Highway pavement quality detection method and system
By fusing multi-frame image information and structural entropy analysis, the problems of low efficiency, insufficient accuracy, and poor environmental adaptability in existing technologies for highway pavement detection have been solved, achieving efficient and intelligent pavement quality detection and evaluation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-30
- Publication Date
- 2026-05-12
AI Technical Summary
Existing highway pavement quality testing technologies suffer from problems such as low testing efficiency, strong data subjectivity, high cost, poor environmental adaptability, and insufficient testing accuracy. In particular, they are difficult to identify mild distortions and hidden defects in complex environments.
A multi-frame image information fusion method is adopted. The minimum interference frame group is selected by the inter-frame brightness gradient change, and regional heterogeneous reconstruction and multi-scale structural entropy analysis are performed. Combined with the support vector machine model, the road surface defect area is identified and the quality level is evaluated.
It improves the stability and accuracy of detection, realizes automated closed-loop processing from image acquisition to quality rating, is suitable for intelligent detection of highway maintenance, and enhances the ability to identify minor hidden structural anomalies.
Smart Images
Figure CN122024059A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to a method and system for detecting the quality of highway pavement. Background Technology
[0002] In the maintenance of road traffic infrastructure, the detection and assessment of pavement quality is a crucial task, directly affecting driving safety, road lifespan, and the scientific basis of maintenance decisions. Currently, highway pavement quality detection mainly relies on manual inspection and visual judgment. This method depends on people carrying inspection tools to visually identify or manually record pavement surface defects (such as cracks, potholes, and subsidence). Although the cost is low, the detection efficiency is extremely low, the data is highly subjective, and it is difficult to achieve high-frequency, large-scale real-time detection, which is not suitable for the current development needs of information technology and intelligentization in highway maintenance.
[0003] In recent years, some high-grade highways have introduced laser scanning equipment or 3D structured light modules to acquire road surface elevation data and construct 3D structural models. However, such equipment is expensive, has high requirements for the operating environment, and its ability to detect early-stage hidden defects such as surface color difference and texture disturbance remains insufficient. Furthermore, its data processing flow is complex, relies on high-performance computing resources, and is difficult to widely apply in basic road sections and large-scale highway networks. Compared to the aforementioned traditional methods, image processing-based pavement inspection technology has attracted widespread attention due to its advantages such as low cost, flexible deployment, and strong adaptability. These methods typically use industrial cameras, vehicle-mounted cameras, or mobile terminals to acquire pavement image sequences and identify potential defect areas through image enhancement, edge detection, and texture analysis.
[0004] However, existing image detection technologies still have shortcomings. In real-world road environments, factors such as sudden changes in lighting, camera shake, and motion blur lead to poor image sequence stability, making it prone to false detections or missed detections. Most methods rely on single-frame images and lack multi-view, multi-time-node data fusion mechanisms, making it difficult to identify hidden defects such as minor distortions and boundary cracks. Summary of the Invention
[0005] This invention provides a method and system for highway pavement quality inspection. It is an image processing-based highway pavement inspection method that integrates multi-frame image information, has structural stability modeling capabilities, and supports multi-level quality grading. This method breaks through the bottlenecks of existing technologies in terms of detection accuracy, stability, and scalability, and achieves more efficient and intelligent highway quality inspection and maintenance.
[0006] A method for testing the quality of highway pavement includes the following steps: S1. Acquire the image frame of the road surface to be tested, and construct the minimum interference frame group based on the image time series in the direction of travel. The minimum interference frame group is selected from the image time series according to the principle of minimizing the change in brightness gradient between frames, so as to weaken the dynamic interference of vehicle shadows and instantaneous light spots. S2: Perform regional heterogeneous reconstruction on the minimum interference frame group. Regional heterogeneous reconstruction includes clustering pixels at the same position based on the pixel position information of each image frame in the same frame group to generate candidate distortion regions that reflect local texture stability and edge distortion features. Introduce multi-scale structural entropy analysis. Multi-scale structural entropy analysis includes calculating the structural disorder entropy value of texture direction at multiple scales from pixel level to block level within the candidate distortion regions, extracting the inflection point feature of the structural disorder entropy value changing with scale, and quantifying the inflection point feature into a structural instability index. The road surface distortion response map is generated by fusing the candidate distortion regions and their corresponding structural instability indices. S3: Based on the consistency of gradient direction, morphological closure degree, and structural instability index of each region in the road surface distortion response map, identify the road surface defect region and comprehensively mark and output the road surface quality level.
[0007] Optionally, in step S1, an image acquisition unit mounted on the detection vehicle continuously acquires images of the road surface under test along the direction of travel at fixed time or displacement intervals to form the image time series; each frame in the image time series is grayscaled and its brightness gradient amplitude map is calculated; the change in brightness gradient amplitude between adjacent frames in the image time series is calculated sequentially, and the change is measured by the difference norm of the gradient amplitude of the corresponding pixel.
[0008] Optionally, S1 further includes setting a change threshold, selecting a continuous frame subsequence in the image time series, such that the brightness gradient change between all adjacent frames in the frame subsequence is lower than the change threshold, and the number of frames contained in the frame subsequence reaches a preset number, and constructing the frame subsequence as the minimum interference frame group.
[0009] Optionally, the generation of the candidate distortion region specifically includes: Perform pixel coordinate alignment and normalization on each image frame in the minimum interference frame group; Based on the aligned image frames, for each position in the road surface preset grid, the pixels from all frames at the corresponding position are aggregated to form a set of pixels at the same position. Unsupervised clustering is performed based on the gray value, gradient direction and neighborhood contrast of each pixel at the same position. Sets whose cluster center and the overall attribute difference of the set exceed a predetermined attribute difference threshold are marked as heterogeneous pixel sets.
[0010] Optionally, S2 further includes merging and morphologically processing the spatially adjacent and similar heterogeneous pixel sets in the road surface spatial domain to generate spatially continuous candidate distortion regions; for each candidate distortion region, performing multi-scale structural entropy analysis, defining multiple analysis scales from a single pixel, a local block to the entire region; at each analysis scale, statistically analyzing the distribution of gradient directions of all pixels within the region, calculating the Shannon entropy value of the distribution, and using it as the structural disorder entropy value at the corresponding analysis scale.
[0011] Optionally, the quantification of the structural instability index includes plotting the change curve of the structural disorder entropy value of the candidate distortion region as the analysis scale increases, identifying the inflection point in the change curve where the rate of change of entropy value changes from rapid to slow down, and linearly combining the scale value corresponding to the inflection point, the average slope ratio of the curve before and after the inflection point, and the entropy value at the inflection point to calculate the structural instability index of the corresponding region. Each candidate distortion region is used as a base layer, and its corresponding structural instability index is used as an intensity layer superimposed on this region. Weighted fusion is then performed to generate the road surface distortion response map.
[0012] Optionally, S3 includes extracting the boundary contour of each candidate distortion region from the road surface distortion response map; calculating the gradient direction consistency of edge pixels within each boundary contour; and simultaneously calculating the ratio of the area of each contour to the area of its smallest circumscribed convex polygon as its morphological closure index.
[0013] Optionally, S3 further includes establishing a defect identification model, wherein the input features of the defect identification model are the gradient direction consistency, morphological closure index, and structural instability index of the corresponding region; and using a pre-trained support vector machine as the defect identification model to output the road surface defect area.
[0014] Optionally, all identified pavement defect areas are sorted and graded according to the magnitude of their structural instability index, and divided into multiple defect levels; the total defect area of each defect level and its spatial distribution density within the test section are calculated; based on the highest level of all defect areas within the test section, the proportion of the total defect area, and the spatial distribution density, the pavement quality level is calculated and output as four levels: excellent, qualified, poor, and dangerous, according to a predefined comprehensive scoring rule.
[0015] A highway pavement quality inspection system for implementing the aforementioned highway pavement quality inspection method includes the following modules: Image acquisition module: used to acquire time series images of the road surface under test; Image preprocessing module: used to construct a minimum interference frame group from the image time series based on the principle of minimizing inter-frame brightness gradient changes; The region reconstruction and structural entropy analysis module is used to perform pixel clustering at the same location on the minimum interference frame group, generate candidate distortion regions, and perform multi-scale structural entropy analysis to extract the structural instability index and construct the road distortion response map. Defect identification and quality level assessment module: used to identify defect areas based on the gradient direction consistency, morphological closure degree and structural instability index of each candidate area in the distortion response map, and output the pavement quality level result.
[0016] The beneficial effects of this invention are: This invention introduces a brightness change rate detection mechanism during the road image acquisition stage, selecting only image sequences with stable brightness changes to form the minimum interference frame group. This significantly suppresses image interference caused by factors such as ambient light fluctuations, motion blur, or jitter, effectively ensuring the accuracy and stability of subsequent feature extraction and image fusion processes. This mechanism, combined with the Sobel operator to extract brightness gradient changes, forms an adaptive frame selection strategy, enabling the construction of a high-quality analytical foundation even under complex road conditions and poor lighting.
[0017] Compared with existing technologies that only use single frames or simple image stacking to extract road surface defect features, this invention proposes a same-location heterogeneous pixel aggregation mechanism. After image registration and perspective transformation, it fuses the corresponding position pixel data of multiple time point images under a unified coordinate system. Combined with an unsupervised clustering algorithm based on brightness / texture / edge direction attributes, it accurately distinguishes potential structurally heterogeneous pixel regions, effectively enhancing the sensitivity and coverage of identifying slight latent structural anomalies.
[0018] This invention introduces a structural instability index calculation method, comprehensively considering key features such as gradient direction consistency, closure index, and structural disturbance degree. Furthermore, it utilizes a support vector machine model to classify and identify defect regions. Finally, it combines the defect region area and distortion level response map to construct a quality-level-oriented classification model, achieving fully automated closed-loop processing from raw image acquisition and defect detection to quality rating. Compared to traditional methods relying on manual interpretation or single-indicator statistics, accuracy and consistency are significantly improved, making it particularly suitable for intelligent detection scenarios in highway pavement maintenance. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a schematic diagram of the detection method flow according to an embodiment of the present invention; Figure 2 This is a block diagram illustrating the execution logic of the method in an embodiment of the present invention. Detailed Implementation
[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.
[0022] like Figures 1-2 As shown, a method for testing the quality of highway pavement includes the following steps: S1. Acquire the image frame of the road surface to be tested, and construct the minimum interference frame group based on the image time series in the direction of travel. The minimum interference frame group is selected from the image time series according to the principle of minimizing the change in brightness gradient between frames, so as to weaken the dynamic interference of vehicle shadows and instantaneous light spots.
[0023] S1 specifically includes: S11. In road surface detection, if the intervals between acquired images are unstable, problems such as inconsistent image overlap rates between frames, drastic changes in target scale, and abrupt changes in viewpoint can easily occur. This can interfere with subsequent image processing steps, especially: Gradient magnitude comparison calculation: If the viewing angle difference between two frames is too large, the calculated brightness gradient change may reflect more the parallax difference than the change in road surface condition. Inaccurate judgment of dynamic interference: Dynamic interference such as vehicle shadows and reflections are only easily identified and eliminated in continuous and stable image sequences.
[0024] Therefore, images must be acquired at a constant pace to ensure that variations between images primarily originate from the actual condition of the road surface itself, rather than noise introduced by the acquisition conditions. Specifically, an image acquisition unit mounted on a detection vehicle continuously acquires road surface images along the direction of travel at fixed time intervals Δt or displacement intervals Δd, resulting in an image time series. ;in, Represents an image time series. Indicates the first Frame image, This represents the total number of frames.
[0025] The time interval Δt is suitable for situations where the vehicle speed is relatively constant, ensuring a consistent sampling rhythm and a smooth time series of generated images. The displacement interval Δd is suitable for detection tasks with strict requirements on the spatial distribution characteristics of the road surface, ensuring that the road surface area covered by each frame of the image has an equidistant sampling attribute, which is beneficial for spatial feature analysis.
[0026] Since the detection vehicle travels along the road, and the image acquisition system is positioned facing forward or directly downward, the image sequence naturally aligns with the vehicle's trajectory. Acquiring images along the direction of travel helps to reconstruct the continuous evolution trend of the road surface structure; analyze quality changes in different areas along the route; and then perform temporal comparisons and local anomaly detection.
[0027] S12. In a raw color image, each pixel typically contains three color channels: red, green, and blue. However, color information is not sensitive to structural texture, while structural information, such as cracks and pit edges, is mainly reflected in abrupt changes in brightness. Therefore, grayscale processing converts a color image into a single-channel image, retaining only brightness information (i.e., grayscale values) for easier subsequent processing. The brightness gradient describes the degree of brightness change with spatial location. The greater the brightness change at a point, the more likely it is to be located at a structural edge or in a region of abrupt texture change. The brightness gradient magnitude map represents the intensity of brightness change at each pixel in the entire image as a numerical value, used for subsequent inter-frame comparison.
[0028] Specifically, this means for each frame of the image Perform grayscale conversion to obtain a grayscale image. And calculate its brightness gradient magnitude map. Its definition is: ; in, Indicates the first The grayscale value of the frame image, Indicates the first Pixels in a frame gradient magnitude, This represents the rate of change of brightness in the horizontal and vertical directions. The rate of change of brightness is obtained based on the Sobel operator, an image edge detection operator used to approximate the rate of change of brightness in the horizontal (x-axis) and vertical (y-axis) directions of an image. Specifically: S21. Perform a convolution operation on the image, using two Sobel templates, i.e., Sobel kernels, to calculate the rate of change in two directions, including the Sobel kernel for the horizontal gradient and the Sobel kernel for the vertical gradient.
[0029] S22. The convolution output is the rate of change of brightness of a pixel in that direction. (This refers to the grayscale image.) Convolutions with the two templates above yield the following results: : Indicates the rate of change of brightness of the pixel in the x-direction; : Indicates the rate of change of brightness of the pixel in the y direction.
[0030] 3. Calculate the magnitude of the brightness gradient, i.e., the combined intensity of the rates of change in the two directions: ; Alternatively, for efficiency reasons in practice, approximate calculations can be used: .
[0031] In S13, each frame of the image in S12 is processed into a brightness gradient magnitude map. This represents the intensity of brightness change for each pixel in an image. Each frame generates such a two-dimensional matrix, where each element represents the degree of brightness change for that pixel. Therefore, by analyzing the gradient magnitude maps of adjacent frames in an image time series, the change in brightness gradient magnitude between frames can be calculated. Its definition is: ;in, For the first With the The amount of brightness gradient change between frames, The Frobenius norm is the square root of the sum of the squared differences between all pixels. Indicates the first With the Gradient magnitude map of the frame.
[0032] If the changes between two frames are small, it indicates that the image shooting conditions were stable and there were no drastic changes in lighting, vehicle shadows, reflections, or other interfering factors. Conversely, if the brightness of certain areas suddenly changes drastically, such as when a vehicle shadow sweeps by, there will be significant differences between their gradient magnitude maps. Therefore, the above method compares frames one by one to determine which frames have stable changes and less interference, thus selecting the cleanest image frame group for subsequent analysis.
[0033] S14, the aforementioned calculation of the brightness gradient change between each pair of adjacent image frames in the image sequence. This is used to measure the visual differences between frames, but to ensure that the selected frame groups are stable enough, a criterion must be set. That is, only when the brightness variation between all consecutive frames is less than a certain threshold are they considered low-interference frames. Therefore, a brightness variation threshold is set. In image time series In the given information, find a continuous subsequence: ; The following conditions must be met: For any They all ; Subsequence length ,in This is the preset minimum number of frame groups.
[0034] Frame subsequences that meet the above conditions These are constructed as minimal interference frame groups for subsequent road surface distortion detection. Brightness change threshold. Based on adaptive settings using statistical characteristics, this includes calculating all... mean and standard deviation Set the threshold to ; It is an empirical coefficient with a value in the range of 0.5-1.5. The above settings are used to adapt to the overall fluctuation level under different scenarios and avoid the threshold being too wide or too narrow.
[0035] Furthermore, even if we find a set of frames with very small changes, if the number of frames is too small, it will not provide enough information for subsequent analysis, such as spatial structure recognition and multi-scale structure entropy extraction. Therefore, it is necessary to set a minimum frame number threshold, which is the so-called preset minimum number of frame groups. The minimum number of frame groups is set based on subsequent processing requirements. If subsequent structural entropy analysis needs to cover at least three scales, such as pixel blocks, local windows, and the entire frame, then it is recommended to... If both stability and redundancy cancellation capabilities are considered, it is recommended to filter out occasional noise or interference from vehicle body obstruction. Values That is, 7 consecutive and stable images are selected as the minimum interference frame group, while taking into account data integrity.
[0036] S2: Perform regional heterogeneous reconstruction on the minimum interference frame group. Regional heterogeneous reconstruction includes clustering pixels at the same position based on the pixel position information of each image frame in the same frame group to generate candidate distortion regions that reflect local texture stability and edge distortion features. Introduce multi-scale structural entropy analysis. Multi-scale structural entropy analysis includes calculating the structural disorder entropy value of texture direction at multiple scales from pixel level to block level within the candidate distortion region, extracting the inflection point feature of the structural disorder entropy value changing with scale, and quantifying the inflection point feature into a structural instability index. The road surface distortion response map is generated by fusing the candidate distortion regions and their corresponding structural instability indices.
[0037] S2 specifically includes: S21, for each frame in the minimum interference frame group Pixel-level coordinate alignment is achieved through perspective projection transformation, and grayscale values are adjusted accordingly. Perform naturalization: ; in, This represents the normalized grayscale image. Indicates the first The grayscale mean and standard deviation of the frame. Represents the unified coordinates after the transformation and alignment.
[0038] Perspective projection transformation, also known as homography transformation, is a commonly used method in image registration. It can transform any quadrilateral region in one image into a target region in another image, simulating the projection geometry under changes in camera viewpoint. A road surface is a relatively flat two-dimensional plane. Images captured by a camera from different angles may have varying viewpoints, but these changes can be mathematically described by a fixed transformation matrix.
[0039] S22, Same-location pixel clustering and heterogeneous discrimination: The core is to perform pixel clustering and heterogeneous discrimination at each fixed position in the aligned image frame. The pixels are analyzed to determine whether the pixel attributes at this location are stable and consistent across different time frames. If the location shows significant inconsistency across different image frames, it may indicate the presence of abnormal texture structures or deformation interference, and thus be considered a candidate point for a possible road surface distortion area.
[0040] Specifically, in the aligned image sequence, for each position in the preset road surface grid... Extract the set of pixels at the same position across all frames: ; in, Indicates the first The pixel attribute vector at that location in the frame image. This represents the spatial coordinates of the aligned image. This is the grayscale value of the pixel. This indicates the gradient direction of the pixel. This indicates the contrast of the pixel's neighborhood. This represents the set of pixels at the same position.
[0041] Furthermore, since we cannot know in advance whether the pixel at a given location will behave consistently across multiple frames, we cannot use fixed rules to determine this. Instead, we employ unsupervised clustering. If the attributes of pixels at that location are highly similar across different frames, the clustering result will be highly concentrated, meaning most points will cluster into one class, indicating strong consistency. Conversely, if the pixel attributes at that location differ significantly, the clustering result will show multiple, more dispersed clusters, indicating that the pixel's behavior is unstable across different frames, potentially due to texture distortion, dynamic interference, or other issues. Therefore, we use an unsupervised clustering algorithm to cluster the set of pixels at the same location. Clustering is performed; unsupervised clustering can automatically identify pixel locations with inconsistent attributes as clues to potential distortion regions. Specifically, this includes defining cluster centers as... If any class satisfies: Then this location is marked as a heterogeneous pixel set, where, The overall mean of the pixel set. Represents Euclidean distance. The attribute difference threshold is set through statistical adaptive adjustment, and all... From the overall covariance matrix, extract the standard deviation of the principal component directions, and set a threshold as follows: ;in, This represents the global standard deviation of all pixel attribute vectors in the entire image sequence, used to measure the normal fluctuation range of overall pixel attributes. The system's sensitivity to anomalies is adjusted accordingly. The aforementioned co-located pixels refer to pixels at the same spatial location in different time frames, i.e., a set of pixels at the same location but in different frames, reflecting temporal differences; heterogeneous pixels refer to these co-located pixels that exhibit significant or inconsistent differences in attributes such as grayscale, orientation, and contrast, displaying abnormal characteristics, and are the result of clustering judgment.
[0042] The unsupervised clustering algorithm using K-means has the following steps: 1. Input data: Put the set Each pixel vector in As a three-dimensional point, it is input into the algorithm.
[0043] 2. Specify the number of clusters K: K-means requires specifying the number of cluster centers K in advance.
[0044] 3. Perform clustering: The algorithm iterates continuously based on Euclidean distance to update the cluster centers. Ultimately, this results in K classes.
[0045] 4. Determine if there is heterogeneity: If the difference between a cluster center and the overall mean exceeds the set attribute difference threshold, then there is significant heterogeneity at that location.
[0046] S23, Candidate Distortion Region Generation: Previously, some heterogeneous pixels were identified in the image. These pixels indicate inconsistencies in the behavior of certain locations across different frames, potentially caused by road surface anomalies or lighting disturbances. However, these points are usually discrete and isolated, and direct analysis could lead to noise or misjudgments. Real road surface structures are often continuously distributed, so it's necessary to aggregate these discrete points into connected regions for more effective subsequent structural stability analysis. Therefore, S23 aggregates all spatially adjacent heterogeneous pixels with similar properties. By performing aggregation and morphological processing, spatially continuous regions are generated. , as candidate distortion regions. Among them: S231, the aggregation rule is based on whether the difference in feature vectors between adjacent pixels is less than a threshold. Aggregation is the process of merging heterogeneous sets of adjacent pixels with similar attributes into a candidate region. There are two basic rules for aggregation: first, spatial proximity, meaning the two pixels must be adjacent in the image, such as vertically, horizontally, or diagonally; second, attribute similarity, meaning the difference between the attribute vectors of the two pixels cannot be too large, i.e., their grayscale values, gradient directions, and contrast characteristics must be similar. When determining attribute similarity, an attribute difference threshold is used. If the attribute difference between two pixels is less than this threshold, they are considered to belong to the same candidate region. Based on these two conditions, the region can be continuously expanded, absorbing surrounding pixels that meet the conditions, until it can no longer be expanded.
[0047] S232, after aggregation, due to image noise or local interference, the resulting region may still have irregular boundaries, small holes or breaks, and many isolated small dots. To solve these problems, image morphological operations are used to further clean and regularize the region shape. These operations are not based on changes in pixel values, but on processing based on the pixel spatial structure. Morphological operations include: Expansion: Expanding the boundary of the region outward to fill the small hole; Erosion: Shrinks the boundary of the region and removes small spots at the edges; Opening operation: Erosion followed by dilation to remove minor noise; Closing operation: First expand, then corrode, to fill small cracks; Connected component extraction: Identify and retain large connected regions while removing scattered small pieces.
[0048] These operations can make the shape of the candidate distortion region more regular, the edges smoother, and better match the spatial distribution characteristics of real road defects.
[0049] S24, Multi-scale structural entropy calculation: In actual road images, some defects, such as fine cracks, are only visible in a very small local area, while other defects, such as potholes and structural collapses, can only be identified in a larger area. In order to comprehensively determine whether there is structural anomaly in a region, it is necessary to analyze it from multiple observation scales, which leads to multi-scale structural entropy analysis, that is, to measure whether the texture direction of the region is consistent under multiple analysis windows of different sizes.
[0050] Specifically, for each candidate region Define the analytical scale set: ; At each scale Below, the area Divide the data into sub-blocks or sliding windows, calculate the gradient direction distribution histogram of pixels within each sub-block, and determine its structural disorder entropy value: ;in, For the first The frequency of each directional interval. The structural disorder entropy value is essentially a quantitative description of the disorder level of pixel gradient direction distribution within a region. If the texture directions within a region are very consistent, the gradient direction distribution will be concentrated in a certain directional interval, resulting in a smaller entropy value. If the structure within a region is chaotic and lacks obvious directionality, the directional distribution will be dispersed, resulting in a larger entropy value. Therefore, the larger the structural entropy value, the more disordered and unstable the texture of this region.
[0051] To calculate structural entropy, we first need to know how the gradient directions of pixels in a region are distributed. Therefore, we first construct a histogram of direction distribution, which is a statistical structure that divides the gradient direction angles into several directional intervals. Then, we count which bins the gradient directions of all pixels at that scale fall into and their frequencies, resulting in a normalized direction probability distribution (the sum of all bins equals 1), i.e., the gradient direction distribution histogram. That is, the first in the histogram The normalized probability values of each bin.
[0052] S25, Inflection Point Extraction and Structural Instability Index Calculation: draw With scale Changing structural entropy curve Specifically, by using scale as the horizontal axis and structural entropy as the vertical axis, a curve is obtained showing how entropy changes with scale. This curve reflects how the degree of order in the regional structure changes as the observation scale increases, and thus, an inflection point is defined. Where the rate of increase in entropy slows down significantly, the following condition is met: ; This is the second derivative threshold, used to determine whether an inflection point holds. The formula indicates that the second derivative is negative and less than 1 / 3. The curve shows a sharp bend that transitions to a gentler curve. This inflection point is a crucial signal of structural instability. If a region has a simple texture, the entropy curve will likely appear very flat. However, if a region contains complex structures, the degree of structural disorder increases rapidly before a certain scale, while the entropy growth slows down after that scale. This point represents the inflection point, where the rate of increase slows significantly. After defining the inflection point, the following three features are further extracted: Feature 1: Inflection Point Scale This represents the observation scale where structural abrupt changes are most pronounced; it is obtained by analyzing the changing trend of the structural entropy curve, essentially finding the location where the rate of entropy growth significantly slows down. Simply put, the second difference is calculated at each scale point, and the point with the fastest decreasing rate of change, the smallest value, and less than a set threshold is selected. That scale, as the inflection point scale; Feature 2, inflection point slope ratio: ;in, The average slope over several scales before the inflection point. This represents the average slope over several scales after the inflection point; a large ratio indicates rapid growth at the beginning and slowdown at the end, suggesting a stronger sense of structural instability. Feature 3, inflection point entropy value: , which represents the degree of disorder in a region at the inflection point scale.
[0053] Final structural instability index Defined as: ;in These are the weighting coefficients.
[0054] Obtaining the second derivative threshold involves first calculating the standard deviation of the second difference result of the entire entropy curve. Then set: ;in It is an adjustment coefficient with a value between 1.5 and 3.0. It can automatically adjust the threshold according to the fluctuation of different curves and has strong adaptability.
[0055] S26, Road surface distortion response map generation: Generate all candidate regions Projected onto the response map as a region layer, along with its corresponding structural instability index. As an intensity layer, a weighted fusion method is used to generate the road surface distortion response map. : ;in, Indicates an indicator function, if exist In the given value, the value is 1; otherwise, it is 0.
[0056] Specifically, S26 involves all candidate regions. The location is mapped back to the original image, which is to draw it on the image; the pixel corresponding to the location of each region is assigned the value of the structural instability index of that region, thus forming an intensity layer; if multiple regions overlap, their intensity values can be merged according to the maximum or average value; finally, a two-dimensional grayscale image is generated, and each pixel value reflects the distortion intensity of the road surface at its corresponding location, which is called the road surface distortion response map.
[0057] S3: Based on the consistency of gradient direction, morphological closure degree, and structural instability index of each region in the road surface distortion response map, identify the road surface defect region and comprehensively mark and output the road surface quality level.
[0058] S31, Structural Feature Extraction: Response Map to Road Surface Distortion Each candidate distortion region in Perform the following processing: S311, Boundary Contour Extraction: Extracting Region Boundaries Using Edge Detection and Connectivity Analysis , represents a set of pixels: ;in, The total number of pixels on the outline. The first on the boundary contour The coordinates of each pixel are used. Specifically, edge detection is first performed on the candidate distortion region to find the locations in the region where the grayscale or response value changes significantly. Then, through connected component analysis, the outer connected edge of the region is determined. Finally, the outermost ring of pixels is retained as the boundary contour of the region.
[0059] S312, Gradient Direction Consistency Calculation: For each pixel on the region boundary, calculate its gradient direction. Define a gradient direction consistency index: The closer the value is to 1, the more consistent the directions; the closer it is to 0, the more chaotic the directions. Indicates the first boundary contour The gradient direction angle of each pixel Indicates the gradient direction The direction vector is mapped onto the unit complex plane. The gradient direction is the direction in which the brightness change at this location mainly changes. If it is a real defect, the gradient directions on the boundary are often regular and close to each other, and the whole looks like a line or a closed contour. If it is noise or a pseudo-anomaly, the boundary directions will be very messy, and the gradient directions will differ greatly from each other.
[0060] The gradient direction calculation for each pixel on the boundary includes calculating the brightness changes in the horizontal and vertical directions at the boundary pixel based on its neighborhood gray values. Based on the ratio of these two changes, the main direction of the brightness change of the pixel is calculated, which is the gradient direction of the boundary pixel.
[0061] S313, Morphological Closure Index: The actual outline area of the region is denoted as... Its minimum circumscribed convex polygon area is The degree of closure is defined as: Specifically, all pixels on the boundary of the region are taken as a point set, and a convex hull is constructed on this point set to obtain the smallest convex polygon that can completely enclose the region and has no indentation. The area of this convex polygon is calculated, which is the area of the smallest circumscribed convex polygon.
[0062] S32, Defect Identification and Judgment: Constructing a Defect Identification Model The following feature vectors are used as input: ;in, For gradient direction consistency activities, As an index of closure degree, This is the structural instability index.
[0063] The defect identification model uses a support vector machine classifier for defect identification.
[0064] For each candidate distortion region Extract the following three types of structural features to form the input vector: Gradient direction consistency : Reflects the degree of directional concentration of the regional boundary; Morphological closure index : Reflects the compactness of the region's outline; Structural instability index It reflects the degree of abrupt changes in the regional texture structure at multiple scales.
[0065] Combine the three elements into a three-dimensional feature vector. .
[0066] Using an existing labeled sample dataset, each sample represents a candidate region, and each sample includes its corresponding feature vector. and manually labeled ,in: This indicates that the area represents a real defect; This indicates that the region is either non-defective or noisy. A support vector machine (SVM) classification algorithm is used to train the samples, constructing a boundary discrimination model. .
[0067] When performing prediction: for candidate regions of new input Extract its corresponding feature vector ,Will The input is fed into the trained classification model, which then outputs a prediction result. If the model output is 1, then the area is determined to be a defect area. ; If the output is 0, it is a non-defect area.
[0068] S33, Defect Level Classification and Statistics: For all identified defect areas... According to the structural instability index Classify: ; For each level : Total area: ; Statistical spatial distribution density: ; This indicates the number of defect areas with a grade of L.
[0069] It is the dividing threshold between minor and moderate defects, using the structural instability index. In the distribution statistics, the mean μ plus 0.5 standard deviations is selected. : ;Pick This is to identify roughly the first 70% to 80% of stable areas as mild, with the remainder gradually transitioning to moderate and severe. It is the dividing threshold between moderate and severe defects, and the structural instability index is selected. The 90th percentile; the quantile method is used to identify the most severe 10% or so areas for high-priority maintenance / early warning.
[0070] S34 is based on the following three macroeconomic indicators: S341, highest defect level The value can be one of mild, moderate, or severe. S342, Defect area percentage: ; It refers to the area of the entire image or the entire detection region. Indicates the first The area of each defect region is obtained by pixel area statistics. After defect identification, the pixel set of each defect region has been marked. The number of all pixels in the region is counted and multiplied by the actual area represented by a single pixel to obtain the area.
[0071] S343, average defect space density: .
[0072] The overall road surface quality of the inspected road section is divided into four levels, using the following scoring rules: serious or The output level is: Dangerous; medium and The output level is: second difference; Mild and and The output level is: Pass; Other cases; output level: Excellent.
[0073] A highway pavement quality inspection system, used to implement the above-mentioned inspection methods, includes the following modules: Image acquisition module: used to acquire time series images of the road surface under test; Image preprocessing module: used to construct a minimum interference frame group from the image time series based on the principle of minimizing inter-frame brightness gradient changes; The region reconstruction and structural entropy analysis module is used to perform pixel clustering at the same location on the minimum interference frame group, generate candidate distortion regions, and perform multi-scale structural entropy analysis to extract the structural instability index and construct the road distortion response map. Defect identification and quality level assessment module: used to identify defect areas based on the gradient direction consistency, morphological closure degree and structural instability index of each candidate area in the distortion response map, and output the pavement quality level result.
[0074] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.
[0075] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for testing the quality of highway pavement, characterized in that, Includes the following steps: S1, acquire the image frame of the road surface to be tested, and construct the minimum interference frame group based on the image time series in the direction of travel. The minimum interference frame group is selected from the image time series based on the principle of minimizing the inter-frame brightness gradient change. S2: Perform regional heterogeneous reconstruction on the minimum interference frame group. Regional heterogeneous reconstruction includes clustering pixels at the same position based on the pixel position information of each image frame in the same frame group to generate candidate distortion regions that reflect local texture stability and edge distortion features. Introduce multi-scale structural entropy analysis. Multi-scale structural entropy analysis includes calculating the structural disorder entropy value of texture direction at multiple scales from pixel level to block level within the candidate distortion regions, extracting the inflection point feature of the structural disorder entropy value changing with scale, and quantifying the inflection point feature into a structural instability index. The road surface distortion response map is generated by fusing the candidate distortion regions and their corresponding structural instability indices. S3: Based on the consistency of gradient direction, morphological closure degree, and structural instability index of each region in the road surface distortion response map, identify the road surface defect region and comprehensively mark and output the road surface quality level.
2. The method for detecting the quality of highway pavement according to claim 1, characterized in that, In step S1, an image acquisition unit mounted on the detection vehicle continuously acquires images of the road surface under test along the direction of travel at fixed time or displacement intervals, thereby forming the image time series. Each frame of the image time series is converted to grayscale and its brightness gradient magnitude map is calculated. The change in brightness gradient magnitude between adjacent frames in the image time series is calculated sequentially, and the change is measured by the difference norm of the gradient magnitude of the corresponding pixel.
3. The method for detecting the quality of highway pavement according to claim 2, characterized in that, S1 further includes setting a change threshold, selecting a continuous frame subsequence in the image time series, such that the brightness gradient change between all adjacent frames in the frame subsequence is lower than the change threshold, and the number of frames contained in the frame subsequence reaches a preset number, and constructing the frame subsequence as the minimum interference frame group.
4. The method for detecting the quality of highway pavement according to claim 1, characterized in that, The generation of the candidate distortion region specifically includes: Perform pixel coordinate alignment and normalization on each image frame in the minimum interference frame group; Based on the aligned image frames, for each position in the road surface preset grid, the pixels from all frames at the corresponding position are aggregated to form a set of pixels at the same position. Unsupervised clustering is performed based on the gray value, gradient direction and neighborhood contrast of each pixel at the same position. Sets whose cluster center and the overall attribute difference of the set exceed a predetermined attribute difference threshold are marked as heterogeneous pixel sets.
5. A method for detecting the quality of highway pavement according to claim 4, characterized in that, S2 further includes merging and morphologically processing the spatially adjacent and similar heterogeneous pixel sets in the road surface spatial domain to generate spatially continuous candidate distortion regions; for each candidate distortion region, performing multi-scale structural entropy analysis and defining multiple analysis scales from a single pixel, a local block to the entire region; At each analysis scale, the distribution of gradient directions of all pixels within the statistical region is calculated, and the Shannon entropy value of the distribution is used as the structural disorder entropy value at the corresponding analysis scale.
6. The method for detecting the quality of highway pavement according to claim 5, characterized in that, The quantification of the structural instability index includes plotting the change curve of the structural disorder entropy value of the candidate distortion region as the analysis scale increases, identifying the inflection point in the change curve where the rate of change of entropy value changes from rapid to slow down, and linearly combining the scale value corresponding to the inflection point, the average slope ratio of the curve before and after the inflection point, and the entropy value at the inflection point to calculate the structural instability index of the corresponding region. Each candidate distortion region is used as a base layer, and its corresponding structural instability index is used as an intensity layer superimposed on this region. Weighted fusion is then performed to generate the road surface distortion response map.
7. The method for detecting the quality of highway pavement according to claim 1, characterized in that, S3 includes extracting the boundary contour of each candidate distortion region from the road surface distortion response map; calculating the gradient direction consistency of edge pixels within each boundary contour; and calculating the ratio of the area of each contour to the area of its smallest circumscribed convex polygon as its morphological closure index.
8. A method for detecting the quality of highway pavement according to claim 7, characterized in that, S3 further includes establishing a defect identification model, wherein the input features of the defect identification model are the gradient direction consistency, morphological closure index and structural instability index of the corresponding region; and a pre-trained support vector machine is used as the defect identification model to output the road surface defect area.
9. A method for detecting the quality of highway pavement according to claim 8, characterized in that, All identified pavement defect areas are sorted and graded according to the magnitude of their structural instability index, and divided into multiple defect levels. The total defect area of each defect level and its spatial distribution density within the test section are calculated. Based on the highest level of all defect areas within the test section, the proportion of the total defect area, and the spatial distribution density, the pavement quality level is calculated and output as four levels: excellent, qualified, poor, and dangerous, according to a predefined comprehensive scoring rule.
10. A highway pavement quality inspection system, used to implement the highway pavement quality inspection method as described in any one of claims 1-9, characterized in that, Includes the following modules: Image acquisition module: used to acquire time series images of the road surface under test; Image preprocessing module: used to construct a minimum interference frame group from the image time series based on the principle of minimizing inter-frame brightness gradient changes; The region reconstruction and structural entropy analysis module is used to perform pixel clustering at the same location on the minimum interference frame group, generate candidate distortion regions, and perform multi-scale structural entropy analysis to extract the structural instability index and construct the road distortion response map. Defect identification and quality level assessment module: used to identify defect areas based on the gradient direction consistency, morphological closure degree and structural instability index of each candidate area in the distortion response map, and output the pavement quality level result.