loss for backpropagation, the class imbalance problem with a small proportion of crack pixels is effectively solved.

[0029] Table 1: Hyperparameter Settings for Semantic Segmentation Network Training

[0030] S302: Based on the preliminary segmentation results, calculate the connected components of the crack centerline, perform attention mechanism to complete the components whose length is less than the preset length threshold, predict the missing segments using the crack direction of the neighboring region, and generate the completed crack map. Calculating the connected components of the crack centerline refers to extracting the crack centerline with a single pixel width based on the geometric center distribution of the crack region in the preliminary segmentation results, and marking the connectivity of the pixel set in the crack centerline that uses eight-neighborhood connection to form multiple connected components of the crack centerline. Based on the preliminary segmentation results, crack centerline extraction and breakpoint completion operations are performed. First, a skeleton thinning algorithm is applied to iteratively peel away the edge pixels of the crack region until the crack region is thinned to a centerline of a single pixel width. Then, based on the eight-neighborhood connection rule, all pixels on the centerline are traversed, and mutually adjacent pixel sets are marked as the same connected component. This process counts the number of pixels contained in each connected component as its length, and sets the length filtering threshold to 15 pixels. For tiny components with a length less than 15 pixels, the process initiates an attention completion mechanism, searching for other crack endpoints within a 20-pixel radius centered on the breakpoint, calculating the Euclidean distance and directional angle between the breakpoints. If the distance between two breakpoints is less than 20 pixels and the directional angle is less than 30 degrees, the process uses a cubic spline interpolation algorithm to calculate the smooth path coordinates connecting the two points and sets the pixel values ​​on the predicted path to 1, thereby connecting the broken crack. For example, if the coordinates of the endpoints of two crack components are detected as 100, 100 and 110, 115 respectively, the calculated distance is 18.02 pixels, which is less than 20 pixels, and the extension direction is consistent, the process automatically generates connecting line segments to generate a complete crack map.

[0031] S303: Based on the completed crack map, extract the centerline direction vector and cross-sectional width distribution of each crack, count the frequency of each type of crack in the image, sort them in descending order of frequency, and obtain the optimized pavement crack identification dataset. Based on the completed crack image, morphological feature quantification and dataset reorganization are performed. Specifically, each complete crack skeleton is traversed, and principal component analysis is applied to calculate the covariance matrix of the crack pixel set. The direction of the first principal component eigenvector is then used as the overall crack orientation angle. Simultaneously, the cross-sectional width of the original segmented region is calculated along the centerline normal direction, and the arithmetic mean and maximum value of the entire crack width are statistically analyzed to describe the cross-sectional width distribution characteristics. Next, based on the orientation angle, cracks are classified into transverse cracks (0-30 degrees), longitudinal cracks (60-90 degrees), and mesh cracks, and the frequency of each type of crack in the current image is counted. Finally, the crack data is sorted and indexed according to the frequency values ​​from largest to smallest, and the structured data containing geometric morphological parameters and classification statistics is integrated to output an optimized pavement crack identification dataset.

[0032] Please see Figure 5 The specific steps for creating a road surface defect safety assessment dataset are as follows: S401: Based on the optimized pavement crack identification dataset, extract the geographic coordinate information of each crack, match it with the road grid database, determine the road segment number and station number, and obtain a spatial location list; Based on the optimized pavement crack identification dataset, spatial geographic information matching is performed. First, the center point coordinates of each identified crack in the image coordinate system are extracted. A coordinate transformation matrix is ​​constructed using GPS positioning data synchronously recorded by the vehicle-mounted acquisition system, mapping the image pixel coordinates to geographic latitude and longitude coordinates. Then, the process accesses a road grid database storing road network topology and station information. A spatial projection algorithm is used to calculate the perpendicular projection point of the crack's latitude and longitude coordinates onto the road centerline vector in the database. The road segment number and specific station location of the crack are determined based on the relative position of the projection point in the road network. For example, if the latitude and longitude of a crack are matched to G105 National Highway, it is calculated to be located at K10 plus 500 meters. Finally, the process integrates all identified cracks and their corresponding road segment numbers and station locations to generate a detailed spatial location list.

[0033] S402: Based on the spatial location list, query the original records in the crack defect management database, extract the crack area, length and type change data of the same station location in the past inspection cycle, and generate crack evolution time series; Based on the spatial location list, a historical evolution data query is initiated to the crack defect management database. For each station number in the list, the process retrieves the original records for that location within the past five inspection cycles, extracting crack area, crack length, and crack type data from each cycle. The extracted multidimensional data is then arranged chronologically to generate a crack evolution time series for a specific spatial location. For example, for station K10 + 500 meters, the crack areas extracted from inspections in January 2023, April 2023, and July 2023 are 0.05 square meters, 0.06 square meters, and 0.08 square meters, respectively. This time series data accurately reflects the dynamic change process of crack defects at that location over time, providing the necessary data foundation for subsequent calculations of the expansion rate.

[0034] S403: Based on the crack evolution time series, calculate the crack propagation rate per unit time, count the number of times the same location is identified in two consecutive inspection cycles, filter out areas where the propagation rate exceeds a preset rate threshold or the number of recurrences exceeds a preset number threshold, and obtain a pavement defect safety assessment dataset. Based on the time series of crack evolution, a pavement defect development trend and safety assessment is performed. First, the crack propagation rate per unit time is calculated. The calculation logic is as follows: obtain the crack area value of the current inspection cycle and the crack area value of the previous cycle, calculate the difference between the two, and then divide by the time interval between the two cycles in days. For example, if the current area is 0.08 square meters, the previous period was 0.06 square meters, and the interval is 90 days, then the propagation rate is calculated to be 0.00022 square meters per day. This process sets a safety threshold of 0.0001 square meters per day. If the calculated propagation rate is greater than this threshold, the crack is considered to be in a state of rapid deterioration. Simultaneously, the process counts the number of times the same station number is identified as a crack in two consecutive inspection cycles, setting a repetition threshold of 3 times. If the repetition count is greater than 3 times, it indicates that the defect is persistent. Finally, this process filters out all areas with excessive propagation rates or excessive repetition counts, classifying them as high-risk objects, and compiling them into a pavement defect safety assessment dataset.

[0035] Please see Figure 6 The specific steps for creating a road maintenance decision support dataset are as follows: S501: Based on the road surface defect safety assessment dataset, the number and total area of ​​cracks within a unit length of road segment are statistically analyzed, crack density index is calculated, compared with preset density threshold, road segments exceeding the standard are screened, and a set of high-density areas is generated. Based on a road surface defect safety assessment dataset, macroscopic density analysis and high-risk area screening are performed. Specifically, for road segments of 1000 meters in length, the total number and area of ​​all cracks marked as high-risk are counted. Next, a crack density index is calculated by dividing the total crack area of ​​the road segment by the total road surface area, yielding a percentage density value. A pre-set density warning threshold of 1.5% is used. The calculated real-time density index is compared to the pre-set threshold; if the density index is greater than 1.5%, the road segment is considered severely damaged. For example, a 1-kilometer road segment with a total road surface area of ​​7000 square meters and a detected crack area of ​​140 square meters yields a density of 2.0%, which is greater than 1.5%, thus the segment is deemed to exceed the threshold. This process aggregates all screened-out exceeding-threshold road segments and their corresponding density data to generate a high-density area set.

[0036] S502: Based on the high-density area set, the DBSCAN clustering algorithm is applied to identify the spatial clustering pattern of cracks, distinguishing between linear extension type, network crack type and random scattered type. Combined with the daily average traffic flow level of the road segment, maintenance priority weights are assigned to generate a road maintenance decision support dataset. Spatial clustering analysis and maintenance decision support were performed on the selected high-density area set. First, the DBSCAN density clustering algorithm was applied to identify the spatial distribution pattern of cracks, setting the cluster radius (Epsilon) to 1.5 meters and the minimum number of contained points (MinPts) to 4. By calculating the Euclidean distance between the centroids of cracks, spatially adjacent cracks were clustered into the same cluster. Based on the geometric morphology of the clustering results, the crack distribution was classified into linear extension type, network crack type, and random scattering type. Next, this process combined the daily average traffic flow level data of the road segment to perform maintenance priority weight calculation. The weight calculation logic is as follows: obtain the normalized crack density value, the normalized expansion rate value, and the traffic flow weight coefficient; multiply the normalized crack density value by a preset density weight of 0.4; multiply the normalized expansion rate value by a preset rate weight of 0.3; multiply the traffic flow weight coefficient by a preset flow weight of 0.3; and finally, add the above three products to obtain the final score. For example, if a road segment has a crack density score of 0.8, a propagation rate score of 0.7, and a traffic weight coefficient of 0.9 for a high-traffic segment, then the priority weight calculation result is 0.8*0.4+0.7*0.3+0.9*0.3=0.80. This calculation result shows that by introducing multi-dimensional weighted calculation, the urgency of different road segments can be accurately distinguished. This process generates a road maintenance decision support dataset based on the final score. The specific evaluation data is shown in Table 2.

[0037] Table 2: Example Table of Road Maintenance Priority Assessment

[0038] As shown in Table 2, although the crack density score of SEG_01 section is slightly lower than that of SEG_03, its traffic flow coefficient is extremely high, resulting in a final calculated score as high as 0.80, which is judged as "immediate repair". This verifies the rationality of the resource allocation of this plan.

[0039] Please see Figure 7 A road surface crack image recognition and processing system, comprising: Based on the collected road surface image data, the image processing module performs illumination normalization, noise suppression and multi-scale texture enhancement to generate a standardized image dataset, which is stored in the image feature database. If synchronization fails, it is cached in the local storage unit to obtain the basic image dataset for crack recognition. The candidate region generation module is based on the crack recognition base image dataset. It calculates local contrast and gradient distribution, performs adaptive threshold segmentation and edge enhancement, marks texture interference regions, and obtains crack candidate region dataset. The semantic segmentation optimization module performs pixel-level classification based on the crack candidate region dataset, performs morphological continuity analysis and context-aware completion, and counts the frequency of crack types to obtain the optimized pavement crack identification dataset. The defect assessment module is based on the optimized pavement crack identification dataset, matches spatial location information, queries original inspection records, calculates crack propagation rate and recurrence frequency, filters risk areas, and obtains pavement defect safety assessment dataset. The maintenance decision module calculates crack density and identifies spatial clustering patterns based on the pavement defect safety assessment dataset. It also allocates maintenance priorities by combining traffic flow data and generates a road maintenance decision support dataset.

[0040] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.

Claims

1. A method for recognizing and processing road surface crack images, characterized in that, Includes the following steps: S1: Based on the collected road surface image data, perform illumination normalization and noise suppression, extract texture enhancement image features, generate a standardized image dataset, store the standardized image data, update the global image processing configuration, detect the data synchronization status, and obtain the basic image dataset for crack recognition. S2: Based on the crack recognition basic image dataset, extract the local contrast index and gradient distribution features of the image, screen potential crack candidate regions, perform adaptive threshold segmentation on low contrast regions, perform edge enhancement on high gradient regions, record texture interference regions, and obtain crack candidate region dataset. S3: Based on the crack candidate region dataset, perform pixel-level classification, analyze the continuity of crack morphology, filter irregular morphological regions, perform perceptual completion on small crack regions, analyze crack direction and width distribution, and obtain an optimized road surface crack identification dataset. S4: Based on the optimized pavement crack identification dataset, detect the spatial location information of cracks, extract inspection records, detect the evolution log of cracks at the same location, and filter out areas of cracks that reappear or expand within a short period of time to obtain a pavement defect safety assessment dataset.

2. The road surface crack image recognition and processing method according to claim 1, characterized in that, The basic image dataset for crack recognition includes illumination intensity information, noise level information, texture enhancement coefficient, multi-scale feature map, and synchronization status identifier. The crack candidate region dataset includes local contrast information, gradient magnitude information, edge enhancement parameters, interference region markers, segmentation threshold configuration, and candidate region coordinates. The optimized pavement crack recognition dataset includes crack pixel classification labels, morphological continuity score, context completion identifier, crack direction vector, width distribution histogram, and crack type index. The pavement defect safety assessment dataset includes spatial location coordinates, original crack records, evolution trend identifier, recurrence count, and expansion rate parameter.

3. The road surface crack image recognition and processing method according to claim 1, characterized in that, The specific steps for obtaining the basic image dataset for crack identification are as follows: S101: Based on the collected road surface image data, including the original RGB image, illumination intensity distribution and environmental noise parameters, white balance correction and gamma correction are performed on the image, wavelet denoising is performed to generate a denoised image, multi-scale Gabor filtering is performed on the denoised image, texture responses with different directions and scales are extracted, and multi-scale feature maps are generated by fusion. S102: Based on the multi-scale feature map, calculate the local variance of the channel, apply contrast stretching to the low variance region, perform dynamic range compression on the high variance region, generate a standardized image dataset, store the standardized image dataset in the image feature database, and generate a unique data identifier. S103: Based on the unique data identifier, check the consistency of the data version between the local cache and the global image processing configuration. If the versions are inconsistent, trigger the incremental synchronization mechanism, upload the difference data block and update the version number in the global configuration. If they are consistent, clear the temporary cache and obtain the crack recognition basic image dataset.

4. The road surface crack image recognition and processing method according to claim 3, characterized in that, The specific steps for creating the crack candidate region dataset are as follows: S201: Based on the crack recognition basic image dataset, calculate the contrast standard range and gradient magnitude benchmark range of the local region of the image, identify regions that exceed the preset lower limit of contrast but are lower than the upper limit of gradient, and mark them as low contrast candidate regions. Identify regions whose gradient magnitude exceeds the preset upper limit of gradient and mark them as high gradient candidate regions to obtain an initial candidate region set. S202: Based on the initial candidate region set, apply the Otsu adaptive thresholding algorithm to the low-contrast candidate regions to generate a binary segmentation mask, apply Canny edge detection to the high-gradient candidate regions and superimpose non-extreme suppression to generate an edge enhancement mask, and merge the two types of masks to form a unified candidate region. S203: Based on the unified candidate region, detect whether there are non-crack interference objects of the same type as the crack texture in the region. If so, generate interference region markers based on the texture spectrum features, and record the interference type and location to obtain the crack candidate region dataset.

5. The road surface crack image recognition and processing method according to claim 4, characterized in that, The specific steps for obtaining the optimized pavement crack identification dataset are as follows: S301: Based on the crack candidate region dataset, crop the candidate region image patch, input it into the pre-trained U-Net++ semantic segmentation network, output the crack probability map of each pixel, perform morphological closing operation on the probability map and fill in the break points to generate a preliminary segmentation result; S302: Based on the preliminary segmentation results, calculate the connected components of the crack centerline, perform attention mechanism to complete the components whose length is less than a preset length threshold, predict the missing segments using the crack direction of the adjacent area, and generate a completed crack map. S303: Based on the completed crack map, extract the centerline direction vector and cross-sectional width distribution of each crack, count the frequency of each type of crack in the image, and sort them in descending order of frequency to obtain the optimized road surface crack identification dataset.

6. The road surface crack image recognition and processing method according to claim 5, characterized in that, The calculation of the connected components of the crack centerline refers to extracting the crack centerline with a single pixel width based on the geometric center distribution of the crack region in the preliminary segmentation result, and marking the connectivity of the pixel set connected by eight neighborhoods in the crack centerline to form multiple connected components of the crack centerline.

7. The road surface crack image recognition and processing method according to claim 5, characterized in that, The specific steps for creating the road surface defect safety assessment dataset are as follows: S401: Based on the optimized pavement crack identification dataset, extract the geographic coordinate information of each crack, match it with the road grid database, determine the road segment number and station number, and obtain a spatial positioning list; S402: Based on the spatial location list, query the original records in the crack defect management database, extract the crack area, length and type change data of the same station location in the past inspection cycle, and generate a crack evolution time series; S403: Based on the crack evolution time series, calculate the crack propagation rate per unit time, count the number of times the same location is identified in two consecutive inspection cycles, filter out areas where the propagation rate exceeds a preset rate threshold or the number of recurrences exceeds a preset number threshold, and obtain a road surface defect safety assessment dataset.

8. The road surface crack image recognition and processing method according to claim 1, characterized in that, The method also includes step S5: S5: Based on the road surface defect safety assessment dataset, extract the crack distribution density of risky road sections, analyze the crack aggregation pattern, screen areas where the crack density exceeds the preset density threshold, and combine traffic flow data to sort the maintenance priorities to obtain a road maintenance decision support dataset. The road maintenance decision support dataset includes crack density distribution maps, clustering pattern categories, a list of risky road sections, maintenance priority scores, and traffic flow association identifiers.

9. The road surface crack image recognition and processing method according to claim 8, characterized in that, The specific steps for the road maintenance decision support dataset are as follows: S501: Based on the road surface defect safety assessment dataset, count the number and total area of ​​cracks within a unit length of road segment, calculate the crack density index, compare it with the preset density threshold, filter out road segments that exceed the standard, and generate a set of high-density areas. S502: Based on the high-density area set, the DBSCAN clustering algorithm is applied to identify the spatial clustering pattern of cracks, distinguishing between linear extension type, network crack type and random scattered type. Combined with the daily average traffic flow level of the road segment, maintenance priority weights are assigned to generate a road maintenance decision support dataset.

10. A road surface crack image recognition and processing system, characterized in that, The system is used to implement the road surface crack image recognition and processing method according to any one of claims 1-9, the system comprising: Based on the collected road surface image data, the image processing module performs illumination normalization, noise suppression and multi-scale texture enhancement to generate a standardized image dataset, which is stored in the image feature database. If synchronization fails, it is cached in the local storage unit to obtain the basic image dataset for crack recognition. Based on the crack recognition base image dataset, the candidate region generation module calculates local contrast and gradient distribution, performs adaptive threshold segmentation and edge enhancement, marks texture interference regions, and obtains a crack candidate region dataset. The semantic segmentation optimization module performs pixel-level classification based on the crack candidate region dataset, performs morphological continuity analysis and context-aware completion, and counts the frequency of crack types to obtain the optimized road surface crack identification dataset. Based on the optimized pavement crack identification dataset, the defect assessment module matches spatial location information, queries original inspection records, calculates crack propagation rate and recurrence frequency, filters risk areas, and obtains a pavement defect safety assessment dataset. The maintenance decision-making module calculates crack density, identifies spatial clustering patterns, and allocates maintenance priorities based on the pavement defect safety assessment dataset, generating a road maintenance decision support dataset.

Images