Road damage detection method based on vision assistance

Abstract

The invention relates to the technical field of computer vision, and discloses a road damage detection method based on vision assistance. The method comprises the following steps: firstly, collecting a pavement grayscale image, and establishing a pixel coordinate and a row-column index with the left upper corner as an original point; gaussian smoothing is carried out on the image, and first-order partial derivative is solved in row and column directions to obtain a gradient field; carrying out weighted accumulation on the gradient in a local window to form a symmetric structure tensor, and taking a characteristic value difference as structure energy; second-order and mixed partial derivatives are solved through second-order difference, and curvature tensor is constructed to obtain curvature energy; standardizing the two types of energy, and weighting the two types of energy into coupling energy according to a self-adaptive proportion; setting a threshold value according to the total graph statistics to generate candidate masks; and outputting structured results such as boundaries, geometric centers and the like through four neighborhood connected domain and area and shape ratio screening. The method gives consideration to cracks and pits, improves the accuracy and adaptability, reduces the manual dependence, and is convenient for large-scale batch application.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a vision-assisted road damage detection method. Background Technology

[0002] With the continuous expansion of transportation infrastructure, the health of road surface structures is becoming increasingly important for socio-economic operation and traffic safety. Urban roads, highways, and roads within industrial parks are prone to various types of surface damage, such as cracks, potholes, and erosion, due to multiple factors including traffic loads, environmental aging, and rainwater erosion. If these damages are not detected and repaired in a timely manner, they will further accelerate road surface deterioration, increase accident risks, and even affect the overall traffic efficiency of the road network. Therefore, developing efficient, accurate, and automated road damage detection technologies has become an important research direction in the fields of intelligent transportation and infrastructure maintenance.

[0003] Currently, traditional road damage detection methods primarily rely on manual inspection and recording. This approach depends on human experience and subjective judgment, resulting in high labor intensity, low efficiency, and difficulty in achieving large-area, all-weather, and standardized damage detection. It is also prone to problems such as missed detections, false detections, and data inconsistencies. To improve detection efficiency and automation, auxiliary detection methods based on mobile devices, sensor arrays, or vehicle-mounted systems have gradually emerged in recent years. For example, physical measurement methods such as vehicle accelerometers, laser rangefinders, and ground radar can be used to indirectly monitor road smoothness or abrupt changes. However, these methods often only reflect local physical characteristics and have limited ability to distinguish microscopic or surface damage such as cracks and shallow potholes. They are also easily affected by various factors such as ground material, vehicle parameters, and installation errors, resulting in poor detection accuracy and scene adaptability. With the advancement of digital imaging and data processing technologies, computer vision-based road damage detection methods have become a research hotspot in academia and industry. Typical technical paths include: threshold segmentation, edge detection, and morphological analysis based on image processing, as well as end-to-end detection models based on frameworks such as convolutional neural networks (CNN) and deep learning in recent years. Traditional image processing methods, such as Sobel / Laplacian edge operators, grayscale histogram segmentation, and region growing, rely on manually set thresholds or simple local features, making them susceptible to changes in lighting, shadow interference, and texture complexity. Their robustness and adaptability are insufficient, making them ill-suited to the complex and ever-changing real-world road environments. Deep learning methods, on the other hand, extract complex features through large-scale sample training, achieving high levels of automated detection capabilities. However, these methods heavily depend on high-quality, fully labeled training data. Their model structure and inference processes are often black boxes, resulting in limited generalization and interpretability. Furthermore, the inference process demands significant computational resources, hindering direct application in lightweight scenarios such as embedded systems and mobile devices. In addition, existing visual detection methods often focus on a single feature dimension, such as extracting only grayscale changes or edge responses, lacking a unified multi-scale, structural descriptive framework for complex road damage morphologies (such as the coexistence of linear cracks and planar potholes). Some studies have attempted to use multi-channel feature fusion or statistical model-based anomaly detection, but these still face significant challenges in practical applications, including strong noise interference, empirical parameter setting, and unstable target segmentation and morphology selection results.

[0004] Therefore, this case aims to propose a vision-assisted road damage detection method. It utilizes grayscale image acquisition and pixel coordinate system establishment, combines Gaussian smoothing noise reduction and multi-order difference operators to extract local structure and curvature information of the image, integrates structure and curvature energies through tensor feature analysis, uses adaptive weights to form a unified significance criterion, and finally accurately identifies and outputs the location and contour of the damaged area through statistical threshold segmentation, connected component analysis and morphological screening. Summary of the Invention

[0005] This invention provides a vision-assisted road damage detection method, which helps to solve the problems mentioned in the background art.

[0006] This invention provides the following technical solution: a road damage detection method based on vision assistance, comprising:

[0007] Acquire a grayscale image of the road surface, determine the discrete pixel indices in the row and column directions, and establish a Cartesian pixel coordinate system with the top left corner as the origin;

[0008] Discrete Gaussian smoothing is applied to the grayscale image, and the gradient field is obtained by calculating the first-order partial derivatives in the row and column directions respectively.

[0009] The gradient field is weighted and accumulated within a preset window to construct a two-dimensional symmetric structure tensor, and the tensor eigenvalues ​​are calculated to form a structural energy representation.

[0010] Second-order partial derivatives and mixed partial derivatives are obtained based on discrete second-order differences, and a symmetric curvature tensor is constructed to obtain a curvature energy characterization.

[0011] The structural energy and curvature energy are standardized to determine the adaptive scaling factor, and the coupling energy is obtained by combining them according to weights.

[0012] Calculate statistics on the coupled energy map, set a significant anomaly threshold, and generate a damage candidate mask;

[0013] Perform four-neighbor connectivity analysis on the mask, and filter out the effective damage area based on the lower limit of area and the lower limit of shape ratio.

[0014] The pixel boundaries and geometric centers of the effective damage area are extracted and the detection results are output in a structured form.

[0015] Optionally, the step of acquiring a grayscale image of the road surface, determining the discrete pixel indices in the row and column directions, and establishing a Cartesian pixel coordinate system with the top left corner as the origin specifically includes:

[0016] A road grayscale image is captured using a vision acquisition system. Let the image be a two-dimensional discrete matrix I. m,n The size is M×N; where m∈[0,M-1] is the row index of the pixel, from top to bottom; n∈[0,N-1] is the column index of the pixel, from left to right; I m,n The value is the grayscale value located at pixel row index m and column index n; M≥7, which is the number of rows in the image; N≥7, which is the number of columns in the image;

[0017] Establish the origin as the top left corner O 0,0 The Cartesian pixel coordinate system is used, with the n-axis along the width direction and the m-axis along the height direction. Each pixel corresponds to a unique coordinate (m, n).

[0018] Optionally, the step of performing discrete Gaussian smoothing on the grayscale image and calculating the first-order partial derivatives in the row and column directions to obtain the gradient field specifically includes:

[0019] S210, Discrete Gaussian kernel normalization:

[0020] For image I m,n A two-dimensional discrete Gaussian kernel with a standard deviation of σ = 3 pixels and supporting {u,v∈[-3,3]} is applied and then discretely normalized, specifically as follows:

[0021]

[0022] Where u and v are the relative row and column displacements of the discrete kernel; Z represents the value of the unnormalized two-dimensional discrete Gaussian kernel at the relative displacement (u,v); σ This is the summation normalization factor of the discrete Gaussian kernel over the supporting domain; is the normalized Gaussian kernel; u, v are the relative row and column shifts of the discrete kernel;

[0023] Smoothing results: in, The grayscale value is the smoothed value.

[0024] S220, Derivative Superscript Correspondence and Pixel Spacing Explicitization:

[0025] Based on the five-point central difference formula, calculate the first-order partial derivatives of the smoothed image in the m and n directions:

[0026]

[0027] in, For I ( s) Discrete first derivative estimation with respect to column axis n; For I ( s) Discrete first derivative estimate with respect to the row axis m; I ( s) is the image after Gaussian smoothing;

[0028] S230, Symmetric mapping processing of boundary indices:

[0029] For all indexes involving out-of-bounds access in S210 and S220, a periodic symmetric reflection rule is uniformly applied: For any row index placeholder r and column index placeholder c, the following is defined:

[0030] P M =2M, r ★ =r mod P M r ★ ∈{0,1,...,P M -1},

[0031]

[0032] P N =2N, c ★ =c mod P N c ★ ∈{0,1,...,P N -1},

[0033]

[0034] Where r and c are placeholders, representing any potentially out-of-bounds row and column integer indices; P M P N The period length is the periodized length of the row and column latitudes, used for symmetrical reflection; r ★ c ★ For mapping indices, arbitrary integer indices are transformed back to the periodic intervals {0,1,...,2M-1} and {0,1,...,2N-1} through modulo operation; φ(r;M) and ψ(c;N) are mapping functions that map out-of-bounds indices back to the valid intervals {0,1,...,M-1} and {0,1,...,N-1} through periodicized symmetric reflection, with segmented definition to ensure endpoint symmetry.

[0035] Optionally, the step of weighted accumulation of the gradient field within a preset window to construct a two-dimensional symmetric structure tensor and calculating tensor eigenvalues ​​to form a structure energy representation specifically includes:

[0036] For each pixel (m,n), define the structure tensor J. m,n It is a two-dimensional symmetric matrix:

[0037] Among them, J m,n Let (m, n) be the structure tensor of the pixel (m, n). For the x components of the structure tensor; The yy component of the structure tensor; These are the xy components of the structure tensor;

[0038] Apply Gaussian kernel smoothing with a standard deviation ρ = 1.5 pixels to a window size of 5×5 pixels, and then discretize and normalize:

[0039]

[0040] Where ρ is the one-dimensional standard deviation used for smoothing within the structure tensor window; Z ρ , The kernel and normalized quantity are the same as those in step S210, but the standard deviation is ρ and the support region is [-2,2].

[0041] Calculate the two eigenvalues ​​of the tensor

[0042]

[0043] Constructing the local structure energy function

[0044] Optionally, the step of obtaining second-order partial derivatives and mixed partial derivatives based on discrete second-order differences, constructing a symmetric curvature tensor, and obtaining a curvature energy characterization specifically includes:

[0045] Calculate the second-order partial derivative using the five-point difference formula:

[0046]

[0047] in, The second-order discrete derivative is estimated along the column axis n; The second-order discrete derivative is estimated along the row axis m; The mixed second-order discrete derivative with respect to m and then with respect to n;

[0048] Constructing a symmetric curvature tensor: Among them, C m,n Let m be a matrix variable, and n be the curvature tensor of the pixel (m,n).

[0049] Constructing the curvature energy function

[0050] Optionally, the standardization of structural energy and curvature energy to determine an adaptive scaling factor, and the weighted combination to obtain coupling energy, specifically includes:

[0051] S510, Construction of Standardized Quantities for the Entire Map:

[0052]

[0053] Where, μ J μ C The overall mean is the arithmetic mean of the structural energy and curvature energy across the entire graph, respectively; σ J σ C The standard deviation of the whole map is a measure of the dispersion of energy across the whole map.

[0054] S520, Constructing dimensionless energy and adaptive scaling factor:

[0055]

[0056] If the denominator is 0, then let in, Standardized energy; This is an adaptive scaling factor;

[0057] S530, Constructing Coupled Energy: Among them, S m,n Let be the coupling energy of pixel (m,n).

[0058] Optionally, the step of calculating statistics on the coupled energy map, setting a significant anomaly threshold, and generating a damage candidate mask specifically includes:

[0059] Calculate the average coupling energy μ of the entire image. S With standard deviation σ S :

[0060]

[0061] Set the significant anomaly threshold to T. S =μ S +2σ S ;

[0062] Generate damage mask map Among them, D m,n =1 indicates a suspected damage point, D m,n =0 indicates a non-damaged point.

[0063] Optionally, the step of performing four-neighbor connectivity analysis on the mask and filtering effective damage regions based on area and shape ratio lower limits specifically includes:

[0064] For mask image D m,n Perform 4-neighborhood connectivity analysis and define the region:

[0065] And connected to a point within the region via several 4-neighborhoods}; where, Let k be the set of pixels connected to the kth 4-neighborhood; k is the region number.

[0066] Define the set of boundary edges of the region:

[0067] Or ((m′,n′) out of bounds)}; where, To and The set of adjacent pixel edges (including boundary edges); (m′,n′) is a pixel coordinate pair, the pixel adjacent to (m,n)4;

[0068] Calculate the perimeter L of the k-th region respectively. k and area A k :

[0069]

[0070] Let the minimum analytical scale for first-order and second-order difference and tensor smoothing be p = 5 pixels. Based on this, define: A min =p 2 =25, Among them, A min η is the lower limit of the area used to eliminate excessively small noise regions. min The lower limit of the shape ratio is used to preserve elongated crack-like regions;

[0071] Morphological screening criteria and branching operations: If A k ≥A min and Then retain The area is considered an effective damage area; otherwise, delete it. Not included in subsequent output;

[0072] After this filtering, the remaining regions are reassigned with the numbers k = {1, ..., K} according to their order of appearance; where K is the total number of valid connected regions after filtering.

[0073] Optionally, the step of extracting pixel boundaries and geometric centers of the effective damaged area and outputting the detection results in a structured form specifically includes:

[0074] Construct a set of pixel boundaries:

[0075] Or ((m′,n′) out of bounds)}; where, Let k be the set of pixel boundary points of the k-th region;

[0076] Obtain the geometric center point: in, Let be the row and column geometric center coordinates of the k-th region;

[0077] Each valid area The output is structured data, with the following fields recorded:

[0078] Region number k, pixel boundary Maximum coupling energy value Geometric center point

[0079] Construct all valid regions to form the final detection result set.

[0080] If K=0 after filtering, then output The conclusion is that no road damage areas satisfying the threshold and morphological constraints were detected in this image.

[0081] The present invention has the following beneficial effects:

[0082] 1. A standardized Cartesian pixel coordinate system with the top-left corner as the origin is introduced into road damage detection, clarifying the spatial meaning of image row and column indices and achieving unique localization of each pixel. By rigidly defining a minimum image size, all subsequent convolution, differencing, and other operations are ensured to be definable even in boundary conditions, completely avoiding artifacts or numerical discontinuities caused by traditional algorithms when processing boundary pixels. Simultaneously, a periodic symmetric reflection mapping is used for out-of-bounds indices, improving computational stability and enhancing the algorithm's adaptability to different resolutions and shooting conditions. These improvements provide a unified and comparable data foundation for subsequent advanced operations such as gradients and tensors, effectively guaranteeing the algorithm's generalization ability across devices and scenarios. Compared to existing detection methods that rely solely on raw pixels or simple indices, this solution has a more rigorous data organization structure, facilitating seamless integration with external systems such as GIS and databases, and enabling practical engineering implementation.

[0083] 2. Image preprocessing is performed using a discrete Gaussian kernel, and the first derivatives in the row and column directions are calculated using a five-point center difference method, effectively improving noise suppression and gradient estimation accuracy. Gaussian smoothing suppresses high-frequency textures and imaging noise, providing stable input for subsequent structural analysis. Compared to the three-point method, the five-point center difference method has higher accuracy in quantization error and edge response, ensuring that boundary information of structures such as cracks and pits can be accurately captured. The innovation lies in the fact that smoothing and gradient extraction form a robust local structural input channel, taking into account both sensitivity to minor damage and tolerance to background clutter.

[0084] 3. Within a preset window, the local gradient field is accumulated using a Gaussian kernel weighting to form a two-dimensional symmetric structural tensor. Eigenvalue decomposition analysis is then used to analyze the intensity of the principal and secondary directions, enabling the quantification of the directionality and prominence of linear damage structures such as cracks. The structural tensor can distinguish between strongly anisotropic linear structures and isotropic uniform regions. The eigenvalue difference, as structural energy, can improve the response capability to anomalies such as micro-cracks and slender breaks. Through Gaussian weighting and eigenvalue analysis, not only is the detection sensitivity for linear damage enhanced, but also non-directional random texture noise is suppressed. Compared with traditional algorithms based on unidirectional gradients or simple edge operators, structural tensor analysis has higher rotational invariance and robustness, and can handle complex road textures, irregular damage, and other practical conditions.

[0085] 4. By calculating second-order partial derivatives in the row, column, and mixed directions using the five-point difference formula, a locally symmetric curvature tensor is constructed, and its absolute determinant is defined as curvature energy. This method can efficiently detect planar and nonlinear damage structures such as pits and spalling, supplementing the advantages of the structure tensor for linear features and achieving compatible coverage of various damage morphologies. The introduction of the curvature tensor overcomes the problem of insufficient response capability of traditional methods to nonlinear and curved surface damage, and is especially suitable for irregular damage, spots, pits, and other situations that are difficult to describe with single-directional features. Compared with schemes that rely solely on first-order derivatives, curvature analysis improves the detection capability of complex surface changes and reduces false alarms caused by texture undulations.

[0086] 5. A method is proposed to standardize structural energy and curvature energy across the entire image and adaptively calculate the fusion weight for each pixel accordingly, synthesizing the two saliency indicators into a unified coupled energy index. This strategy can dynamically allocate discrimination weights based on the actual type of pavement damage (e.g., crack-dominated or pothole-dominated), avoiding the failure of single channels and improving the overall adaptive capability of detection. The unsupervised determination mechanism of the adaptive scaling factor avoids the drawbacks of traditional methods that require manual parameter tuning or external annotation training.

[0087] 6. By statistically modeling the entire coupled energy map, an unsupervised threshold of the mean plus twice the standard deviation is used to automatically segment candidate damage regions, eliminating the need for manual annotation or preset empirical parameters. This approach incorporates the concept of statistical anomaly detection, automatically adapting to energy distributions in different images and environments, reducing the subjectivity of parameter selection. Automatic segmentation of suspected damage regions using full-image statistical indicators effectively suppresses false detections caused by moderate-intensity noise while ensuring high recall for highly salient targets. Compared to traditional mask generation methods based on threshold settings or simple binarization, this method has stronger adaptability and universality, reducing manual intervention and improving the level of detection automation.

[0088] 7. Based on the candidate mask, four-neighborhood connectivity analysis is introduced, and geometric and morphological screening is performed by combining area and shape ratio lower limits. This effectively eliminates isolated noise, small spots, and non-target areas that do not conform to engineering definitions. The setting of area and shape ratio lower limits combines the minimum analytical scale of first-order and second-order difference and tensor smoothing, achieving flexible retention of different types of damage. This method maintains a high recognition rate for typical pavement damage such as thin cracks and small pothole edges, while minimizing the false alarm rate. A strict binary rule is used to determine the retention or rejection of regions, ensuring the traceability and consistency of the final output. Compared with previous algorithms that rely on manual post-processing or subjective experience screening, this method has a high degree of standardization and is easy to deploy and promote in batches.

[0089] 8. For each valid connected region, structured fields are output, including pixel boundaries, maximum coupling energy value, and geometric center point, achieving full output of spatial, intensity, and location information. The final results are organized in a data structure format, facilitating visualization, statistical report generation, and seamless integration with road maintenance systems or GIS platforms. When there is no valid target, clear empty results and conclusions are output, completely avoiding data gaps and processing anomalies caused by misunderstandings. This scheme improves the engineering usability and system integration of the detection results, providing a standardized and operable data foundation for subsequent intelligent assessment, decision-making, and management. Unlike existing detection results that only provide image binary masks or require manual secondary processing, this method achieves automated, standardized, and structured detection data output, reducing the risk of human intervention and misoperation. Attached Figure Description

[0090] Figure 1 This is a schematic diagram of the process of the present invention.

[0091] Figure 2 This is a schematic diagram of the Cartesian rectangular pixel coordinate system of the present invention.

[0092] In the diagram: 1-Image, 2-Origin, 3-Image height, 4-Image width, 5-m-axis, 6-n-axis. Detailed Implementation

[0093] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0094] Example, refer to Figure 1 A vision-assisted road damage detection method includes:

[0095] Acquire a grayscale image of the road surface, determine the discrete pixel indices in the row and column directions, and establish a Cartesian pixel coordinate system with the top left corner as the origin;

[0096] Discrete Gaussian smoothing is applied to the grayscale image, and the gradient field is obtained by calculating the first-order partial derivatives in the row and column directions respectively.

[0097] The gradient field is weighted and accumulated within a preset window to construct a two-dimensional symmetric structure tensor, and the tensor eigenvalues ​​are calculated to form a structural energy representation.

[0098] Second-order partial derivatives and mixed partial derivatives are obtained based on discrete second-order differences, and a symmetric curvature tensor is constructed to obtain a curvature energy characterization.

[0099] The structural energy and curvature energy are standardized to determine the adaptive scaling factor, and the coupling energy is obtained by combining them according to weights.

[0100] Calculate statistics on the coupled energy map, set a significant anomaly threshold, and generate a damage candidate mask;

[0101] Perform four-neighbor connectivity analysis on the mask, and filter out the effective damage area based on the lower limit of area and the lower limit of shape ratio.

[0102] The pixel boundaries and geometric centers of the effective damage area are extracted and the detection results are output in a structured form.

[0103] A systematic, end-to-end vision-assisted road damage detection method is proposed. By clearly defining the steps: First, acquire grayscale images of the road surface and establish discrete pixel indices and a standard coordinate system in the row and column directions. This step ensures accurate image data acquisition and spatial positioning, providing a standardized spatial foundation for all subsequent image processing operations. Next, noise is reduced through discrete Gaussian smoothing, and first-order partial derivatives are calculated in two orthogonal directions to extract gradient information, effectively enhancing the perception of weak structures. Statistical analysis of the gradient field is performed through weighted accumulation of a preset window to construct a structure tensor, and then principal and secondary eigenvalues ​​are extracted to form structural energy. This mechanism can sensitively identify features such as linear cracks. Furthermore, by combining second-order difference to construct a curvature tensor and extract curvature energy, nonlinear and planar damage features such as potholes and spalling can be effectively captured. The structural energy and curvature energy are standardized and adaptively weighted and fused to obtain a coupled energy index sensitive to different damage types. By setting anomaly thresholds through statistical modeling, this method automatically segments damaged areas. Finally, through connected component analysis and morphological constraint filtering, effective damaged areas are obtained, and structured detection results are output. Compared to existing methods, this approach is compatible with both linear and nonlinear damage types, improving detection accuracy and adaptability, reducing the need for manual parameter adjustments, and enhancing automation and batch processing capabilities. This solution covers the entire process from data acquisition and processing to output, transforming road maintenance work from traditional reliance on experience and manual inspection to intelligent detection based on scientific algorithms and standardized data. This improves work efficiency and detection reliability, making it particularly suitable for large-area, multi-type, and multi-scenario practical applications.

[0104] Reference Figure 2 The process of acquiring a grayscale image of the road surface, determining the discrete pixel indices in the row and column directions, and establishing a Cartesian pixel coordinate system with the top left corner as the origin specifically includes:

[0105] A road grayscale image is captured using a vision acquisition system. Let the image be a two-dimensional discrete matrix I. m,nThe size is M×N; where m∈[0,M-1] is the row index of the pixel, from top to bottom; n∈[0,N-1] is the column index of the pixel, from left to right; I m,n Let m be the gray value located at pixel row index m and column index n; M≥7, which is the number of rows in the image; N≥7, which is the number of columns in the image; establish a unified data object and numerical range to provide a clear input domain and unit guarantee for subsequent convolution, difference, tensor and statistical operations, avoid dimensional confusion and facilitate the comparison of results between different images; give a hard constraint on the minimum image size to ensure that all boundary accesses can be defined under symmetric reflection mapping, and avoid numerical holes and artifacts caused by out-of-bounds access;

[0106] Establish the origin as the top left corner O 0,0 The Cartesian pixel coordinate system is used, with the n-axis along the width direction and the m-axis along the height direction. Each pixel corresponds to a unique coordinate (m, n).

[0107] The specific implementation methods for image acquisition and spatial indexing are further refined. By explicitly requiring the use of a visual acquisition system to capture grayscale images of roads and setting a lower limit on image size, it ensures that all pixels can be effectively processed, preventing boundary processing failures or anomalies caused by excessively small images. Simultaneously, a unified coordinate system and data objects are constructed, providing a clear input / output domain and numerical foundation for subsequent operations such as convolution, differencing, tensor analysis, and statistics. In particular, the hard constraint on minimum size and the boundary processing scheme for symmetric reflection mapping not only guarantee the algorithm's stability but also prevent numerical holes, artifacts, and false detections caused by boundary index out-of-bounds errors. Compared to most existing technologies that do not standardize the management of spatial indexing systems and boundary conditions, this method improves the algorithm's rigor, stability, and cross-platform compatibility. It also facilitates the subsequent engineering applications and system integration of the detection results, making it suitable for deployment across different acquisition devices and application environments.

[0108] The process of performing discrete Gaussian smoothing on the grayscale image and calculating the first-order partial derivatives in both the row and column directions to obtain the gradient field specifically includes:

[0109] S210, Discrete Gaussian kernel normalization:

[0110] For image I m,n A two-dimensional discrete Gaussian kernel with a standard deviation of σ = 3 pixels and supporting {u,v∈[-3,3]} is applied and then discretely normalized, specifically as follows:

[0111]

[0112] Where u and v are the relative row and column displacements of the discrete kernel; Z represents the value of the unnormalized two-dimensional discrete Gaussian kernel at the relative displacement (u,v); σThis is the summation normalization factor of the discrete Gaussian kernel over the supporting domain; is the normalized Gaussian kernel; u, v are the relative row and column shifts of the discrete kernel;

[0113] Smoothing results: in, The grayscale value is the smoothed value.

[0114] While maintaining the overall brightness energy, noise and high-frequency textures are suppressed, making the discrete difference insensitive to noise and providing a more stable input for the structure tensor and second derivative.

[0115] S220, Derivative Superscript Correspondence and Pixel Spacing Explicitization:

[0116] Based on the five-point central difference formula, calculate the first-order partial derivatives of the smoothed image in the m and n directions:

[0117]

[0118] in, For I ( s) Discrete first derivative estimation with respect to column axis n; For I ( s) Discrete first derivative estimate with respect to the row axis m; I ( s) is the image after Gaussian smoothing;

[0119] Extracting directional change information provides local gradient input for the structure tensor; five-point center difference improves approximation accuracy and suppresses quantization noise;

[0120] S230, Symmetric mapping processing of boundary indices:

[0121] For all indexes involving out-of-bounds access in S210 and S220, a periodic symmetric reflection rule is uniformly applied: For any row index placeholder r and column index placeholder c, the following is defined:

[0122] P M =2M, r ★ =r mod P M r ★ ∈{0,1,...,P M -1},

[0123]

[0124] P N =2N, c ★ =c mod P N c ★ ∈{0,1,...,P N -1},

[0125]

[0126] Where r and c are placeholders, representing any potentially out-of-bounds row and column integer indices; P M P N The period length is the periodized length of the row and column latitudes, used for symmetrical reflection; r ★ c ★ For mapping indices, arbitrary integer indices are transformed back to the periodic intervals {0,1,...,2M-1} and {0,1,...,2N-1} through modulo operation; φ(r;M) and ψ(c;N) are mapping functions that map out-of-bounds indices back to the valid intervals {0,1,...,M-1} and {0,1,...,N-1} through periodicized symmetric reflection, with segmented definition to ensure endpoint symmetry;

[0127] To ensure that the smoothness and difference have consistent statistical properties at the boundaries and inside, avoid artificial increases or decreases in energy, and prevent artificial boundary features from interfering with subsequent detection.

[0128] Image preprocessing and gradient extraction are refined into three steps: Gaussian smoothing for noise reduction, five-point central difference differentiation, and boundary symmetric reflection mapping. Gaussian smoothing effectively suppresses high-frequency noise and texture interference in the image, ensuring that subsequent structural and difference analyses are not excessively affected by noise, thus improving the stability and accuracy of damage features. The five-point central difference formula is more accurate than the traditional three-point difference formula, better preserving image structural information and being particularly sensitive to details with gradual changes, such as cracks and edges. Simultaneously, a periodic symmetric reflection mapping rule is uniformly applied to all pixel indices involving boundary crossings, ensuring equal treatment for boundary pixels and internal pixels, eliminating numerical discontinuities such as energy anomalies and fictitious increases / decreases at boundaries that occur in traditional algorithms. The synergistic effect of these steps makes the entire image, from initial acquisition and preprocessing to gradient analysis, highly robust and consistent, reducing false positives and false negatives.

[0129] The step of weighted accumulation of the gradient field within a preset window to construct a two-dimensional symmetric structure tensor and calculating tensor eigenvalues ​​to form a structure energy representation specifically includes:

[0130] For each pixel (m,n), define the structure tensor J. m,n It is a two-dimensional symmetric matrix:

[0131] Among them, J m,n Let (m, n) be the structure tensor of the pixel (m, n). For the x components of the structure tensor; The yy component of the structure tensor; The xy components of the structural tensor are encoded in matrix form to provide an analytical carrier for anisotropy and principal direction, laying the foundation for the identification of linear structures such as cracks.

[0132] Apply Gaussian kernel smoothing with a standard deviation ρ = 1.5 pixels to a window size of 5×5 pixels, and then discretize and normalize:

[0133]

[0134] Where ρ is the one-dimensional standard deviation used for smoothing within the structure tensor window; Z ρ , The kernel and normalized quantity are the same as those in step S210, but the standard deviation is ρ and the support region is [-2,2].

[0135] Gradient information is aggregated with Gaussian weights within a 5×5 local area to suppress noise and enhance local consistency, making tensor features robust to small perturbations and rotation-adaptive.

[0136] Calculate the two eigenvalues ​​of the tensor

[0137]

[0138] Decoupling the intensity of the local structure in the primary and secondary directions makes it easier to characterize the prominence of the linear structure by the difference in eigenvalues, and reduces the sensitivity to specific directional angles.

[0139] Constructing the local structure energy function The anisotropic intensity is measured by the eigenvalue gap, which achieves a higher response at linear cracks, while suppressing the response in regions with uniform or isotropic textures.

[0140] By weighted accumulation of the gradient field within a window, a two-dimensional symmetric structure tensor is constructed, and the structural energy of the principal and secondary directions is extracted through eigenvalue decomposition. This process uses Gaussian kernel weighting to ensure the smoothness and robustness of local statistics, enhancing the ability to identify linear and anisotropic structures (such as cracks). The structure tensor can distinguish the intensity of local principal and secondary directions, and the difference in eigenvalues ​​is used to measure the structural energy. This not only improves the response to crack-dominated regions but also automatically suppresses the response in regions without principal directions or with isotropic properties, effectively reducing false alarms in non-damage areas such as textures and spots. Furthermore, this method exhibits strong rotation invariance, adapting to damage characteristics in different directions. Compared to traditional methods that rely solely on gradient modulus and edge detection, the structure tensor method can more accurately distinguish subtle and hidden linear damage, achieving high-sensitivity detection of cracks in real-world complex pavement environments and solving the problem of poor adaptability of existing methods to atypical damage.

[0141] The process of obtaining second-order partial derivatives and mixed partial derivatives based on discrete second-order differences, constructing a symmetric curvature tensor, and obtaining a curvature energy representation specifically includes:

[0142] Calculate the second-order partial derivative using the five-point difference formula:

[0143]

[0144] in, The second-order discrete derivative is estimated along the column axis n; The second-order discrete derivative is estimated along the row axis m; It is a mixed second-order discrete derivative with respect to m and then with respect to n; it captures brightness curvature and bending information, and the second-order terms are significant at pits (concave / convex) and sharp turns, which helps to highlight nonlinear structures such as pits and broken edges in subsequent curvature energy;

[0145] Constructing a symmetric curvature tensor: Among them, C m,n For matrix variables, the curvature tensor of pixel (m,n); integrating the second derivative into a symmetric tensor provides a unified framework for constructing curvature intensity using invariants such as determinants, facilitating rotation-independent analysis;

[0146] Constructing the curvature energy function The absolute value of the determinant of the curvature tensor is used to characterize the local "Gaussian curvature type" intensity. The response is enhanced at areas with planar anomalies such as pits and spalling, which complements the structural energy advantage over cracks and achieves a complementary effect.

[0147] This paper proposes a method to obtain second-order and mixed partial derivatives based on a five-point difference formula, and then construct a symmetric curvature tensor and curvature energy. The second-order derivative primarily reflects the curvature features of an image, effectively detecting planar and nonlinear damage such as potholes and spalling, which are often key indicators of road surface deterioration. First-order structural features, however, are often insensitive to these issues. By integrating the second-order derivatives into a symmetric tensor and using the absolute value of the determinant to measure the curvature energy, direction-independent detection of complex morphological changes can be achieved, effectively compensating for the limitation of first-order structural energy, which is only sensitive to linear features. Furthermore, this method is rotation-independent, obtaining consistent responses in anomalous regions such as potholes and damaged edges from various orientations. Compared to traditional methods relying on manually designed filters or simple edge detection, the curvature tensor method improves the detection capability for nonlinear and non-uniform damage, further enhancing the overall detection breadth and accuracy of the solution.

[0148] The standardization of structural energy and curvature energy, determination of adaptive scaling factors, and weighted combination to obtain coupling energy specifically include:

[0149] S510, Construction of Standardized Quantities for the Entire Map:

[0150]

[0151] Where, μ J μ C The overall mean is the arithmetic mean of the structural energy and curvature energy across the entire graph, respectively; σ J σ C The standard deviation of the whole image corresponds to the degree of dispersion of energy across the whole image; it obtains energy statistics that are comparable across images, which are used for subsequent standardization to offset the effects of image brightness, material albedo and illumination differences on the energy scale.

[0152] S520, Constructing dimensionless energy and adaptive scaling factor:

[0153]

[0154] If the denominator is 0, then let in, Standardized energy; It is an adaptive scaling factor; it normalizes the two types of energy to the same dimensionless scale and adaptively allocates weights according to the local relative salience of pixels, so as not to over-amplify either channel, and to strengthen the corresponding channel at the dominant mechanism.

[0155] S530, Constructing Coupled Energy: Among them, S m,n The coupling energy of pixel (m,n) is used; pixel-level fusion of structural and curvature evidence forms a unified significance measure that takes into account both cracks (linear anisotropy) and pits (curvature anomalies).

[0156] This paper proposes a full-image standardization method for structural and curvature energies, followed by weight allocation using an adaptive scaling factor to ultimately form a unified coupled energy. Firstly, standardization eliminates energy scale differences caused by external factors such as image brightness and material reflection, improving the comparability and universality of the results. The introduction of the adaptive scaling factor allows the algorithm to automatically adjust the contribution of structural and curvature energies based on the actual damage type, ensuring dual sensitivity to both linear and planar damage while avoiding misjudgments caused by a single dominant energy. This coupled energy serves as a saliency criterion, providing a more comprehensive and accurate basis for subsequent threshold segmentation and target extraction. Compared to traditional methods relying on a single channel or fixed weighting strategies, this fusion approach enhances the algorithm's adaptability to complex and variable real-world scenarios and reduces the subjectivity of parameter selection and the difficulty of manual adjustment.

[0157] The calculation of statistics on the coupled energy map, setting a significant anomaly threshold, and generating a damage candidate mask specifically includes:

[0158] Calculate the average coupling energy μ of the entire image. S With standard deviation σ S :

[0159] To characterize the overall distribution level and dispersion of coupling energy, providing a statistical basis for a unified, unsupervised threshold;

[0160] Set the significant anomaly threshold to T. S =μ S +2σ S Using two standard deviations as the deviation limit, high-energy anomaly regions can be separated from the background without external annotation or prior knowledge, thus suppressing false alarms under moderate noise.

[0161] Generate damage mask map Among them, D m,n =1 indicates a suspected damage point, D m,n =0 indicates a non-damaged point; the continuous saliency map is discretized into a set of candidate damaged pixels, laying a discretized foundation for subsequent geometric and morphological analysis (connected regions, perimeter, area).

[0162] By performing full-image statistical analysis on the coupled energy map, anomaly thresholds are automatically set, and damage candidate masks are generated. The automatic calculation of the full-image statistical mean and standard deviation enables adaptive threshold setting, dynamically adjusting according to the actual distribution of different images, effectively avoiding false positives and false negatives caused by static thresholds. Through unsupervised anomaly detection, regions exceeding the threshold are automatically separated, reducing the need for manual intervention and parameter tuning. The generated mask map provides high-quality candidate regions for subsequent connected component analysis and morphological screening, ensuring the completeness and effectiveness of the final detected targets. Unlike traditional methods that rely on manually setting thresholds or empirical parameters, this method improves the automation level of detection and its adaptability to practical applications, providing a technical foundation for batch processing and unattended applications.

[0163] The step of performing four-neighbor connectivity analysis on the mask, and filtering out effective damage regions based on area and shape ratio lower limits, specifically includes:

[0164] For mask image D m,n Perform 4-neighborhood connectivity analysis and define the region:

[0165] And connected to a point within the region via several 4-neighborhoods}; where, Let k be the set of pixels connected to the kth 4-neighborhood; k is the region number.

[0166] Define the set of boundary edges of the region:

[0167] Or ((m′,n′) out of bounds)}; where, To and The set of adjacent pixel edges (including boundary edges); (m′,n′) is a pixel coordinate pair, the pixel adjacent to (m,n)4;

[0168] Calculate the perimeter L of the k-th region respectively. k and area A k :

[0169]

[0170] Organize pixel-level results into target-level objects and define geometric quantities, explicitly including image border boundaries, to facilitate subsequent removal of noise and non-target areas using shape constraints;

[0171] Let the minimum analytical scale for first-order and second-order difference and tensor smoothing be p = 5 pixels, and define accordingly: Among them, A min η is the lower limit of the area used to eliminate excessively small noise regions. min The lower limit of the shape ratio is used to preserve elongated crack-like regions;

[0172] Occasional small spots and blocky interference are filtered out with a geometric threshold of consistent scale, while retaining the geometric features of cracks and small pit edges that conform to the "slender / narrow" shape.

[0173] Morphological screening criteria and branching operations: If A k ≥A min and Then retain The area is considered an effective damage area; otherwise, delete it. Not included in subsequent output;

[0174] After this filtering, the remaining regions are reassigned with the numbers k = {1,...,K} according to their order of appearance; where K is the total number of valid connected regions after filtering.

[0175] Each candidate is given a decision on whether to stay or leave based on a clear binary rule, and is uniformly numbered to ensure the determinism and traceability of the output set.

[0176] By performing four-neighbor connectivity analysis on the mask image and filtering based on lower limits for area and shape ratio, the final effective damage areas are obtained. Four-neighbor connectivity analysis accurately identifies and segments each independent damage target, ensuring the spatial independence and traceability of each detection result. The dual lower limit constraints of area and shape ratio effectively filter out isolated small noises, spots, and atypical targets, focusing on retaining truly engineering-significant areas such as slender cracks and planar pits. This measure not only improves the reliability of the detection results but also facilitates subsequent statistical analysis and engineering evaluation.

[0177] The step of extracting pixel boundaries and geometric centers of the effective damaged area and outputting the detection results in a structured form specifically includes:

[0178] Construct a set of pixel boundaries:

[0179]

[0180] Or ((m′,n′) out of bounds)}; where, Let k be the set of pixel boundary points of the k-th region;

[0181] Obtain the geometric center point: in, Let be the row and column geometric center coordinates of the k-th region;

[0182] Each valid area The output is structured data, with the following fields recorded:

[0183] Region number k, pixel boundary Maximum coupling energy value Geometric center point

[0184] Output the spatial contour, intensity representative value and location information of each target in a consumable data structure, which facilitates visualization annotation, statistical reports and interface with external systems (GIS / maintenance system);

[0185] Construct all valid regions to form the final detection result set. The standardized test result set is formed as the final output, supporting batch export, overlay display and subsequent processes (measurement, evaluation, decision-making);

[0186] If K=0 after filtering, then output The conclusion is given that no road damage areas satisfying the threshold and morphological constraints were detected in this image; in the absence of a target, clear and definite empty results and textual conclusions are given to avoid the business side from misunderstanding that the processing has failed or that the data is missing.

[0187] The method specifies structured data output for effective damage areas, including extracting pixel boundary point sets, calculating geometric center points, and outputting information such as maximum energy values ​​and region numbers. Through this series of refined processes, the detection results not only possess spatial positioning accuracy but also allow for subsequent statistical analysis and visualization, facilitating seamless integration with external GIS systems and road maintenance management platforms. Simultaneously, the standardized data structure output facilitates historical data archiving, batch evaluation, and multi-scenario comparison. Furthermore, the method stipulates that when no regions meeting threshold and morphological constraints are detected, an empty result and a clear conclusion should be output, completely avoiding data anomalies and project workflow interruptions caused by misunderstandings.

[0188] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0189] 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 technical principles 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 road damage detection method based on vision assistance, characterized in that, include: Acquire a grayscale image of the road surface, determine the discrete pixel indices in the row and column directions, and establish a Cartesian pixel coordinate system with the top left corner as the origin; Discrete Gaussian smoothing is applied to the grayscale image, and the gradient field is obtained by calculating the first-order partial derivatives in the row and column directions respectively. The gradient field is weighted and accumulated within a preset window to construct a two-dimensional symmetric structure tensor, and the tensor eigenvalues ​​are calculated to form a structural energy representation. Second-order partial derivatives and mixed partial derivatives are obtained based on discrete second-order differences, and a symmetric curvature tensor is constructed to obtain a curvature energy characterization. The structural energy and curvature energy are standardized to determine the adaptive scaling factor, and the coupling energy is obtained by combining them according to the weights. Calculate statistics on the coupled energy map, set a significant anomaly threshold, and generate a damage candidate mask; Perform four-neighbor connectivity analysis on the mask, and filter out the effective damage area based on the lower limit of area and the lower limit of shape ratio. The pixel boundaries and geometric centers of the effective damage area are extracted and the detection results are output in a structured form.

2. The road damage detection method based on vision assistance according to claim 1, characterized in that, The process of acquiring a grayscale image of the road surface, determining the discrete pixel indices in the row and column directions, and establishing a Cartesian pixel coordinate system with the top left corner as the origin specifically includes: A road grayscale image is captured using a vision acquisition system. Let the image be a two-dimensional discrete matrix I. m,n The size is M×N; where m∈[0,M-1] is the row index of the pixel, from top to bottom; n∈[0,N-1] is the column index of the pixel, from left to right; I m,n The value is the grayscale value located at pixel row index m and column index n; M≥7, which is the number of rows in the image; N≥7, which is the number of columns in the image; Establish the origin as the top left corner O 0,0 The Cartesian pixel coordinate system is used, with the n-axis along the width direction and the m-axis along the height direction. Each pixel corresponds to a unique coordinate (m, n).

3. The road damage detection method based on vision assistance according to claim 2, characterized in that, The process of performing discrete Gaussian smoothing on the grayscale image and calculating the first-order partial derivatives in both the row and column directions to obtain the gradient field specifically includes: S210, Discrete Gaussian kernel normalization: For image I m,n A two-dimensional discrete Gaussian kernel with a standard deviation of σ = 3 pixels and supporting {u,v∈[-3,3]} is applied and then discretely normalized, specifically as follows: Where u and v are the relative row and column displacements of the discrete kernel; Z represents the value of the unnormalized two-dimensional discrete Gaussian kernel at the relative displacement (u,v); σ This is the summation normalization factor of the discrete Gaussian kernel over the supporting domain; is the normalized Gaussian kernel; u, v are the relative row and column shifts of the discrete kernel; Smoothing results: in, The grayscale value is the smoothed value. S220, Derivative Superscript Correspondence and Pixel Spacing Explicitization: Based on the five-point central difference formula, calculate the first-order partial derivatives of the smoothed image in the m and n directions: in, For I ( s) Discrete first derivative estimation with respect to column axis n; For I ( s) Discrete first derivative estimate with respect to the row axis m; I ( s) is the image after Gaussian smoothing; S230, Symmetric mapping processing of boundary indices: For all indexes involving out-of-bounds access in S210 and S220, a periodic symmetric reflection rule is uniformly applied: For any row index placeholder r and column index placeholder c, the following is defined: P M =2M,r ＊ =rmod P M ,r ＊ ∈{0,1,...,P M -1}, P N =2N,c ＊ =cmod P N ,c ＊ ∈{0,1,...,P N -1}, Where r and c are placeholders, representing any potentially out-of-bounds row and column integer indices; P M P N The period length is the periodized length of the row and column latitudes, used for symmetrical reflection; r ＊ c ＊ For mapping indices, arbitrary integer indices are transformed back to the periodic intervals {0,1,...,2M-1} and {0,1,...,2N-1} through modulo operation; φ(r;M) and ψ(c;N) are mapping functions that map out-of-bounds indices back to the valid intervals {0,1,...,M-1} and {0,1,...,N-1} through periodicized symmetric reflection, with segmented definition to ensure endpoint symmetry.

4. The road damage detection method based on vision assistance according to claim 3, characterized in that, The step of weighted accumulation of the gradient field within a preset window to construct a two-dimensional symmetric structure tensor and calculating tensor eigenvalues ​​to form a structure energy representation specifically includes: For each pixel (m,n), define the structure tensor J. m,n It is a two-dimensional symmetric matrix: Among them, J m,n Let (m, n) be the structure tensor of the pixel (m, n). For the x components of the structure tensor; The yy component of the structure tensor; These are the xy components of the structure tensor; Apply Gaussian kernel smoothing with a standard deviation ρ = 1.5 pixels to a window size of 5×5 pixels, and then discretize and normalize: Where ρ is the one-dimensional standard deviation used for smoothing within the structure tensor window; Z ρ , The kernel and normalized quantity are the same as those in step S210, but the standard deviation is ρ and the support region is [-2,2]. Calculate the two eigenvalues ​​of a tensor Constructing the local structure energy function 5. The road damage detection method based on vision assistance according to claim 4, characterized in that, The process of obtaining second-order partial derivatives and mixed partial derivatives based on discrete second-order differences, constructing a symmetric curvature tensor, and obtaining a curvature energy representation specifically includes: Calculate the second-order partial derivative using the five-point difference formula: in, The second-order discrete derivative is estimated along the column axis n; The second-order discrete derivative is estimated along the line axis m; The mixed second-order discrete derivative with respect to m and then with respect to n; Constructing a symmetric curvature tensor: Among them, C m,n Let m be a matrix variable, and n be the curvature tensor of the pixel (m,n). Constructing the curvature energy function 6. The road damage detection method based on vision assistance according to claim 5, characterized in that, The standardization of structural energy and curvature energy, determination of adaptive scaling factors, and weighted combination to obtain coupling energy specifically include: S510, Construction of Standardized Quantities for the Entire Map: Where, μ J μ C The overall mean is the arithmetic mean of the structural energy and curvature energy across the entire graph, respectively; σ J σ C The standard deviation of the entire graph is a measure of the dispersion of energy across the entire graph. S520, Constructing dimensionless energy and adaptive scaling factor: If the denominator is 0, then let in, Standardized energy; This is an adaptive scaling factor; S530, Constructing Coupled Energy: Among them, S m,n Let be the coupling energy of pixel (m,n).

7. The road damage detection method based on vision assistance according to claim 6, characterized in that, The calculation of statistics on the coupled energy map, setting a significant anomaly threshold, and generating a damage candidate mask specifically includes: Calculate the average coupling energy μ of the entire image. S With standard deviation σ S : Set the significant anomaly threshold to T. S =μ S +2σ S ; Generate damage mask map Among them, D m,n =1 indicates a suspected damage point, D m,n =0 indicates a non-damaged point.

8. The road damage detection method based on vision assistance according to claim 7, characterized in that, The step of performing four-neighbor connectivity analysis on the mask, and filtering out effective damage regions based on area and shape ratio lower limits, specifically includes: For mask image D m,n Perform 4-neighborhood connectivity analysis and define the region: And connected to a point within the region via several 4-neighborhoods}; where, Let k be the set of pixels connected to the kth 4-neighborhood; k is the region number. Define the set of boundary edges of the region: in, To and The set of adjacent pixel edges (including boundary edges); (m′,n′) is a pixel coordinate pair, the pixel adjacent to (m,n)4; Calculate the perimeter L of the k-th region respectively. k and area A k : Let the minimum analytical scale for first-order and second-order difference and tensor smoothing be p = 5 pixels. Based on this, define: A min =p 2 =25, Among them, A min This is the lower limit of the area used to eliminate excessively small noise regions; η min The lower limit of the shape ratio is used to preserve elongated crack-like regions; Morphological screening criteria and branching operations: If A k ≥A min and Then retain The area is considered an effective damage area; otherwise, delete it. Not included in subsequent output; After this filtering, the remaining regions are reassigned with the numbers k = {1, ..., K} according to their order of appearance; where K is the total number of valid connected regions after filtering.

9. A road damage detection method based on vision assistance according to claim 8, characterized in that, The step of extracting pixel boundaries and geometric centers of the effective damaged area and outputting the detection results in a structured form specifically includes: Construct a set of pixel boundaries: Or ((m′,n′) out of bounds)}; where, Let k be the set of pixel boundary points of the k-th region; Obtain the geometric center point: in, Let be the row and column geometric center coordinates of the k-th region; Each valid area The output is structured data, with the following fields recorded: Region number k, pixel boundary Maximum coupling energy value Geometric center point Construct all valid regions to form the final detection result set. If K=0 after filtering, then output The conclusion is that no road damage areas satisfying the threshold and morphological constraints were detected in this image.

Images