Bridge concrete surface damage identification method and system based on texture analysis

Abstract

This invention discloses a method and system for identifying apparent damage to bridge concrete based on texture analysis, belonging to the interdisciplinary field of bridge engineering inspection and computer vision technology. The method includes: acquiring surface images of bridge components using a drone and generating orthophotos; extracting contrast, correlation, and entropy features based on the gray-level co-occurrence matrix; extracting histogram features based on local binary patterns; detecting corrosion and seepage areas based on the HSV color space; fusing multi-dimensional features and inputting them into a multi-label classifier, simultaneously outputting pixel-level segmentation masks for multiple types of damage such as cracks, spalling, exposed rebar corrosion, and seepage; generating a damage level score and maintenance recommendation report according to bridge technical condition assessment standards. This invention integrates complementary texture features to achieve simultaneous identification of multiple types of damage, outputs pixel-level segmentation results, and automatically generates an assessment report that conforms to engineering specifications.

Description

Technical Field

[0001] This invention relates to the intersection of bridge engineering inspection and computer vision technology, and in particular to a method and system for identifying apparent damage to bridge concrete based on texture analysis. Background Technology

[0002] As a crucial component of transportation infrastructure, the structural safety of bridges directly impacts public safety and property. With the increasing service life of bridges in my country, many are gradually entering a period of frequent structural damage. Regular inspections and health assessments of bridge concrete structures have become essential for ensuring safe bridge operation. During long-term service, bridge concrete structures are affected by various factors such as loads, environmental erosion, and material deterioration, resulting in various types of surface damage, including cracks, spalling, exposed reinforcement corrosion, and seepage. These surface damages not only affect the bridge's appearance and durability but may also reflect potential safety hazards within the structure. Therefore, timely and accurate identification and assessment of concrete surface damage are of great significance for bridge maintenance and management.

[0003] Traditional bridge concrete damage detection mainly relies on manual visual inspection. Inspectors need to be in close contact with the surface of bridge components, identifying and recording damage through visual observation or with the aid of simple tools. This traditional method has several limitations: First, manual inspection is inefficient and labor-intensive, making it difficult to meet the needs of large-scale bridge inspections; second, the results are greatly affected by the inspectors' experience and subjective judgment, making consistency and repeatability difficult to guarantee; third, manual inspection poses significant safety risks and implementation difficulties for hard-to-reach areas such as high piers and the bottom of long-span beams.

[0004] In recent years, with the rapid development of computer vision and deep learning technologies, automatic concrete damage identification methods based on image processing have received widespread attention. Among existing technologies, such as Chinese Patent Publication No. CN113096088A, a deep learning-based concrete structure detection method is disclosed. This method preprocesses concrete structure damage images to form a crack image dataset, inputs the crack image dataset into an improved neural network for training, and uses a ResNet34 model pre-trained on ImageNet as the basis for transfer learning. It extracts basic features by freezing the parameters of the first two modules, and subsequent modules fine-tune the extraction of crack image features, ultimately outputting the classification result of concrete structure damage identification. This method reduces the dependence of the neural network on the number of concrete crack images to a certain extent, lowers the computational cost, and has strong robustness and generalization ability.

[0005] However, the aforementioned existing technologies still have the following shortcomings: First, this method mainly identifies a single damage type, cracks, and cannot simultaneously handle complex situations where multiple damage types such as spalling, exposed rebar corrosion, and seepage are intertwined, while actual bridge inspection often requires comprehensive judgment of multiple damage types; Second, this method uses image-level classification to output overall classification results, failing to provide pixel-level damage area segmentation information, and cannot accurately locate damage boundaries and quantify damage area; Third, this method relies on end-to-end deep neural network feature learning, failing to fully utilize the statistical characteristics of concrete surface texture, and not deeply exploring the texture differences between different damage types; Fourth, this method is not integrated with bridge technical condition assessment standards, and cannot directly output damage level scores and maintenance recommendations that meet engineering practice requirements.

[0006] Therefore, there is an urgent need to develop an intelligent method for identifying apparent damage to bridge concrete that can integrate multi-dimensional texture analysis technology, simultaneously identify multiple types of damage, output pixel-level segmentation results, and automatically generate a damage assessment report that meets the evaluation criteria. Summary of the Invention

[0007] To address the technical problem of complex types of apparent damage to bridge concrete and the difficulty in automatic identification due to the intertwining of multiple types of defects in existing technologies, this invention provides a method and system for identifying apparent damage to bridge concrete based on texture analysis.

[0008] The present invention provides a method for identifying apparent damage to bridge concrete based on texture analysis, comprising the following steps:

[0009] Image acquisition and preprocessing steps: A drone equipped with a high-definition camera is used to acquire full-coverage images of the bridge component surface to obtain the original image sequence; the original image sequence is processed by image stitching and distortion correction to generate an unfolded orthophoto of the bridge component surface.

[0010] Gray-level co-occurrence matrix feature extraction steps: convert the unfolded orthophoto into a grayscale image, calculate the texture feature parameters of the image based on the gray-level co-occurrence matrix algorithm, the texture feature parameters include contrast parameters, correlation parameters and entropy parameters, and construct the gray-level co-occurrence matrix feature vector with the contrast parameters, correlation parameters and entropy parameters.

[0011] Local binary pattern feature extraction steps: Apply the local binary pattern operator to the grayscale image for texture encoding, statistically analyze the histogram distribution of the local binary pattern encoded values, and generate a local binary pattern histogram feature vector.

[0012] Color space feature extraction steps: Convert the unfolded orthophoto from RGB color space to HSV color space, segment the image according to the preset corrosion and seepage color gamut threshold range, and extract the color space feature vector of the corrosion and seepage area.

[0013] Feature fusion and multi-label classification steps: The gray-level co-occurrence matrix feature vector, the local binary pattern histogram feature vector, and the color space feature vector are fused to form a multi-dimensional fused feature vector; The multi-dimensional fused feature vector is input into a pre-trained multi-label classifier, which simultaneously outputs pixel-level segmentation masks for various types of damage such as cracks, spalling, exposed reinforcement and corrosion, and seepage.

[0014] Damage assessment and report generation steps: Based on the pixel-level segmentation mask, the area proportion and spatial distribution information of various types of damage are statistically analyzed; according to the bridge technical condition assessment standards, the area proportion of various types of damage is weighted and scored to generate a component damage level score; a maintenance recommendation report is generated based on the component damage level score.

[0015] This invention also provides a bridge concrete apparent damage identification system based on texture analysis, comprising:

[0016] The image acquisition and preprocessing module is used to control the high-definition camera mounted on the UAV to acquire full-coverage images of the bridge component surface, obtain the original image sequence, and perform image stitching and distortion correction on the original image sequence to generate an unfolded orthophoto of the bridge component surface.

[0017] The gray-level co-occurrence matrix feature extraction module is connected to the image acquisition and preprocessing module. It is used to convert the unfolded orthophoto into a gray-level image and calculate and output the gray-level co-occurrence matrix feature vector, which includes contrast parameters, correlation parameters and entropy parameters, based on the gray-level co-occurrence matrix algorithm.

[0018] The local binary pattern feature extraction module is connected to the image acquisition and preprocessing module. It is used to perform texture encoding on the grayscale image using the local binary pattern operator and output the local binary pattern histogram feature vector.

[0019] The color space feature extraction module is connected to the image acquisition and preprocessing module and is used to convert the unfolded orthophoto image to the HSV color space and extract the color space feature vector according to the rust and seepage color gamut threshold range.

[0020] The multi-label classification and segmentation module is connected to the gray-level co-occurrence matrix feature extraction module, the local binary pattern feature extraction module, and the color space feature extraction module, respectively. It is used to fuse the gray-level co-occurrence matrix feature vector, the local binary pattern histogram feature vector, and the color space feature vector to form a multi-dimensional fused feature vector. The module outputs pixel-level segmentation masks for various types of damage, such as cracks, peeling, exposed reinforcement corrosion, and seepage, through a pre-trained multi-label classifier.

[0021] The damage assessment and report generation module is connected to the multi-label classification and segmentation module. It is used to statistically analyze the area proportion and spatial distribution information of various types of damage, and generate component damage level scores and maintenance recommendation reports based on the bridge technical condition assessment standards.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] This invention integrates three complementary texture feature extraction methods—gray-level co-occurrence matrix, local binary mode, and color space analysis—to comprehensively characterize the texture characteristics of different types of damage on concrete surfaces. Among them, the gray-level co-occurrence matrix feature is sensitive to changes in surface roughness, the local binary mode feature is sensitive to changes in local texture structure, and the color space analysis is sensitive to color changes caused by corrosion and seepage. The synergistic effect of these three methods enables accurate differentiation of various types of damage, such as cracks, spalling, exposed rebar corrosion, and seepage.

[0024] This invention employs a multi-label classifier architecture, which can simultaneously output pixel-level segmentation masks for multiple types of damage. This solves the problem that traditional single-label classification methods cannot handle the intertwining of multiple types of damage, and provides accurate damage region localization and area quantification information.

[0025] This invention combines damage identification results with bridge technical condition assessment standards to automatically generate component damage level scores and maintenance recommendation reports that meet engineering practice requirements, thereby improving the engineering application value of the detection results. Attached Figure Description

[0026] Figure 1 This is a flowchart of the bridge concrete apparent damage identification method based on texture analysis provided in the embodiments of the present invention;

[0027] Figure 2 This is an architecture diagram of a bridge concrete apparent damage identification system based on texture analysis provided in an embodiment of the present invention. Detailed Implementation

[0028] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0029] See Figure 1 This invention provides a method for identifying apparent damage to bridge concrete based on texture analysis. This method integrates multi-dimensional texture feature analysis technology to achieve intelligent identification and assessment of various types of damage on the surface of bridge concrete components. The steps are described in detail below.

[0030] Step S1. Image acquisition and preprocessing steps.

[0031] In one embodiment of the present invention, the image acquisition and preprocessing step is the basic step of the entire damage identification process. Its main purpose is to obtain high-quality surface image data of bridge components and integrate the scattered multi-view images into a unified orthophoto image, so as to provide reliable data input for subsequent texture feature analysis.

[0032] In terms of image acquisition, this embodiment uses a drone equipped with a high-definition camera to collect full-coverage images of the bridge component surface. Preferably, the drone is equipped with a gimbal stabilization system to ensure the camera maintains a stable attitude during flight, reducing image blurring caused by aircraft vibration or attitude changes. In one embodiment of the invention, the high-definition camera has a resolution of not less than 4000×3000 pixels, and a full-frame sensor camera is preferred to obtain higher image quality and a wider dynamic range. The focal length of the camera lens is selected according to the actual inspection requirements. For the inspection of conventional bridge components, a wide-angle lens with a focal length of 24mm to 35mm is recommended to control the degree of distortion while ensuring a sufficient field of view.

[0033] Regarding UAV flight planning, this embodiment develops differentiated flight path planning strategies for different types of bridge components. For pier components, the UAV adopts a circling flight mode, taking multiple circling shots within a range of 0.5m to 1.5m from the pier surface, with a different pitch angle for each shot, to cover the entire surface of the pier from bottom to top. For cap beams and main beam components, the UAV adopts a ribbon-like flight mode, flying at a constant speed along the longitudinal direction of the component while the camera remains facing the component surface. Preferably, the distance between the UAV and the surface of the bridge component is controlled within the range of 0.3m to 2m when acquiring images. This distance range can obtain sufficient coverage area while ensuring image clarity. In one embodiment of the present invention, the forward overlap rate between adjacent images is set to no less than 70%, and the lateral overlap rate is set to no less than 60%. The high overlap rate design ensures sufficient feature matching points during subsequent image stitching.

[0034] In image stitching processing, this embodiment first performs feature point detection and description on the acquired original image sequence. Preferably, a scale-invariant feature transformation algorithm or an accelerated robust feature extraction algorithm is used for feature point extraction. These two algorithms have good invariance to image scale and rotation changes. Subsequently, feature matching between adjacent images is performed based on the extracted feature points. The nearest neighbor distance ratio method is used to eliminate mismatched points and retain reliable matching point pairs. After obtaining sufficient matching point pairs, a random sample consensus algorithm is used to estimate the homography transformation matrix between adjacent images. This matrix describes the geometric mapping relationship between the two images.

[0035] Regarding distortion correction, this embodiment addresses both radial and tangential distortion caused by wide-angle lenses. Radial distortion primarily manifests as barrel or pincushion-shaped deformation in the image edge regions, while tangential distortion is caused by the lens not being perfectly parallel to the sensor plane. In one embodiment of this invention, the following distortion correction model is employed:

[0036] .

[0037] .

[0038] in: and These are the corrected pixel coordinates, in pixels; and These are the original pixel coordinates, in pixels. The radial distance from a pixel to the image center is calculated as follows: The unit is pixels; , , is the radial distortion coefficient, where The dimensions are , The dimensions are , The dimensions are , making , , All are dimensionless quantities, with the same dimensions as the constant term 1, and their values ​​typically range from 1 to 1. This information is obtained from camera calibration. , denoted as the tangential distortion coefficient, where and The dimensions are This ensures that the tangential distortion correction term has the same dimensions as the main term, and its value range is typically [value range missing]. This is also obtained through camera calibration. In the above distortion correction model, by setting the dimensions of each distortion coefficient to the corresponding negative power of the pixel, the terms within the parentheses in the formula are... All are dimensionless quantities, ensuring dimensional consistency in the formula: the left side of the formula The dimension of is pixel, and the terms on the right are... The dimensions are completely consistent. This distortion correction model can effectively eliminate the influence of lens distortion on geometric measurement accuracy, and the geometric accuracy of the corrected image can reach the sub-pixel level.

[0039] After image stitching and distortion correction, this embodiment generates an unfolded orthophoto of the bridge component surface. An orthophoto is a frontal view image that eliminates projection distortion, where each pixel corresponds to a fixed physical size in real space. Preferably, the generated unfolded orthophoto has a resolution of at least 0.5 mm / pixel to ensure clear identification of minute cracks wider than 0.2 mm. The generation of the orthophoto provides a unified and standardized data foundation for subsequent multi-dimensional texture feature extraction.

[0040] In one embodiment of the present invention, the unfolding method of the orthophoto differs for different types of bridge components. For cylindrical piers, a cylindrical unfolding method is used, unfolding the cylindrical surface into a rectangular plane along the generatrix direction. The width of the unfolded orthophoto corresponds to the pier perimeter, and the height corresponds to the pier height. For rectangular cross-section cap beams and main beams, orthophotos of the top, bottom, and side surfaces are generated respectively, with each surface maintaining an independent coordinate system. During the unfolding process, an inverse distance weighted interpolation method is used to process pixel resampling to reduce the impact of unfolding deformation on texture features.

[0041] Preferably, in this embodiment, a mapping relationship between pixel coordinates and component physical coordinates is established simultaneously during the orthophoto generation process and recorded as a georeferenced file. This mapping relationship is used for subsequent damage localization, converting the identified pixel positions of the damaged area into actual physical positions on the component, facilitating maintenance personnel to quickly locate the damaged parts.

[0042] Step S2. Gray-level co-occurrence matrix feature extraction step.

[0043] In one embodiment of the present invention, the gray-level co-occurrence matrix (GLCM) feature extraction step aims to extract feature parameters reflecting the statistical characteristics of concrete surface texture from the unfolded orthophoto. The GLCM is a second-order statistic describing the spatial distribution characteristics of image texture. By statistically analyzing the frequency of pixel pairs with specific gray values ​​and spatial relationships in an image, it reveals features such as roughness, directionality, and regularity of image texture.

[0044] First, the unfolded orthophoto is converted into a grayscale image. For the original RGB color image, a weighted average method is used for grayscale conversion, and the conversion formula is as follows:

[0045] .

[0046] in: coordinates The grayscale value at that location ranges from 1 to 1. Dimensionless; , , These represent the red, green, and blue channel values ​​at that coordinate, with values ​​ranging from [value range missing]. The weighting coefficients are determined based on the differences in human eye sensitivity to different colors, which can produce grayscale images with good visual effects.

[0047] Subsequently, a gray-level co-occurrence matrix is ​​constructed. In one embodiment of the present invention, the number of gray levels in the gray-level image is set to be... Preferably The value is 256. Gray-level co-occurrence matrix. Defined as: at the direction angle Above, at a distance of The two pixels have grayscale values ​​of respectively and The probability of pixel pairs appearing. The formula for constructing the gray-level co-occurrence matrix is:

[0048] .

[0049] in: The elements of the normalized gray-level co-occurrence matrix are dimensionless and take values ​​ranging from 1 to 2. ; and These are the grayscale values ​​of the pixel pairs, with a value range of [value range missing]. ; The pixel pitch is measured in pixels. In this embodiment, the value ranges from 1 to 5 pixels, with a preferred value of 1. The direction angle is expressed in degrees. In this embodiment, four directions are included: 0 degrees, 45 degrees, 90 degrees, and 135 degrees. This indicates the number of pixel pairs that meet the condition. This represents the total number of pixel pairs included in the statistics, used for normalization.

[0050] In one embodiment of the present invention, based on the constructed gray-level co-occurrence matrix, the following three key texture feature parameters are calculated:

[0051] The formula for calculating the contrast parameter is: ,in: The contrast parameter is dimensionless and its value range is [value range missing]. ; The square of the grayscale difference is used as a weighting factor. The contrast parameter reflects the intensity of local grayscale changes in an image, and its physical meaning lies in characterizing the degree of roughness change of the concrete surface. When there is spalling or honeycomb pitting on the concrete surface, the local grayscale changes are drastic, and the contrast value increases significantly; while the contrast value of a smooth and intact concrete surface is relatively small. According to experimental statistics, the contrast value of a normal concrete surface is usually in the range of 50 to 150, while the contrast value of a spalled area can reach 200 to 400.

[0052] The formula for calculating the correlation parameter is: ,in: This is a correlation parameter, dimensionless, with a value range of [value missing]. ; and These are the mean values ​​in the row and column directions of the gray-level co-occurrence matrix, respectively, calculated using the following formula: and ; and These represent the corresponding standard deviations. The correlation parameter reflects the degree of linear correlation of image texture in a specific direction, and its physical meaning lies in characterizing the directional characteristics of linear damage such as cracks. When there are cracks extending along a specific direction in the image, the correlation value in that direction is close to 1, while the correlation value in the vertical direction is lower.

[0053] The formula for calculating the entropy parameter is: ,in: This is the entropy parameter, in bits, with a value range of... ; It is a very small positive number, and its value is [value]. This is used to avoid the zero-value problem in logarithmic operations. The entropy parameter reflects the complexity and disorder of the image texture, and its physical meaning lies in characterizing the distribution characteristics of damage on the concrete surface. When the concrete surface is uniform and smooth, the probability distribution of the gray-level co-occurrence matrix is ​​relatively concentrated, and the entropy value is small; while when there are multiple irregular peelings or crack networks on the surface, the probability distribution tends to be uniform, and the entropy value increases significantly.

[0054] Preferably, to enhance the robustness of the features, this embodiment averages the gray-level co-occurrence matrix feature parameters calculated in the four directions to obtain direction-invariant feature parameters:

[0055] , , Finally, the gray-level co-occurrence matrix feature vector is constructed using the contrast parameter, correlation parameter, and entropy parameter after directional averaging. The feature vector has a dimension of 3 and is used for subsequent feature fusion processing.

[0056] In one embodiment of the present invention, a sliding window approach is used for local gray-level co-occurrence matrix (GLCM) feature extraction to obtain feature representations with higher spatial resolution. Specifically, the sliding window size is set to 64×64 pixels, and the window sliding step size is 32 pixels. The GLCM feature parameters are calculated independently for each window, ultimately forming a feature map that spatially corresponds to the original image. The choice of window size requires a trade-off between feature stability and spatial resolution: a window that is too small results in insufficient statistical samples and large fluctuations in feature values; a window that is too large reduces spatial resolution and makes it difficult to capture local damage. The 64×64 pixel window size selected in this embodiment can contain enough pixel samples for reliable statistical analysis while maintaining sensitivity to the boundaries of damaged areas.

[0057] Preferably, this embodiment also introduces a multi-scale gray-level co-occurrence matrix (GLCM) feature extraction strategy. GLCM features are calculated at three scales: the original resolution, half-resolution, and quarter-resolution. The features from the three scales are then concatenated to form a 9-dimensional multi-scale GLCM feature vector. Multi-scale features can simultaneously capture high-frequency texture changes from subtle cracks and low-frequency texture changes from large-area peeling, improving the expressive power of the features.

[0058] Step S3. Local binary pattern feature extraction step.

[0059] In one embodiment of the present invention, the local binary pattern feature extraction step aims to extract feature information describing the local texture structure from a grayscale image. Local binary pattern is an efficient texture description operator that encodes the local texture structure into binary values ​​by comparing the grayscale relationship between the center pixel and its neighboring pixels. It has the advantages of simple computation and good illumination invariance.

[0060] This embodiment employs a rotation-invariant uniform local binary pattern operator for texture encoding. Preferably, the number of neighborhood sampling points for the local binary pattern operator... The number of sampling points can be set to 8 or 16; this embodiment preferably uses 8 sampling points; sampling radius The sampling radius is set to 1 to 3 pixels; in this embodiment, a sampling radius of 1 pixel is preferred. The basic local binary pattern coding calculation formula is:

[0061] ,in: Center pixel The local binary pattern encoded value at a given location is an unsigned integer, and its value range is [value range missing]. ; This is the number of neighborhood sampling points, dimensionless, and takes the value of 8 or 16. The sampling radius is expressed in pixels and ranges from 1 to 3. The grayscale value of the center pixel, with a value range of [value range missing]. ; For the first The gray values ​​of each neighboring sampling point are obtained by circular sampling, and their coordinates are calculated using the following formula: ; For symbolic functions, it is defined as: For sampling points with non-integer coordinates, bilinear interpolation is used to calculate their gray values.

[0062] To achieve rotation invariance, this embodiment further employs a rotation-invariant uniform local binary pattern. The determination of the uniform pattern is based on the number of transitions in the encoded value, defined as: ,in: The number of jumps is dimensionless and its value range is [value range missing]. When the number of transitions is no more than two, the encoded value is marked as a uniform mode, with a clear meaning of local texture structure; when the number of transitions exceeds two, it is marked as a non-uniform mode and uniformly classified into one category. For In this case, there are 58 uniform patterns (including 2 with 0 transitions and 56 with 2 transitions), plus 1 non-uniform pattern, for a total of 59 pattern categories.

[0063] Rotation invariance is achieved by cyclically shifting the uniform pattern to its minimum value:

[0064] .

[0065] in: For rotationally invariant, uniform, locally binary model values, dimensionless, for The range of values ​​is There are 10 possible values. This processing makes the local binary pattern invariant to image rotation, enhancing the robustness of the features.

[0066] After encoding the local binary patterns of the entire image, the frequency of each pattern category is counted, and a local binary pattern histogram is generated. The histogram normalization formula is: ,

[0067] in: For the first The normalized frequency of the pattern, dimensionless, with a value range of . ; For the first The number of times a pattern appears in an image; For rotationally invariant uniform local binary modes, the total number of modes is [number]. ), Ultimately, the local binary pattern histogram feature vectors... , dimension This feature vector can effectively characterize the local texture structure of concrete surfaces: for spalling and honeycomb pitted areas, due to the presence of numerous tiny pits and irregular edges on the surface, the local binary pattern histogram shows a relatively uniform distribution; while for intact and smooth concrete surfaces, the histogram distribution is concentrated in a few pattern categories.

[0068] Step S4. Color space feature extraction step.

[0069] In one embodiment of the present invention, the color space feature extraction step is specifically designed for the detection of corrosion seepage damage. Corrosion seepage is a surface discoloration phenomenon caused by the corrosion of internal steel reinforcement or the penetration of external rust into the concrete surface, typically manifesting as yellowish-brown or reddish-brown contamination bands. The main characteristic of this type of damage is color change rather than texture change; therefore, color space analysis methods are required for detection.

[0070] First, the unfolded orthophoto image is converted from the RGB color space to the HSV color space. The HSV color space decomposes color information into three independent components: hue, saturation, and lightness, which better aligns with human color perception and makes it easier to set color thresholds. The RGB to HSV conversion formula is: , ,

[0071] .

[0072] in: , , These are the normalized RGB components, dimensionless, with a value range of [value range missing]. ; This refers to the hue component, measured in degrees, with a value range of [value range missing]. ; The saturation component is dimensionless and its value range is [value range missing]. ; The value is the lightness component, which is dimensionless and has a range of values. .

[0073] In one embodiment of the present invention, a color gamut threshold range is set for pixel-level segmentation based on the yellowish-brown characteristic of rust seepage. The specific setting of the rust seepage color gamut threshold range in the HSV color space is: hue component... The range of values ​​is This range covers a chromatic area from orange-yellow to brownish-red, corresponding to the typical color of rust; saturation component The range of values ​​is Exclude gray-white areas with excessively low saturation and artificially painted areas with excessively high saturation; lightness component The range of values ​​is Eliminate overly dark shadow areas and overly bright reflective areas.

[0074] The formula for calculating the segmentation mask of the corrosion and seepage zone is: ,in: For pixels The value of the rust segmentation mask at the location can be either 0 or 1; , This refers to the hue threshold range; , This refers to the saturation threshold range; , This refers to the brightness threshold range.

[0075] Furthermore, in order to filter out isolated noise points and smooth the boundaries of rusted areas, this embodiment performs morphological processing on the initial segmentation mask: ,in: A mask is used to divide the rust after treatment; For morphological erosion operations; For morphological dilation operations; As a corrosive structural element, it is used A circular structural element of pixels; This is an expansion structuring element, identical to the corrosion structuring element. This morphological opening operation effectively eliminates isolated noise points smaller than the structuring element size while preserving the shape of larger connected regions.

[0076] Finally, color space feature vectors are extracted based on the rust segmentation mask, including the average hue of the rusted area. Average saturation Percentage of rusted area Equal statistical measures constitute the feature vector of the color space. The dimension is 3.

[0077] Step S5. Feature fusion and multi-label classification steps.

[0078] In one embodiment of the present invention, the feature fusion and multi-label classification steps fuse the aforementioned three complementary texture features and simultaneously output pixel-level segmentation masks of multiple types of damage through a multi-label classifier.

[0079] First, the gray-level co-occurrence matrix feature vector, the local binary mode histogram feature vector, and the color space feature vector are concatenated and fused to form a multi-dimensional fused feature vector: ,in: To fuse feature vectors, the dimension is (when (Time). The three features describe the surface properties of concrete from different perspectives and have good complementarity: the gray-level co-occurrence matrix feature reflects the macroscopic texture statistical characteristics, the local binary mode feature reflects the microscopic texture structural characteristics, and the color space feature reflects the color change characteristics.

[0080] In one embodiment of the present invention, the multi-label classifier is implemented using a deep convolutional neural network structure, comprising two main components: a feature encoder and a multi-branch segmentation decoder. The feature encoder receives a multi-dimensional fused feature vector as auxiliary input, which is fused with the original image features to extract high-level semantic features. The multi-branch segmentation decoder includes decoding branches equal to the number of damage types. In this embodiment, there are four branches, corresponding to four types of damage: cracks, spalling, exposed reinforcement and corrosion, and seepage. Each decoding branch independently outputs a pixel-level segmentation mask corresponding to the damage type.

[0081] The multi-label classifier is trained using a multi-task learning strategy, and the total loss function is defined as: ,in: The total loss function value is dimensionless. For the first Binary cross-entropy loss for damage-like classes; The Dice loss is used to address class imbalance issues. The weighting coefficients for various types of damage are dimensionless and set according to the sample proportion of each type of damage in the training data. The weighting coefficients for cracks, spalling, exposed rebar corrosion, and seepage are set to [values ​​to be filled in]. , , , ; The Dice loss weight is dimensionless and has a value of 0.5.

[0082] The formula for calculating the binary cross-entropy loss is: ,in: This represents the total number of pixels in the image. For the first The first pixel The actual label for the type of damage, with a value of 0 or 1; The model predicts the first The probability of damage type, with a value range of 100%. .

[0083] The trained multi-label classifier infers from the input image and outputs pixel-level segmentation masks for four types of damage: cracks, spalling, exposed reinforcement and corrosion, and seepage, denoted as follows: , , and Each mask has the same spatial resolution as the input image, and the pixel value is 0 or 1, indicating whether the pixel belongs to the corresponding damage type.

[0084] Step S6. Damage assessment and report generation steps.

[0085] In one embodiment of the present invention, the damage assessment and report generation step is based on the aforementioned multi-type damage pixel-level segmentation mask, and performs damage area statistics, grade assessment and maintenance suggestion report generation.

[0086] First, the area percentage of various types of damage is statistically analyzed based on the pixel-level segmentation mask. Let the physical resolution of the unfolded orthophoto image of the component surface be... (Unit: m / pixel), total number of pixels in the image is The formulas for calculating the area and area percentage of various types of damage are as follows: , ,in: For the first The area of ​​the type of damage, in m 2 ; For the first The pixel in the first The value (0 or 1) in the damage mask; The total surface area of ​​the component, in m². 2 ; For the first The percentage of area affected by this type of lesion, expressed in percent.

[0087] In one embodiment of the present invention, a component damage level score is generated by weighting the area proportion of various types of damage according to the bridge technical condition assessment standard. Preferably, the bridge technical condition assessment standard adopts the assessment method specified in the Highway Bridge and Culvert Maintenance Specification JTG 5120. The formula for calculating the component damage level score is as follows: ,in: The component damage level is scored, with a value range of [value range missing]. The higher the score, the better the condition of the component. For the first The weighting coefficients for damage types are dimensionless and set according to the degree of impact of each type of damage on structural safety. The weighting coefficients for cracks, spalling, exposed rebar corrosion, and seepage are set to [values ​​to be filled in]. , , , ,satisfy ; For the first The mapping function from the proportion of damage area to the deduction value adopts a piecewise linear function form.

[0088] Taking crack damage as an example, its deduction mapping function is:

[0089] .

[0090] The piecewise function design reflects the nonlinear relationship between the proportion of damaged area and the degradation of structural performance: when the damaged area is small, the impact on the structure is limited and the deduction increases slowly; when the damaged area exceeds the critical value, the deduction increases rapidly, reflecting the rapid deterioration of structural performance.

[0091] Similarly, the deduction mapping function for spalling damage is:

[0092] .

[0093] in: The percentage of area affected by spalling damage is expressed as a percentage. The threshold for spalling damage is set slightly more leniently than that for crack damage. This is because minor surface spalling has a relatively small impact on the structural load-bearing capacity, but large-area spalling may lead to the loss of the concrete cover for the reinforcing steel, which requires serious attention.

[0094] The deduction mapping function for exposed rib corrosion damage is:

[0095] .

[0096] in: The percentage of exposed rebar corrosion damage is expressed as a percentage. Exposed rebar corrosion is one of the most serious types of damage to concrete structures. Even a small percentage requires serious attention. Therefore, the deduction function for this type of damage has the steepest slope and the lowest initial deduction threshold. When the percentage of exposed rebar corrosion exceeds 3%, the deduction value quickly reaches the upper limit, indicating that the component is in a severely deteriorated state.

[0097] The deduction mapping function for dialysis damage is:

[0098] .

[0099] in: The percentage represents the area affected by dialysis damage. Dialysis damage primarily reflects issues with the density and durability of concrete. While its impact on structural safety is relatively small when occurring alone, it is often an early sign of internal structural deterioration and thus has significant warning value. Therefore, the weighting coefficient for dialysis damage is set to the minimum, but the detection threshold is set relatively leniently to facilitate the timely detection of potential problems.

[0100] In one embodiment of the present invention, when calculating the damage level score of a component, the mutual influence and superposition effect between multiple types of damage also need to be considered. When multiple types of damage exist in the same area, a correction coefficient is used for adjustment. ,in: The revised component damage level score; The superposition correction factor is dimensionless and has a value of 0.05. The number of regions with overlapping damage is determined by a logical AND operation of the pixel-level mask. This correction mechanism reflects the accelerated degradation effect of structural performance caused by the superposition of multiple types of damage.

[0101] Finally, a repair recommendation report is generated based on the component damage level score. Repair priority is determined based on two preset thresholds:

[0102] .

[0103] in: The first threshold is set at 60 points. A score below this value indicates that the component is severely damaged and requires emergency treatment. The second threshold is set at 80 points. A score higher than this indicates that the component is in good condition and only needs to be observed periodically.

[0104] Preferably, the maintenance recommendation report includes the following: component identification information (including component type, location number, inspection date, etc.); heat maps of various types of damage, visually displaying the spatial distribution characteristics of the damage; damage level assessment results, including the area ratio of various types of damage and the overall score; maintenance priority recommendations, providing specific maintenance suggestions based on the scoring results.

[0105] See Figure 2 This invention also provides a bridge concrete apparent damage identification system based on texture analysis. This system is used to implement the damage identification method described in the above embodiments. The system includes an image acquisition and preprocessing module 1, a gray-level co-occurrence matrix feature extraction module 2, a local binary pattern feature extraction module 3, a color space feature extraction module 4, a multi-label classification and segmentation module 5, and a damage assessment and report generation module 6.

[0106] The image acquisition and preprocessing module 1 is the system's data input terminal. Its main functions include controlling the drone's flight and camera capture, receiving and storing raw image data, and performing image stitching and distortion correction processing. In one embodiment of the present invention, this module includes a drone flight control subunit, an image receiving subunit, and an image preprocessing subunit. The drone flight control subunit is responsible for planning the flight path and controlling the drone to fly along a preset trajectory, while simultaneously triggering the high-definition camera to capture images at a designated location. The image receiving subunit receives image data acquired by the drone in real time via a wireless image transmission link, or reads the image data from a memory card after the flight. The image preprocessing subunit performs processing operations such as feature point extraction, image registration, distortion correction, and orthophoto generation, outputting an orthophoto of the unfolded surface of the bridge component.

[0107] Preferably, the image acquisition and preprocessing module 1 adopts a modular design, supporting the integration with different models of UAVs and camera equipment. This module also features flight mission management capabilities, allowing the saving and retrieval of flight parameters from historical inspection missions, facilitating consistent acquisition conditions during periodic inspections of the same bridge. The generated orthophotos are stored in standard image formats (such as TIFF or PNG) along with geographic coordinate information and physical resolution parameters, providing necessary metadata for subsequent processing.

[0108] The gray-level co-occurrence matrix feature extraction module 2 is connected to the image acquisition and preprocessing module 1. It receives the unfolded orthophoto as input and outputs the gray-level co-occurrence matrix feature vector. The core functions of this module include gray-level image conversion, gray-level co-occurrence matrix construction, and texture feature parameter calculation.

[0109] In one embodiment of the present invention, the module first converts a color orthophoto into a 256-level grayscale image, and then constructs a grayscale co-occurrence matrix according to preset pixel spacing and orientation angle parameters. Preferably, the module supports configuring pixel spacing parameters within the range of 1 to 5 pixels, and supports any combination of four orientation angles: 0 degrees, 45 degrees, 90 degrees, and 135 degrees. Based on the constructed grayscale co-occurrence matrix, the module calculates three texture feature parameters: contrast, correlation, and entropy, and performs orientation averaging to obtain orientation-invariant feature values.

[0110] This module is optimized for large-size orthophotos, employing a sliding window approach for local feature extraction. The window size is configurable from 32×32 pixels to 128×128 pixels, and the stride between adjacent windows is configurable from 1 / 2 to 1 / 4 of the window size to achieve a balance between feature resolution and computational efficiency. The gray-level co-occurrence matrix feature vectors output by the module are organized into feature maps according to their spatial location, maintaining a correspondence with the original image.

[0111] The Local Binary Pattern Feature Extraction Module 3 is connected to the Image Acquisition and Preprocessing Module 1. It receives grayscale images as input and outputs local binary pattern histogram feature vectors. The core functions of this module include local binary pattern encoding calculation and histogram statistics.

[0112] In one embodiment of the present invention, the module implements a rotation-invariant uniform local binary pattern operator, supporting two configurations: 8 sampling points and 16 sampling points, and a sampling radius configuration ranging from 1 to 3 pixels. The module performs local binary pattern encoding calculations pixel by pixel on the grayscale image, and then counts the occurrence frequency of each pattern category within a preset sliding window to generate a normalized histogram feature vector.

[0113] Preferably, this module employs a lookup table acceleration technique to optimize computational efficiency. It pre-calculates the local binary pattern encoding values ​​corresponding to all possible neighborhood grayscale combinations, and quickly retrieves the encoding results at runtime via table lookup, avoiding repetitive comparisons and bitwise operations. The local binary pattern histogram feature vectors output by the module are also organized into feature map form according to spatial location.

[0114] The color space feature extraction module 4 is connected to the image acquisition and preprocessing module 1. It receives color orthophotos as input and outputs color space feature vectors and a preliminary segmentation mask for corrosion and seepage. The core functions of this module include color space conversion, threshold segmentation, and morphological processing.

[0115] In one embodiment of the present invention, the module first converts the RGB color image into an HSV color space representation, and then performs pixel-level segmentation of the image according to a preset rust and seepage color gamut threshold range to obtain a preliminary rust and seepage area mask. Preferably, the color gamut threshold range can be adjusted by the user according to the actual detection environment to adapt to different lighting conditions and differences in concrete material. The module also performs morphological opening operations to denoise and smooth the preliminary mask.

[0116] The color space feature vector output by this module includes statistical measures such as the average hue, average saturation, and area ratio of the rusted region, which are used for subsequent feature fusion. Simultaneously, the output rust and seepage segmentation mask serves as a reference input for a multi-label classifier, helping to improve classification accuracy.

[0117] The multi-label classification and segmentation module 5 is connected to the gray-level co-occurrence matrix feature extraction module 2, the local binary pattern feature extraction module 3, and the color space feature extraction module 4, respectively. It receives three feature vectors as input and outputs pixel-level segmentation masks for multiple types of damage. This module is the core processing unit of the system and integrates a pre-trained deep learning model.

[0118] In one embodiment of the present invention, the module first concatenates and fuses three feature vectors to form a multi-dimensional fused feature vector. The fused feature vector and the original image are then input into a pre-trained multi-label classifier network. The network adopts an encoder-decoder architecture, where the encoder extracts multi-scale semantic features, and the decoder outputs pixel-level segmentation masks for four types of damage—cracks, spalling, exposed reinforcement and corrosion, and seepage—through four parallel branches.

[0119] Preferably, this module supports online model updates and incremental learning. As new labeled data accumulates, the model can be fine-tuned and optimized to continuously improve recognition accuracy. The module also provides a confidence score output function, giving a confidence score for the classification result of each pixel, facilitating manual verification by users to identify areas where the model's prediction is uncertain.

[0120] The damage assessment and report generation module 6 is connected to the multi-label classification and segmentation module 5. It receives multiple types of damage segmentation masks as input and outputs component damage level scores and maintenance recommendation reports. This module is the system's decision output end, transforming technical inspection results into engineering application information.

[0121] In one embodiment of the present invention, the module first calculates the area and area ratio of various types of damage based on the segmented mask, and then calculates the damage level score of the component according to the built-in bridge technical condition assessment standards. The assessment standard parameters are configurable and support assessment according to different specifications. Based on the scoring results, the module automatically determines the maintenance priority and generates a structured maintenance recommendation report.

[0122] Preferably, this module supports multiple report output formats, including complete inspection reports in PDF format, data statistics tables in Excel format, and standard data interfaces that can be imported into bridge management systems. The report content includes sections such as basic component information, inspection condition records, damage distribution heatmaps, damage statistics tables, grade assessment results, and maintenance recommendations, meeting the practical needs of bridge maintenance management.

[0123] The system provided in this embodiment of the invention uses standardized data interfaces for communication between its modules, supporting independent upgrades and replacements of modules. The overall system architecture is designed with scalability in mind, allowing for easy addition of new damage type identification modules or integration of data from other detection devices in the future.

[0124] In one embodiment of the present invention, the system further includes a data management module and a user interaction module. The data management module is responsible for storing, indexing, and querying the detection data, supporting data organization and retrieval by multiple dimensions such as bridge number, component type, and detection date. The detection data is stored in a structured database format, and raw images and intermediate processing results are managed using a distributed file system, supporting efficient access to massive amounts of data. The data management module also provides data backup and recovery functions to ensure the security and integrity of the detection data.

[0125] The user interaction module provides a graphical user interface (GUI) that supports functions such as task configuration, real-time progress monitoring, result visualization, and report export. The interface is responsive and can be used on desktop computers, tablets, and large-screen displays. For result visualization, the module allows damage identification results to be overlaid on orthophotos, with different types of damage labeled in different colors for easy visualization of damage distribution. Users can manually verify and correct the identification results through the interface; the corrected results can then be used for incremental learning and optimization of the model.

[0126] Preferably, the system supports both cloud deployment and local deployment modes. In cloud deployment mode, computationally intensive image processing and deep learning inference tasks are executed on a cloud server, and the client only needs to upload the raw data and download the processing results, making it suitable for mobile terminals with limited processing capabilities. In local deployment mode, all processing tasks are completed on the local workstation without a network connection, making it suitable for application scenarios with high data security requirements. The system provides a unified application programming interface, and the two deployment modes are transparent to the user, allowing for flexible switching according to actual needs.

[0127] In one embodiment of the present invention, the system also integrates a historical data comparison and analysis function. When periodically inspecting the same bridge component, the system automatically loads historical inspection data, compares and analyzes the current inspection results with historical data, and calculates the area change rate and spatial expansion of various types of damage. The historical comparison and analysis function can reveal the development and evolution trend of damage, providing data support for predictive maintenance decisions. The system uses image registration technology to align orthophotos from different periods to a unified coordinate system, ensuring the accuracy of damage area comparison.

[0128] In addition, the system provides a data interface with the bridge management system, supporting the export of damage detection results in a standard format that conforms to the bridge management system's data specifications, thus achieving seamless integration with existing bridge maintenance management processes. The data interface supports both incremental updates and batch imports to meet the data exchange needs of different application scenarios.

[0129] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for identifying apparent damage of a bridge concrete based on texture analysis, characterized by, Includes the following steps: Image acquisition and preprocessing steps: A drone equipped with a gimbal stabilization system and a high-definition camera with a resolution of no less than 4000×3000 pixels is used to acquire full-coverage images of the bridge component surface. The distance between the drone and the bridge component surface during image acquisition is 0.3m to 2m. The forward overlap rate between adjacent images is no less than 70%, and the lateral overlap rate is no less than 60%, to obtain the original image sequence. The original image sequence is then processed by image stitching and distortion correction to generate an unfolded orthophoto of the bridge component surface. The gray-level co-occurrence matrix feature extraction steps are as follows: the unfolded orthophoto is converted into a grayscale image; the texture feature parameters of the image are calculated based on the gray-level co-occurrence matrix algorithm; the pixel spacing used when constructing the gray-level co-occurrence matrix is ​​1 to 5 pixels; the directional angles include four directions: 0 degrees, 45 degrees, 90 degrees, and 135 degrees; the gray level of the gray-level co-occurrence matrix is ​​256 levels; the texture feature parameters include contrast parameters, correlation parameters, and entropy parameters, wherein the contrast parameter is used to characterize the intensity of local gray-level changes in the image to reflect the degree of change in the roughness of the concrete surface; the correlation parameter is used to characterize the regularity of the image texture in any of the four directions to reflect the linear characteristics of the crack; and the entropy parameter is used to characterize the complexity of the image texture to reflect the disordered characteristics of the spalling area. The gray-level co-occurrence matrix feature vector is constructed using the contrast parameter, correlation parameter, and entropy parameter. Local Binary Pattern Feature Extraction Steps: The grayscale image is texture-encoded using a rotation-invariant uniform local binary pattern operator. The local binary pattern operator has 8 or 16 neighborhood sampling points and a sampling radius of 1 to 3 pixels. The rotation-invariant uniform local binary pattern operator achieves rotation invariance by counting the number of jumps in the encoded values. Encoded values ​​with no more than 2 jumps are marked as uniform patterns. The histogram distribution of the local binary pattern encoded values ​​is statistically analyzed to generate a local binary pattern histogram feature vector. Color space feature extraction steps: Convert the unfolded orthophoto from RGB color space to HSV color space, segment the image according to the preset corrosion and seepage color gamut threshold range, wherein the hue component H in the HSV color space ranges from 10 to 40 degrees, the saturation component S ranges from 0.2 to 0.8, and the lightness component V ranges from 0.3 to 0.9, and extract the color space feature vector of the corrosion and seepage area; Feature fusion and multi-label classification steps: The gray-level co-occurrence matrix feature vector, the local binary pattern histogram feature vector, and the color space feature vector are concatenated and fused to form a multi-dimensional fused feature vector; The multi-dimensional fused feature vector is input into a pre-trained multi-label classifier. The multi-label classifier adopts a deep convolutional neural network structure including a feature encoder and a multi-branch segmentation decoder. The multi-branch segmentation decoder includes decoding branches equal to the number of damage types. Each decoding branch independently outputs a pixel-level segmentation mask corresponding to the damage type, and simultaneously outputs pixel-level segmentation masks for multiple types of damage such as cracks, spalling, exposed reinforcement corrosion, and leaching. Damage assessment and report generation steps: Based on the pixel-level segmentation mask, the area proportion and spatial distribution information of various types of damage are statistically analyzed; according to the bridge technical condition assessment standards specified in Highway Bridge and Culvert Maintenance Specification JTG 5120, the area proportion of various types of damage is weighted and scored. Preset weight coefficients are assigned to the area proportions of crack damage, spalling damage, exposed rebar and corrosion damage, and seepage damage, and then the weighted sum is calculated to generate a component damage level score; a maintenance recommendation report is generated based on the component damage level score. When the component damage level score is lower than a preset first threshold, the maintenance priority is emergency maintenance; when the component damage level score is between the first threshold and a preset second threshold, the maintenance priority is routine maintenance; when the component damage level score is higher than the second threshold, the maintenance priority is observation and monitoring.

2. The method of claim 1, wherein, In the image acquisition and preprocessing steps, for the pier components, the UAV adopts a circling flight mode, taking multiple circling shots within a range of 0.5m to 1.5m from the pier surface, with the central axis of the pier as the center; for the cap beam and main beam components, the UAV adopts a ribbon flight mode, flying at a constant speed along the longitudinal direction of the component, while the camera keeps facing the surface of the component to take pictures; the generated unfolded orthophoto image has a resolution of not less than 0.5mm / pixel.

3. The method of claim 1, wherein, In the gray-level co-occurrence matrix feature extraction step, the average value of the gray-level co-occurrence matrix feature parameters calculated in four directions is taken to obtain the contrast parameter, correlation parameter and entropy parameter that remain unchanged in direction. Local gray-level co-occurrence matrix feature extraction is performed using a sliding window method with a sliding window size of 64×64 pixels and a window sliding step size of 32 pixels.

4. The method of claim 1, wherein, In the color space feature extraction step, the initial segmentation mask of the rust and seepage area is subjected to morphological opening operation processing. The corrosion structure element and the expansion structure element are both 3×3 pixel circular structure elements to filter out isolated noise points and smooth the boundary of the rust area. The color space feature vector includes the average hue, average saturation and rust area ratio of the rust area.

5. The method of claim 1, wherein, In the feature fusion and multi-label classification steps, the feature encoder receives the multi-dimensional fused feature vector and extracts high-level semantic features; the training of the multi-label classifier adopts a multi-task learning strategy, and the total loss function includes the weighted sum of the binary cross-entropy loss and Dice loss of various types of damage. The loss weight coefficients for cracks, spalling, exposed reinforcement corrosion, and seepage are set to 1.0, 1.2, 1.5, and 1.3, respectively.

6. The method of claim 1, wherein, In the damage assessment and report generation steps, the weighting coefficients for cracks, spalling, exposed rebar corrosion, and seepage are set to 0.35, 0.25, 0.30, and 0.10, respectively, and the sum of each weighting coefficient is equal to 1; the component damage level score ranges from 0 to 100 points; when multiple types of damage exist in the same area, a superposition correction coefficient is used to adjust the component damage level score, and the superposition correction coefficient is set to 0.

05.

7. The method of claim 1, wherein, The maintenance recommendation report includes component identification information, heat maps of various types of damage distribution, damage level assessment results, and corresponding maintenance priority recommendations; the first threshold is 60 points, and the second threshold is 80 points.

8. The method of claim 1, wherein, In the image acquisition and preprocessing steps, a distortion correction model is used to correct the radial and tangential distortions generated by the wide-angle lens. The radial and tangential distortion coefficients in the distortion correction model are obtained by camera calibration. During the orthophoto generation process, a mapping relationship between pixel coordinates and component physical coordinates is established and recorded as a geographic reference file, which is used to convert the pixel positions of the identified damaged areas into actual physical positions on the components.

9. The method of claim 1, wherein, In the gray-level co-occurrence matrix feature extraction step, gray-level co-occurrence matrix features are calculated at three scales: the original resolution, 1 / 2 resolution, and 1 / 4 resolution. The features at the three scales are concatenated to form a multi-scale gray-level co-occurrence matrix feature vector.

10. A bridge concrete apparent damage identification system based on texture analysis, used to implement the method described in any one of claims 1-9, characterized in that, include: The image acquisition and preprocessing module is used to control a drone equipped with a gimbal stabilization system to carry a high-definition camera with a resolution of not less than 4000×3000 pixels to acquire full-coverage images of the bridge component surface, obtain the original image sequence, and perform image stitching and distortion correction processing on the original image sequence to generate an unfolded orthophoto of the bridge component surface. The gray-level co-occurrence matrix feature extraction module is connected to the image acquisition and preprocessing module. It is used to convert the unfolded orthophoto into a gray-level image, calculate and output a gray-level co-occurrence matrix feature vector based on the gray-level co-occurrence matrix algorithm, including contrast parameter, correlation parameter and entropy parameter. The contrast parameter reflects the degree of change in the roughness of the concrete surface, the correlation parameter reflects the linear characteristics of the crack, and the entropy parameter reflects the disordered characteristics of the spalling area. The local binary pattern feature extraction module is connected to the image acquisition and preprocessing module. It is used to perform texture encoding on the grayscale image using a rotation-invariant uniform local binary pattern operator and output the local binary pattern histogram feature vector. The color space feature extraction module, connected to the image acquisition and preprocessing module, is used to convert the unfolded orthophoto image to the HSV color space and extract the color space feature vector based on the rust and seepage color gamut threshold range of the hue component H (ranging from 10 to 40 degrees), the saturation component S (ranging from 0.2 to 0.8), and the lightness component V (ranging from 0.3 to 0.9). The multi-label classification and segmentation module is connected to the gray-level co-occurrence matrix feature extraction module, the local binary pattern feature extraction module, and the color space feature extraction module, respectively. It is used to fuse the gray-level co-occurrence matrix feature vector, the local binary pattern histogram feature vector, and the color space feature vector to form a multi-dimensional fused feature vector. The module outputs a pixel-level segmentation mask for various types of damage, including cracks, spalling, exposed reinforcement corrosion, and seepage, through a pre-trained multi-label classifier including a feature encoder and a multi-branch segmentation decoder. The multi-branch segmentation decoder includes decoding branches equal to the number of damage types. The damage assessment and report generation module, connected to the multi-label classification and segmentation module, is used to statistically analyze the area proportion and spatial distribution information of various types of damage. Based on the bridge technical condition assessment standards specified in Highway Bridge and Culvert Maintenance Specification JTG 5120, the area proportion of various types of damage is assigned a preset weight coefficient and then weighted and summed to generate a component damage level score and maintenance suggestion report. Based on the comparison results of the component damage level score with the first threshold and the second threshold, the module outputs maintenance priority suggestions for emergency maintenance, routine maintenance, or observation and monitoring.

Images