Methods and Systems for Detection and Color Matching Analysis of Asphalt Pavement Repair Areas

Abstract

This invention discloses a method and system for detecting and color matching analysis of asphalt pavement repair areas, belonging to the field of intelligent detection technology for road maintenance engineering. The method includes: identifying the repair area and outputting the bounding box coordinates using the YOLO target detection model; extracting representative RGB color values ​​from the repair area and the original pavement area using the K-Means clustering algorithm; converting the RGB color values ​​to the Lab color space and calculating the color difference ΔE value based on the CIE color difference formula; and storing the detection results in a database to support historical traceability. The system includes an image acquisition module, a target detection module, a color clustering analysis module, a color difference calculation module, a data storage module, and a visualization interaction module. This invention realizes the integrated processing of repair area detection and color matching evaluation, providing quantitative data support for pavement repair quality assessment.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology for road maintenance engineering, and in particular to a method and system for detecting and color matching analysis of asphalt pavement repair areas. Background Technology

[0002] Over long-term use, asphalt pavements develop cracks and potholes due to factors such as vehicle loads, temperature cycles, and water damage, requiring repair work to restore their service function. However, a significant color difference often exists between the repaired area and the original pavement. This color mismatch not only affects the overall aesthetics of the road but may also indicate problems with the repair material mix or construction techniques. Currently, road maintenance management departments primarily rely on manual visual judgment to assess the color matching of repaired areas, which suffers from high subjectivity, low efficiency, and inconsistent assessment standards. In practical engineering applications, manual assessment is further affected by factors such as ambient lighting conditions, differences in the color vision of assessors, and fatigue levels, resulting in poor consistency and repeatability of assessment results, making it difficult to serve as an objective basis for quality control.

[0003] Chinese patent CN113379765A discloses a method for extracting road repair areas and determining repair types. This method uses a YOLOv5 network to coarsely extract repair areas from road images, employs superpixel segmentation to refine the extraction of block and strip repair areas, and utilizes a Gaussian mixture model to identify high-contrast and low-contrast areas to adapt to different lighting conditions. This method can effectively detect and mark repair areas, providing an automated means of identifying road defects. However, this method has the following technical shortcomings: First, it only focuses on the detection and segmentation of repair areas, without addressing color matching analysis between the repair area and the original road surface, thus failing to provide quantitative data support for repair quality assessment; second, it uses grayscale image processing, losing color information and failing to obtain the true color features of the repair area and the original road surface; third, it lacks a persistent storage mechanism for detection results, making it unable to support historical data tracing and statistical analysis.

[0004] To address the aforementioned technical challenges, those skilled in the art urgently need an intelligent analysis method capable of quantitatively assessing color matching levels while simultaneously detecting repair areas. Existing color analysis technologies are primarily applied in textile printing and dyeing, food testing, and other fields, with their application in road maintenance remaining largely unexplored. Specifically, how to effectively integrate target detection technology with color clustering analysis technology, how to establish a color difference assessment model that conforms to human visual perception characteristics, and how to construct a closed-loop system architecture for detection-analysis-storage remain pressing technical problems to be solved. Summary of the Invention

[0005] To address the aforementioned technical problems and shortcomings of existing technologies, this invention provides a method and system for detecting and color matching in asphalt pavement repair areas. By constructing a collaborative framework for deep learning target detection and color clustering analysis, it achieves automatic identification of repair areas and quantitative evaluation of color matching degree, effectively solving the technical problems of strong subjectivity and low efficiency in traditional manual evaluation methods.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] The method for detecting and color matching in asphalt pavement repair areas includes the following steps: Step S1 is the target detection step, which involves acquiring pavement image data and preprocessing it, identifying targets in the repair area using the YOLO target detection model, and outputting the target bounding box coordinates and confidence scores; Step S2 is the color clustering feature extraction step, which determines the repair area and the original pavement area based on the target bounding box, and extracts representative RGB color values ​​for each area using the K-Means clustering algorithm; Step S3 is the color difference quantization calculation step, which converts the RGB color values ​​to the Lab color space, calculates the color difference ΔE value based on the CIE Lab color difference formula, and determines the color matching degree level; Step S4 is the data storage and traceability step, which encapsulates and stores the detection results in a database to support historical traceability and statistical analysis. These four steps form a tight data flow loop, with the output of the previous step serving as the input for the next step, achieving an integrated processing flow of detection, analysis, and storage.

[0008] An asphalt pavement repair area detection and color matching analysis system, including:

[0009] The image acquisition module is used to acquire road surface image data and transmit the road surface image data to the target detection module;

[0010] The target detection module, connected to the image acquisition module, is used to preprocess the road image data and identify the target in the repair area through the YOLO target detection model, and output the target bounding box coordinates and confidence values ​​to the color clustering analysis module.

[0011] The color clustering analysis module is connected to the target detection module. It is used to determine the repair area and the original road surface area based on the target bounding box coordinates, and to extract representative RGB color values ​​through the K-Means clustering algorithm and transmit them to the color difference calculation module.

[0012] The color difference calculation module is connected to the color clustering analysis module. It is used to convert representative RGB color values ​​to Lab color space and calculate the color difference ΔE value. It determines the color matching degree level according to the preset color difference threshold and transmits it to the data storage module.

[0013] The data storage module, connected to the color difference calculation module, is used to receive the detection results and generate detection result records to store in the database storage system;

[0014] The visualization and interaction module is connected to the target detection module, color clustering analysis module, color difference calculation module, and data storage module, respectively, and is used to provide a human-computer interaction interface to display detection results and historical data.

[0015] Compared with existing technologies, the advantages of this invention are as follows: First, by synergistically coupling YOLO target detection with K-Means color clustering algorithm, integrated processing of repair area detection and color feature extraction is achieved, with a single image processing time of no more than 3 seconds; Second, color difference quantization is performed using Lab color space and CIE color difference formula, and the calculation results conform to the characteristics of human eye perception, with a color difference calculation accuracy better than ±0.5; Third, a complete detection-analysis-storage closed-loop mechanism is established, supporting historical data traceability and statistical analysis, providing data support for the optimization of repair coating formulations; Fourth, the system architecture adopts a layered design with a high degree of module decoupling, facilitating functional expansion and system maintenance. Attached Figure Description

[0016] Figure 1 This is a flowchart of the asphalt pavement repair area detection and color matching analysis method described in this invention.

[0017] Figure 2 This is an architecture diagram of the asphalt pavement repair area detection and color matching analysis system described in this invention. Detailed Implementation

[0018] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, the asphalt pavement repair area detection and color matching analysis method provided by this invention includes four core steps, which form a data flow-driven closed-loop collaborative relationship. Each step is described in detail below.

[0020] Step S1: Target detection step in the repair area.

[0021] The technical objective of this step is to automatically identify repair area targets from road surface images and obtain their location information, providing accurate region localization for subsequent color analysis. The entire detection process includes four sub-stages: image acquisition, preprocessing, model inference, and result filtering.

[0022] In the image acquisition sub-stage, the system acquires road image data through a vehicle-mounted camera, drone aerial photography equipment, or a handheld acquisition terminal. Preferably, the present invention uses a color image with a resolution of not less than 1920×1080 pixels as input to ensure the accuracy of subsequent color analysis. In one embodiment, the pixel size of the image acquisition device is set to 4096×2160 pixels, using sRGB color space encoding with a bit depth of 24 bits. During image acquisition, relatively uniform lighting conditions should be ensured to avoid strong shadows or backlighting that could interfere with subsequent processing.

[0023] In the preprocessing sub-stage, the system performs a series of standardization operations on the acquired road image data. First, size normalization is performed, scaling the original image to the input size required by the target detection model. In a preferred embodiment of the invention, the input size is set to 640×640 pixels, and bilinear interpolation is used for image scaling to balance processing speed and detection accuracy. Next, pixel value normalization is performed, mapping pixel values ​​from the [0, 255] interval to the [0.0, 1.0] interval. This operation eliminates brightness differences between different images, improving the model's generalization ability. Finally, channel order conversion is performed, converting the image from the conventional BGR format to RGB format to adapt to the input requirements of deep learning frameworks.

[0024] In the model inference sub-stage, the preprocessed image is fed into the pre-trained YOLO object detection model for inference computation. The YOLO model used in this invention comprises three main components: a feature extraction backbone network, a multi-scale feature fusion network, and a detection head network. The backbone network adopts a CSPDarknet structure, which improves feature extraction efficiency through cross-stage local connectivity design. In one embodiment of this invention, the backbone network contains five convolutional layer groups, each consisting of several stacked CBS and CSP modules. The CBS module contains convolutional layers, batch normalization layers, and SiLU activation functions, with the convolutional kernel size typically set to 3×3 or 1×1. The multi-scale feature fusion network adopts a PANet structure, which achieves the fusion of features at different scales through bidirectional feature transfer from top to bottom and bottom to top, effectively improving the detection capability for targets in patched regions of different sizes. The detection head network is responsible for outputting the final detection results, including target class probability, bounding box regression parameters, and target confidence.

[0025] The multi-scale feature extraction process of convolutional neural networks can be understood as follows: the backbone network generates feature maps at multiple scales at different depths. Shallow feature maps have high resolution but weak semantic information, while deep feature maps have low resolution but rich semantic information. The PANet structure fuses these features at different scales through a feature pyramid approach, enabling the detector to perceive both large and small targets simultaneously. In the practical application scenario of this invention, the size of the target in the repair area varies greatly, ranging from tens of pixels to thousands of pixels. Therefore, multi-scale feature fusion is of great significance for improving detection recall.

[0026] In the result filtering sub-stage, the system post-processes the detection results output by the model. First, non-maximum suppression (NMS) is performed to eliminate overlapping detection boxes. In one embodiment of this invention, the IoU threshold for NMS is set to 0.45. Then, the detection results are filtered based on the confidence score, retaining only targets with a confidence score higher than a preset confidence threshold. This invention sets the preset confidence threshold to a value between 0.5 and 0.8, preferably 0.65. When the confidence score is greater than or equal to 0.65, the detection result is determined to be a valid repair area target; otherwise, the detection result is considered a false detection and filtered out. After filtering, the system outputs the target bounding box coordinates (including four parameters: top-left x-coordinate, top-left y-coordinate, bounding box width, and bounding box height) and the confidence score for the valid repair area targets. This output data will serve as the key input for the next step.

[0027] In one specific embodiment of the present invention, the YOLOv5s model is used as the target detection backbone, and this model is trained using transfer learning on a pavement repair area dataset. The training dataset contains 5000 labeled images, covering various repair types such as block repair, strip repair, and circular repair. During training, the SGD optimizer is used, with an initial learning rate of 0.01, a momentum coefficient of 0.937, a weight decay coefficient of 0.0005, and 300 epochs of training. After training, the model achieves an mAP@0.5 of 87.3% on the test set, meeting the requirements for engineering applications.

[0028] Preferably, the target detection model of the present invention also employs data augmentation techniques to enhance the model's generalization ability. Data augmentation strategies include: random horizontal flipping with a flip probability of 0.5; random brightness adjustment, with brightness varying from 0.8 to 1.2 times the original brightness; random contrast adjustment, with contrast varying from 0.8 to 1.2 times the original contrast; random saturation adjustment, with saturation varying from 0.7 to 1.3 times the original saturation; and Mosaic data augmentation, stitching four training images together to increase the scale diversity of the target. These data augmentation operations can effectively simulate factors such as lighting changes and weather effects in real-world road scenes, improving the model's robustness in complex environments.

[0029] In another embodiment of the invention, model quantization technology can be employed to reduce inference latency and memory usage. Specifically, the trained FP32 precision model is converted into an INT8 precision model, and statistical estimation of quantization parameters is performed using a calibration dataset. The quantized model can improve inference speed by more than 2 times while maintaining a detection accuracy loss of less than 1%, making it particularly suitable for real-time detection scenarios deployed on edge computing devices. The model deployment formats supported by this invention include mainstream inference framework formats such as ONNX, TensorRT, and OpenVINO, facilitating flexible deployment on different hardware platforms.

[0030] Step S2: Color clustering feature extraction step.

[0031] The technical objective of this step is to extract representative color features from the detected repair area and its surrounding original road surface area, providing a data foundation for subsequent color difference calculations. The entire feature extraction process includes three sub-stages: region localization, pixel sampling, and color clustering.

[0032] In the region localization sub-stage, the system determines the area to be analyzed for color based on the target bounding box coordinates output in step S1. For the repair area, the rectangular region defined by the target bounding box is directly used as the analysis range. For the original road surface area, a certain distance needs to be extended outward from the target bounding box to obtain the original road surface pixels around the repair area. In a preferred embodiment of the present invention, the preset extension distance is set to a value between 0.1 and 0.3 times the length of the shorter side of the target bounding box. For example, when the bounding box size is 200×150 pixels, the extension distance can be set to 15 to 45 pixels.

[0033] The specific region expansion algorithm is as follows: Let the coordinates of the upper left corner of the target bounding box be... Width is The height is The expansion factor is The bounding box parameters of the original road surface area are calculated as follows:

[0034] .

[0035] .

[0036] .

[0037] .

[0038] in: , These are the x and y coordinates of the top-left corner of the expanded bounding box, respectively, in pixels; , These are the width and height of the expanded bounding box, respectively, in pixels; The expansion coefficient is dimensionless and its value range is [value range missing]. The preferred value is 0.2; The function ensures that the expanded coordinates do not exceed the image boundaries. The technical advantage of this expansion algorithm is that, through adaptive expansion distance, the original road surface sampling range can be dynamically adjusted according to the size of the repair area, ensuring that enough original road surface pixels are collected for color analysis, while avoiding the introduction of other interference factors due to an excessively large sampling range.

[0039] To obtain valid pixels from the original road surface area, the repair area needs to be excluded from the expanded bounding box region. Specifically, a mask with the same shape as the repair area bounding box is created within the expanded bounding box region. Pixels within the mask region are marked as invalid, and only pixels outside the mask region are analyzed subsequently. This ensures that the pixels in the original road surface area do not contain any repair area pixels.

[0040] In the pixel sampling sub-stage, the system extracts pixel data from the repair area and the original road surface area respectively. For the repair area, the RGB values ​​of all pixels within the target bounding box are extracted. For the original road surface area, the RGB values ​​of all pixels within the extended bounding box that do not belong to the repair area are extracted. In one embodiment of the present invention, when the number of pixels is too large, a uniform sampling strategy can be used to reduce computational complexity. For example, the maximum number of sampled pixels can be set to 10,000, and when the actual number of pixels exceeds this threshold, sample pixels are obtained according to an equally spaced sampling method.

[0041] In the color clustering sub-stage, the system employs the K-Means clustering algorithm to process the sampled pixel data into color clusters. The K-Means algorithm is a classic unsupervised clustering algorithm whose core idea is to divide data points into K clusters, minimizing the sum of the distances between pixels within a cluster and the cluster center. In the application scenario of this invention, each pixel can be considered a data point in three-dimensional space, with its three dimensions corresponding to the values ​​of the R, G, and B color channels, respectively.

[0042] The specific execution process of K-Means clustering is as follows: First, K pixels are randomly selected as initial cluster centers, or a K-Means++ initialization strategy is used to obtain a better initial distribution. Then, an iterative optimization phase begins, with each iteration containing two sub-steps: an assignment step and an update step. In the assignment step, the Euclidean distance between each pixel and each cluster center is calculated, and the pixel is assigned to the nearest cluster. In the update step, the mean of all pixels within each cluster is recalculated, and this mean is used as the new cluster center. The iterative process continues until the convergence condition is met.

[0043] In a preferred embodiment of the present invention, the number of clusters K is set to an integer between 3 and 5, preferably 3. The reason for choosing K=3 is that the color of the asphalt pavement repair area is usually relatively uniform, mainly black or dark gray, while the original pavement may appear light gray or grayish-white due to aging and oxidation. Setting 3 clusters can effectively distinguish between the main color, secondary color, and noise color, extracting representative color features without reducing computational efficiency due to an excessive number of clusters. The maximum number of iterations is set to an integer between 100 and 300, preferably 200, to ensure that the algorithm converges to a stable solution. Clustering convergence threshold. The value is set to a value between 0.0001 and 0.001, preferably 0.0005. Convergence is determined when the change in cluster centers between two adjacent iterations is less than this threshold.

[0044] After clustering is completed, the system counts the number of pixels within each cluster. And calculate the pixel proportion of each cluster. : ,in: Let be the pixel percentage of the k-th cluster, dimensionless, and its value range is . ; The number of pixels contained in the k-th cluster is dimensionless and takes the value of a positive integer. This represents the total number of clusters. The technical meaning of this formula is that the pixel percentage reflects the coverage of the color represented by that cluster within the entire area; a higher percentage indicates that the color is more representative.

[0045] This invention provides two methods for determining representative RGB color values. The first method is the primary color extraction method: the cluster center corresponding to the cluster with the largest pixel proportion is determined as the primary representative RGB color value. This method is simple and direct, suitable for repair areas or original road surfaces with relatively uniform colors. The second method is the weighted fusion method: the cluster centers of each cluster are weighted and fused using pixel proportion as the weight. The calculation formula is as follows: ,in: The weighted representative RGB color values ​​are a three-dimensional vector, and the values ​​of each component range from 1 to 2. ; The pixel percentage of the k-th cluster; Let be the RGB value of the cluster center of the k-th cluster, which is a three-dimensional vector. The technical advantage of this formula is that, through weighted fusion, it can comprehensively consider the distribution of multiple colors within the region, obtaining more overall representative color features.

[0046] In practical applications of this invention, a strategy combining the primary color extraction method and the weighted fusion method is preferred: when the proportion of primary color pixels is greater than 0.6, the primary color extraction method is used; otherwise, the weighted fusion method is used. This adaptive selection strategy can take into account the processing needs of both single-color and complex-color scenes.

[0047] In a preferred embodiment of the present invention, the color clustering analysis further includes a pixel pre-screening step. Since road surface images may contain occlusions, shadow areas, or reflective areas, pixels in these areas can interfere with the color analysis results. Therefore, the present invention screens pixels before clustering analysis: First, pixels with excessively high brightness values ​​(greater than 250) or excessively low brightness values ​​(less than 10) are excluded; these pixels typically correspond to strongly reflective or deeply shadowed areas. Second, pixels with abnormally high saturation (such as colored marking lines) are excluded; these pixels do not belong to the normal color range of asphalt pavement. Finally, pixels are screened based on spatial location, excluding pixels located within 1% of the bounding box edge to avoid boundary effects. The pre-screened pixel data better reflects the true color characteristics of the repair area and the original road surface.

[0048] In another embodiment of the invention, the system also supports multi-scale color analysis. Since there may be color inhomogeneity within the repair area (such as uneven mixing of repair materials or construction marks), single-scale color analysis may not accurately characterize the overall color features. Therefore, this invention employs an image pyramid method for multi-scale analysis: first, color clustering is performed at the original resolution; then, the image is downsampled to 1 / 2 resolution and 1 / 4 resolution for further color clustering; finally, the clustering results from the three scales are combined to determine the final representative color. This multi-scale analysis method can effectively suppress the influence of local color anomalies on the overall analysis results, obtaining a more stable and reliable color feature representation.

[0049] Step S3: Color difference quantization calculation step.

[0050] The technical objective of this step is to quantify the color difference between the repaired area and the original road surface, and output a color difference value that conforms to the characteristics of human visual perception. The entire calculation process includes three sub-stages: color space conversion, color difference value calculation, and matching level determination.

[0051] In the color space conversion sub-stage, the system converts the representative RGB color values ​​obtained in step S2 to the Lab color space. The Lab color space is a device-independent color model proposed by the CIE (International Commission on Illumination) in 1976. Its characteristic is that the calculated color difference has a good correspondence with the color difference perceived by the human eye. The Lab color space contains three components: the L component represents lightness, with a value range of... Where 0 represents black and 100 represents white; the 'a' component represents the red-green axis, with a value range of [value missing]. Positive values ​​lean towards red, negative values ​​lean towards green; the b* component represents the yellow-blue axis, and its value range is... Positive values ​​tend to be yellowish, while negative values ​​tend to be bluish.

[0052] The conversion from RGB to Lab requires two intermediate steps. First, the RGB color values ​​are converted to the XYZ color space: ,in: , , represents the tristimulus values ​​in the XYZ color space, which are dimensionless; , , The input is the RGB color value, and the range of values ​​is... ; For gamma correction function; This is the sRGB standard conversion matrix. The gamma correction function is defined as follows:

[0053] ,in: These are the normalized color component values, with a range of [value range missing]. This piecewise function is used to convert sRGB non-linear encoding into linear values. The sRGB standard transformation matrix is ​​defined as follows:

[0054] .

[0055] This transformation matrix is ​​a standard parameter derived by the CIE Standards Committee based on the characteristics of human color vision.

[0056] Then, the XYZ color values ​​are converted to the Lab color space: , , ,

[0057] in: The value of is the lightness component in the Lab color space, which is dimensionless and has a range of values. ; The red and green axis components are dimensionless and have a range of values ​​of [value range missing]. ; The yellow and blue axis components are dimensionless and have a range of values ​​of [value range missing]. ; , , To reference the XYZ values ​​of the white point, this invention uses a D65 standard light source, with a corresponding reference value of , , ; A linear approximation of the cube root function:

[0058] .

[0059] in: This piecewise function is used to address the numerical stability problem when the XYZ values ​​are close to zero.

[0060] In the color difference calculation sub-stage, the system calculates the color difference between the repair area and the original road surface based on the CIE Lab color difference formula. value:

[0061] ,in: The color difference value is a CIE 1976 Lab standard, dimensionless, and its range is [value range missing]. ; For difference in brightness, among which and These are the brightness values ​​for the repaired area and the original road surface, respectively. The difference between the red and green axis components; This represents the difference between the yellow and blue axis components. The technical principle behind this formula is that the Lab color space is an approximately uniform color space, and the Euclidean distance can effectively reflect the degree to which the human eye perceives color differences.

[0062] According to CIE standards and industrial practice, color difference The correspondence between values ​​and human visual perception is as follows: When At times, the color difference is almost imperceptible; when At times, careful observation can reveal the differences; when At times, moderate color difference; when At times, there are noticeable color differences; when At that time, the color difference was significant.

[0063] In a preferred embodiment of the present invention, in addition to the basic CIE 1976 Lab color difference formula, the system also supports the use of the CIEDE2000 color difference formula for more accurate color difference calculation. The CIEDE2000 color difference formula is an improved version of the color difference formula released by the CIE International Commission on Illumination in 2000. Compared to the CIE 1976 formula, it exhibits better consistency in human visual perception when handling low-saturation colors and blue areas. The CIEDE2000 formula introduces correction factors such as lightness weighting functions, chromaticity weighting functions, hue weighting functions, and interaction terms, which can more accurately reflect the differences in human visual sensitivity to color differences in different color regions. The system of the present invention can select to use either the CIE 1976 formula or the CIEDE2000 formula through configuration parameters. The default is to use the CIE 1976 formula to ensure calculation efficiency, while for applications requiring higher accuracy, the system can switch to the CIEDE2000 formula.

[0064] In another embodiment of the invention, the system also supports channel-specific color difference analysis. Besides calculating the overall color difference... In addition to the value, the system also outputs the brightness difference. Red-green difference Difference between yellow and blue Three component values. This channel-based analysis helps users gain a deeper understanding of the specific sources of color differences between the repaired area and the original pavement. For example, if Larger value and A smaller value indicates a primary difference in brightness, which may be due to variations in the amount of asphalt or mineral powder content in the repair material compared to the original pavement. A large value indicates a hue deviation in the yellow-blue direction, which may be due to the aging degree of the repair material or differences in the proportion of certain components in the formula. Channel-wise color difference analysis can provide more targeted guidance for the precise adjustment of the repair material formula.

[0065] In the matching level determination sub-stage, the system determines the color difference based on the calculated color difference. The numerical value is compared with a preset color difference threshold to determine the color matching level. This invention defines four matching levels: when... When judged as an excellent matching level, it means that the repaired area is highly consistent with the original road surface color, and the repair quality is excellent; when When the match is rated as good, it indicates that the color difference is small and the repair effect is good; when When it is judged as an acceptable matching level, it means that there is a certain color difference, but it is within an acceptable range; when If the color difference is significant, it is judged as a mismatch level, indicating that the repair material formula or construction process needs to be optimized.

[0066] In one embodiment of the present invention, when multiple repair areas exist in the same image, the system calculates the color difference value between each repair area and its corresponding original road surface area, and outputs the color difference analysis results for all repair area targets. Furthermore, an image-level comprehensive color difference index can be calculated, for example, by using a weighted average to fuse the color difference values ​​of each target. ,in: This is the weighted average color difference value; This represents the number of patched-up targets detected in the image. Let be the area (in pixels) of the i-th repair target area; Let be the color difference value corresponding to the i-th repair area. This formula uses area as the weight, so that larger repair areas contribute more to the overall color difference index, which meets the needs of actual engineering evaluation.

[0067] Step S4: Data storage and traceability steps.

[0068] The technical goal of this step is to store the detection and analysis results in a structured manner and establish a complete data traceability mechanism to support historical data querying and statistical analysis. The entire storage process includes three sub-stages: timestamp generation, data encapsulation, and database writing.

[0069] In the timestamp generation sub-stage, the system obtains the current system time and generates a detection timestamp. This invention uses the Unix timestamp format, specifically the total number of seconds from January 1, 1970, 00:00:00 UTC to the current time. The timestamp accuracy is down to the millisecond level, uniquely identifying the time node of each detection operation. Simultaneously with timestamp generation, the system also obtains the target type identifier of the detected image. This identifier is used to distinguish different types of patch areas (such as block patch, strip patch, circular patch, etc.). The target type identifier is typically output by the YOLO model during the inference process in step S1. This invention allows for customization of the target type classification system according to actual application requirements.

[0070] In the data encapsulation sub-stage, the system integrates and encapsulates the data generated in each step into a unified detection result record. The detection result record is structured in JSON format and includes the following fields: unique record identifier (UUID format), detection timestamp (Unix timestamp), source image file path, target sequence number (target number in a single image), target type identifier, target bounding box coordinates (x, y, w, h quadruplets), detection confidence score, RGB color values ​​of the repaired area (triplets), RGB color values ​​of the original road surface (triplets), Lab color values ​​of the repaired area (triplets), Lab color values ​​of the original road surface (triplets), color difference ΔE value, and color matching degree level. This structured data format facilitates subsequent data querying and statistical analysis.

[0071] In the database writing sub-stage, the system writes the encapsulated detection result records into the database storage system. The database storage system of the present invention preferably adopts a relational database, such as SQLite, MySQL or PostgreSQL. The structure design of the data table is as follows: the table name is detection_results, which includes the id field (primary key, auto-incrementing integer), the record_uuid field (record unique identifier, VARCHAR(36) type), the detection_time field (detection timestamp, BIGINT type), the image_path field (image path, VARCHAR(512) type), the target_index field (target sequence number, INTEGER type), the target_type field (target type, VARCHAR(64) type), the bbox_x field (boundary box x coordinate, FLOAT type), the bbox_y field (boundary box y coordinate, FLOAT type), the bbox_w field (boundary box width, FLOAT type), and the b The fields are: box_h (boundary box height, FLOAT type), confidence (confidence level, FLOAT type), patch_rgb_r (patch area R value, INTEGER type), patch_rgb_g (patch area G value, INTEGER type), patch_rgb_b (patch area B value, INTEGER type), road_rgb_r (original road surface R value, INTEGER type), road_rgb_g (original road surface G value, INTEGER type), road_rgb_b (original road surface B value, INTEGER type), delta_e (color difference value, FLOAT type), and match_level (matching level, VARCHAR(32) type).

[0072] The database storage system supports various query operations, including queries by time range, by target type, by matching level, and by color difference range. In one embodiment of the invention, the system provides an SQL query interface, allowing users to execute custom SQL statements to meet flexible data analysis needs. Furthermore, the system supports data export functionality, enabling the export of detection results to common formats such as CSV, Excel, or JSON, facilitating integration with third-party data analysis tools.

[0073] To ensure data integrity and consistency, the database storage system of this invention employs a transaction mechanism for data write operations. Each detection result record is written as an independent transaction, and the transaction is only committed after all fields have been successfully written; otherwise, a rollback operation is performed to avoid incomplete data records. In high-concurrency scenarios, the system uses connection pooling technology to manage database connections, improving data write throughput.

[0074] In one embodiment of the present invention, the data storage module also supports incremental backup functionality. The system automatically creates incremental backup files for the database at preset time intervals (default 1 hour). These backup files are named according to date and time and stored in a designated backup directory. When the primary database fails, data can be quickly restored based on the most recent backup file, with a maximum data loss not exceeding one hour of detection records. Furthermore, the system also supports remote backup functionality, automatically uploading backup files to a cloud storage server for off-site disaster recovery.

[0075] In another embodiment of the invention, the data storage module also integrates data analysis functions. The system can generate statistical reports based on historical inspection data, including: statistically analyzing the distribution trends of each matching level over time to assess changes in the overall road maintenance quality; statistically analyzing the average color difference value by repair type to compare the color matching effects of different repair types; and statistically analyzing the inspection results by geographical location to identify the spatial distribution patterns of color matching problems. These statistical analysis functions can help road maintenance managers grasp the repair quality status from a macro perspective and formulate more targeted maintenance strategies.

[0076] In a preferred embodiment of the invention, the database storage system also supports data archiving and cleanup functions. When the data volume exceeds a preset threshold (e.g., 1 million records), the system automatically archives older historical data to cold storage and deletes the archived data from the main database to maintain query efficiency. The archiving strategy can be configured according to actual needs, for example, retaining data from the most recent year in the main database.

[0077] like Figure 2 As shown, the asphalt pavement repair area detection and color matching analysis system provided by this invention includes six functional modules, which are connected through data interfaces to form a complete data processing link. The system adopts a layered architecture design, including a visualization interaction layer, a business logic layer, and a data layer. The modules are highly decoupled, which facilitates functional expansion and system maintenance.

[0078] The image acquisition module 1 is located at the system's input end and is responsible for acquiring road surface image data and transmitting it to the target detection module. In one embodiment of the invention, the image acquisition module supports multiple image input methods, including local file import, real-time acquisition via USB camera, and network image stream reception. The image acquisition module 1 also includes an image preview function, allowing users to preview the image content immediately after acquisition and confirm that the image quality meets the requirements for subsequent processing before submitting it to the detection process. The core data output of the image acquisition module 1 is road surface image data, in RGB matrix format, with a resolution range supporting 640×480 to 4096×2160 pixels.

[0079] The target detection module 2 is connected to the image acquisition module 1, receives road surface image data, and performs target detection tasks in the repair area. The core function of this module implements the technical solution described in step S1 of the corresponding method embodiment, including three processing stages: image preprocessing, YOLO model inference, and detection result filtering. In one embodiment of the present invention, the target detection module 2 uses GPU-accelerated inference, and the detection time for a single image does not exceed 0.8 seconds (based on an NVIDIA RTX 3060 graphics card test environment). The output data of the target detection module 2 includes a list of target bounding box coordinates and a list of confidence scores, which are simultaneously transmitted to the color clustering analysis module 3 and the visualization interaction module 6.

[0080] The color clustering analysis module 3 is connected to the target detection module 2. It determines the analysis area based on the received target bounding box coordinates and performs color feature extraction. The core function of this module implements the technical solution described in step S2 of the corresponding method embodiment, including three processing stages: region localization, pixel sampling, and K-Means color clustering. The color clustering analysis module also receives raw image data from the image acquisition module to obtain pixel color information within a specified area. The output data of this module consists of representative RGB color values ​​of the repair area and the original road surface area, in a three-dimensional vector format (R, G, B), with each component ranging from 0 to 255 integers.

[0081] The color difference calculation module 4 is connected to the color clustering analysis module 3, receiving representative RGB color values ​​and performing color difference quantization calculation tasks. The core function of this module implements the technical solution described in step S3 of the corresponding method embodiment, including three processing stages: RGB to Lab color space conversion, CIE color difference formula calculation, and matching level determination. The output data of the color difference calculation module 4 includes Lab color values, color difference ΔE values, and color matching degree levels. These data are transmitted to the data storage module and the visualization interaction module. In one embodiment of the present invention, the color difference calculation module 4 uses NumPy vectorized calculation, with a color difference calculation time of no more than 1 millisecond for a single target, supporting batch processing of multiple targets to improve overall processing efficiency. The color difference calculation module 4 also provides a calibration function, allowing users to calibrate the color difference calculation results using a standard color chart to eliminate the influence of imaging device color deviation on the calculation results.

[0082] In a preferred embodiment of the present invention, the color difference calculation module 4 also integrates an ambient light compensation function. Since the lighting conditions in actual road scenes vary considerably, the colors of directly acquired images may deviate from the true colors. Therefore, the present invention adds a white balance correction step before color difference calculation. This step estimates the color temperature of the scene's light source by identifying neutral gray areas in the image, and then compensates for the image colors accordingly. This ambient light compensation mechanism can effectively improve the consistency and accuracy of color difference calculation results under different lighting conditions.

[0083] The data storage module 5 is connected to the color difference calculation module 4, receiving the detection and analysis results and performing data persistence storage tasks. The core function of this module implements the technical solution described in step S4 of the corresponding method embodiment, including three processing stages: timestamp generation, data encapsulation, and database writing. In one embodiment of the present invention, the data storage module 5 uses SQLite as an embedded database engine, with the database file stored on the local disk, supporting efficient storage and querying of millions of records. The data storage module 5 also provides a data export interface, supporting the export of detection results to CSV format for further analysis in tools such as Excel.

[0084] The visualization interaction module 6 is connected to the target detection module 2, color clustering analysis module 3, color difference calculation module 4, and data storage module 5, respectively, and is responsible for providing a human-computer interaction interface to display detection results and historical data. In a preferred embodiment of the present invention, the visualization interaction module uses the PyQt5 framework to build a graphical user interface. The main interface includes four main parts: an image display area, a detection result list area, a color information display area, and a historical data query area. The image display area supports overlaying target bounding boxes and color annotations on the original image, allowing users to intuitively observe the location and color characteristics of the detection results. The detection result list area displays detailed information of all detected targets in the current image in tabular form, including fields such as target number, target type, confidence level, color difference value, and matching level. The color information display area uses a color block visualization method to intuitively display the representative colors of the repair area and the original road surface, allowing users to directly perceive the degree of color difference through color comparison. The historical data query area provides a multi-condition combined query function, allowing users to retrieve historical detection records by time range, target type, matching level, and other conditions, and supports statistical display of query results in chart form.

[0085] In a preferred embodiment of the present invention, the visualization interaction module 6 also provides a real-time monitoring function. When the system continuously processes multiple images in a pipeline mode, the real-time monitoring interface dynamically updates key indicators such as the current processing progress, the number of detected targets, and the average color difference value. The monitoring interface also displays the distribution trend of each matching level over a recent period in the form of a line graph, helping users to promptly detect anomalies. In addition, the system supports setting alarm rules; when the proportion of detected mismatched level targets exceeds a preset threshold, an alarm prompt will automatically pop up and a notification message will be sent.

[0086] In another embodiment of the invention, the visualization interaction module 6 also supports batch export functionality. Users can select multiple historical detection records and export them with one click into a complete report document containing the original image, the detection annotation image, and detailed data. The report document is output in PDF format and includes detection summary information, detailed information tables for each target, and color difference distribution statistical charts, etc., which can be directly used for acceptance document archiving or quality analysis report preparation.

[0087] In the system architecture of this invention, the modules adopt a loosely coupled design principle, and communication between modules is carried out through a unified data interface specification. This design enables the system to have good scalability; for example, the object detection model can be replaced (upgraded from YOLOv5 to YOLOv8) or new color analysis algorithms (such as deep learning-based color semantic segmentation) can be added without modifying the code of other modules. In one embodiment of this invention, the system also supports a plug-in extension mechanism, allowing third-party developers to develop new functional plug-ins according to predefined interface specifications and dynamically load them into the system to extend its functionality. This plug-in architecture greatly enhances the system's flexibility and customizability, meeting the personalized needs of different users.

[0088] The system performance indicators of this invention are as follows: the end-to-end processing time for a single image does not exceed 3 seconds, of which target detection takes about 0.8 seconds, color clustering analysis takes about 1.5 seconds, color difference calculation takes about 0.1 seconds, and data storage takes about 0.2 seconds; the accuracy of patched area recognition reaches over 85%, and the mAP@0.5 is 87.3% on 5000 test images; the accuracy of color difference ΔE calculation is better than ±0.5, verified by calibration and comparison with a standard colorimeter; the maximum number of targets that can be processed simultaneously on a single image is 20, and a batch processing strategy is adopted when this number is exceeded; a single table in the database supports storing more than 1 million records, and the query response time is less than 100 milliseconds.

[0089] In practical applications of this invention, the complete data processing flow can be summarized into four stages: image input, feature extraction, color analysis, and result output. Preferably, a pipelined parallel processing mechanism is adopted between each stage, allowing the color analysis stage of the current image to be performed simultaneously with the feature extraction stage of the next image, thereby improving the overall throughput of the system.

[0090] During the image input stage, the system first performs format verification on the input image to ensure the integrity and readability of the image file. The image formats supported by this invention include common formats such as JPEG, PNG, BMP, and TIFF. For JPEG format images, the system uses a standard JFIF decoder for decompression, preserving the original color space information during decoding to avoid color distortion. For PNG format images, the system supports 8-bit and 16-bit color depth image processing; for 16-bit images, automatic bit depth conversion is performed to unify subsequent processing. In one embodiment of this invention, the system also supports direct reading of RAW format images. Through a professional RAW decoding library, the system can obtain the sensor's raw data, thereby obtaining more accurate color information.

[0091] In the feature extraction stage, in addition to the bounding box information output by the target detection module, the system can also extract auxiliary features to improve the accuracy of color analysis. Preferably, the present invention further calculates the shape features of the target based on the target bounding box, including parameters such as the target's aspect ratio, area ratio, and convex hull area ratio. These shape features can be used to distinguish different types of repair areas, such as block repairs and strip repairs, thereby providing type prior information for subsequent color analysis. For example, strip repairs are usually distributed along the longitudinal direction of the road, and their color distribution may exhibit a gradual change due to the construction direction; block repairs are usually more uniform in color. By combining shape features and color features, the system can obtain more comprehensive description information of the repair area.

[0092] In the color analysis stage, this invention provides several optional color analysis strategies in addition to basic K-Means color clustering. In one alternative embodiment, the system employs a Gaussian mixture model for color modeling, which provides more flexible cluster boundaries compared to the K-Means algorithm. The Gaussian mixture model assumes that pixel colors follow a mixture of multiple Gaussian distributions, and iteratively solves for the parameters of each Gaussian component using an expectation-maximization algorithm. The advantage of this method is that it can automatically determine the optimal number of clusters and handle complex scenarios where color distributions overlap. In another alternative embodiment, the system uses a histogram-based color quantization method, dividing the RGB color space into a predefined number of color bins, and achieving rapid characterization of color distribution by counting the number of pixels in each bin. This method has high computational efficiency and is suitable for applications with high real-time requirements.

[0093] During the results output phase, the system supports multiple output formats to meet the needs of different application scenarios. For manual review scenarios, the system generates annotated visual images, overlaying detection bounding boxes, color blocks, and color difference values ​​onto the original image. For automated integration scenarios, the system outputs structured data in JSON format via a RESTful API interface, facilitating integration with upper-level business systems. For batch processing scenarios, the system supports generating summary report documents, presenting detailed information on each detected target in tabular form, and including statistical charts to show the color difference distribution.

[0094] To ensure stable system operation, this invention incorporates a robust exception handling and fault tolerance mechanism. In the image acquisition stage, if the input image is corrupted or its format is unsupported, the system generates an error log and provides feedback to the user regarding the specific error cause, while simultaneously skipping that image and continuing with subsequent tasks. In the object detection stage, if memory overflow or computation timeout occurs during model inference, the system automatically reduces the input image resolution and retryes inference; if it fails after a maximum of three attempts, the image is marked as a processing failure. In the color clustering stage, if the K-Means algorithm fails to converge within the maximum number of iterations, the system uses the current cluster centers as the final result and adds a convergence status marker to the output data to inform the user. In the data storage stage, if the database connection fails, the system temporarily caches the detection results to local files, automatically performing batch write operations once the database is restored.

[0095] In a preferred embodiment of the present invention, the system also has a self-checking function, which automatically checks key items such as whether the target detection model file is complete, whether the database connection is normal, and whether the GPU device is available during startup. The self-check results are recorded in log form, which facilitates maintenance personnel to quickly locate the cause of system failure. In addition, the system supports configuring alarm thresholds. When the detection success rate is lower than a preset value or the processing queue backlog exceeds the threshold, an alarm notification is automatically sent to remind maintenance personnel to intervene in a timely manner.

[0096] The asphalt pavement repair area detection and color matching analysis method and system of this invention can be widely applied to multiple aspects of road maintenance management. During the repair construction acceptance phase, construction units can use this system to evaluate the color matching of the repaired pavement, quantifying the repair quality level as an acceptance criterion. During routine inspections, maintenance management departments can deploy this system for automated pavement detection, batch analyzing color changes in repair areas to assess the durability of repair materials. In the material formulation optimization phase, researchers can use historical data collected by this system to analyze the color matching patterns between different repair material formulations and the original pavement, guiding adjustments to the repair material formulation.

[0097] To verify the technical effectiveness of this invention, the inventors conducted practical application tests and effect verification in a provincial highway maintenance project. The test dataset contained 2000 high-resolution pavement images, covering various types such as block repair and strip repair, collected from actual road scenes under different lighting and weather conditions. Test results showed that the target detection module achieved a recall rate of 91.2%, a precision rate of 88.7%, and a mAP@0.5 of 87.3% on the test set; the average L2 distance between the representative colors extracted by the color clustering module and the manually labeled primary colors was 8.3 (RGB space), meeting the accuracy requirements for engineering applications; the correlation coefficient between the output results of the color difference calculation module and the standard colorimeter measurement value reached 0.96, and the 95th percentile of the absolute error of color difference was 0.42, better than the design target of ±0.5; the overall system processing speed reached 25 images per minute (based on the NVIDIA RTX3060 test environment), meeting the efficiency requirements for large-scale batch detection.

[0098] Compared to the pavement repair area extraction method disclosed in prior art document CN113379765A, this invention has the following significant technical advantages. At the technical solution level, the prior art document only implements the detection and segmentation of the repair area, and the use of grayscale image processing results in the loss of color information; while this invention adds color clustering analysis and color difference quantification calculation functions on the basis of target detection, realizing a complete technical closed loop of detection-analysis. At the application value level, the output results of the prior art document only inform the user of the location and type of the repair area; while the output results of this invention also include quantified data on the color difference between the repair area and the original pavement, which can directly guide the adjustment and optimization of the repair material formula. At the system architecture level, the prior art document does not address the persistent storage of detection results; while this invention designs a complete data storage and traceability mechanism, supporting the query and analysis of historical data and the generation of statistical reports.

[0099] In summary, this invention, by constructing a collaborative framework of deep learning object detection and color clustering analysis, achieves automatic identification of asphalt pavement repair areas and quantitative evaluation of color matching accuracy. This fills a gap in existing technologies for color matching analysis of repair areas, possessing significant engineering application value and promising industrialization prospects. The technical solution of this invention has been practically applied, verified, and promoted in multiple provincial highway maintenance projects, achieving significant social and economic benefits and providing strong technical support for promoting the intelligent and digital transformation and upgrading of the road maintenance industry.

[0100] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for detecting and color-matching asphalt pavement repair areas, characterized in that, Includes the following steps: Step S1: Target detection in the repair area. Acquire road surface image data and preprocess the road surface image data to obtain a preprocessed image. Input the preprocessed image into a pre-trained YOLO target detection model. Use a convolutional neural network to extract multi-scale features from the preprocessed image and identify targets in the repair area. Output the target bounding box coordinates and confidence scores of the targets in the repair area. Filter valid targets in the repair area based on the comparison results between the confidence scores and the preset confidence threshold. Step S2: Color clustering feature extraction step. Based on the target bounding box coordinates output in step S1, the repair area is determined, and the original road surface area is obtained by expanding outward by a preset distance according to the target bounding box coordinates. A mask with the same shape as the repair area bounding box is created within the expanded bounding box area. Pixels within the mask area are marked as invalid pixels. Only pixels outside the mask area are analyzed to ensure that the pixels in the original road surface area do not contain any repair area pixels. Pixels in the repair area and the original road surface area are pre-screened to exclude pixels with brightness values ​​greater than 250 or less than 10, pixels with abnormally high saturation, and pixels located within 1% of the bounding box edge. K- The Means clustering algorithm performs color clustering on the pixel data of the pre-screened repair area and the original road surface area. Pixels in the repair area and the original road surface area are divided into K clusters based on color similarity. The cluster center of each cluster is calculated to obtain the representative RGB color values ​​of the repair area and the original road surface area. Specifically, the number of pixels in each cluster is counted and the pixel proportion of each cluster is calculated. When the pixel proportion of the cluster with the largest pixel proportion is greater than 0.6, the cluster center corresponding to that cluster is determined as the primary representative RGB color value. Otherwise, the pixel proportion is used as the weight to perform weighted fusion of the cluster centers of each cluster to obtain a weighted representative RGB color value. Step S3: Color difference quantization calculation step. The representative RGB color values ​​of the repair area and the original road surface area obtained in step S2 are converted to the Lab color space to obtain the corresponding Lab color values. The color difference between the repair area and the original road surface area is calculated based on the CIE Lab color difference formula. Numerical values, based on color difference The color matching degree is determined by comparing the numerical value with a preset color difference threshold; wherein, the color matching degree includes: when the color difference... A value less than 2.0 is considered an excellent match; when the color difference... A value greater than or equal to 2.0 and less than 3.5 is considered a good match. When the color difference... A value greater than or equal to 3.5 and less than 5.0 is considered an acceptable matching level. When the color difference... A value greater than or equal to 5.0 is considered a mismatch level. Step S4: Data storage and traceability step, generate detection timestamp and obtain target type identifier, and store detection timestamp, target type identifier, target bounding box coordinates, representative RGB color values ​​and color difference. The numerical values ​​are encapsulated as detection result records, and these records are written into a database storage system to support historical data tracing and statistical analysis.

2. The method for detecting and color matching analysis of asphalt pavement repair areas according to claim 1, characterized in that, The preprocessing operation in step S1 includes: normalizing the size of the road image data to meet the input size requirements of the target detection model; normalizing the pixel value of the normalized image to a preset value range; and converting the channel order of the normalized image to adapt to the channel format requirements of the target detection model.

3. The method for detecting and color matching analysis of asphalt pavement repair areas according to claim 1, characterized in that, In step S1, the preset confidence threshold for the confidence value is set to a value between 0.5 and 0.8; in step S2, the preset expansion distance is set to a value between 0.1 and 0.3 times the length of the short side of the target bounding box.

4. The method for detecting and color matching analysis of asphalt pavement repair areas according to claim 1, characterized in that, In step S2, the number of clusters K in the K-Means clustering algorithm is set to an integer between 3 and 5, the maximum number of iterations is set to an integer between 100 and 300, and the clustering convergence threshold is set to a value between 0.0001 and 0.

001.

5. The method for detecting and color matching analysis of asphalt pavement repair areas according to claim 1, characterized in that, The step S3 of converting representative RGB color values ​​to Lab color space includes: first, converting RGB color values ​​to XYZ color space to obtain XYZ intermediate values, and then converting XYZ intermediate values ​​to Lab color space to obtain Lab color values; wherein the RGB to XYZ conversion uses the sRGB standard conversion matrix, and the XYZ to Lab conversion uses the D65 standard light source white point parameters.

6. The method for detecting and color matching analysis of asphalt pavement repair areas according to claim 1, characterized in that, In step S1, the YOLO target detection model comprises three parts: a feature extraction backbone network, a multi-scale feature fusion network, and a detection head network. The backbone network adopts the CSPDarknet structure, the multi-scale feature fusion network adopts the PANet structure, and the detection head network outputs the target category probability, bounding box regression parameters, and target confidence.

7. The method for detecting and color matching analysis of asphalt pavement repair areas according to claim 1, characterized in that, In step S4, the database storage system adopts a relational database. The data table recording the detection results includes a primary key identifier field, a detection timestamp field, a target type field, a bounding box coordinate field, a repair area RGB color field, an original road surface RGB color field, a color difference value field, and a matching level field.

8. The method for detecting and color matching analysis of asphalt pavement repair areas according to claim 1, characterized in that, When there are multiple repair areas in the same image, the system calculates the color difference value between each repair area and its corresponding original road surface area, and uses a weighted average method to fuse the color difference values ​​of each target to obtain a comprehensive color difference index at the image level, with the area of ​​each repair area target as the weight.

9. An asphalt pavement repair area detection and color matching analysis system, used to implement the asphalt pavement repair area detection and color matching analysis method according to any one of claims 1-8, characterized in that, include: The image acquisition module is used to acquire road surface image data and transmit the road surface image data to the target detection module; The target detection module, connected to the image acquisition module, is used to preprocess the road image data and identify the target in the repair area through the YOLO target detection model, and output the target bounding box coordinates and confidence values ​​to the color clustering analysis module. The color clustering analysis module is connected to the target detection module. It is used to determine the repair area and the original road surface area based on the target bounding box coordinates, create a mask in the expanded bounding box area to exclude pixels in the repair area, perform pre-screening on the pixels, extract representative RGB color values ​​through the K-Means clustering algorithm and transmit them to the color difference calculation module. The color difference calculation module, connected to the color clustering analysis module, is used to convert representative RGB color values ​​to the Lab color space and calculate the color difference. The numerical value determines the color matching degree level based on the preset color difference threshold and is transmitted to the data storage module. The data storage module, connected to the color difference calculation module, is used to receive the detection results and generate detection result records to store in the database storage system; The visualization and interaction module is connected to the target detection module, color clustering analysis module, color difference calculation module, and data storage module, respectively, and is used to provide a human-computer interaction interface to display detection results and historical data.

Images