CN118823481A - A bridge detection image recognition method and system based on artificial intelligence - Google Patents

Abstract

The present application provides an artificial intelligence-based bridge detection image recognition method and system, including: for a bridge image after visibility enhancement, using a multi-scale Retinex algorithm to enhance contrast, highlighting the outline of the bridge structure, balancing the image brightness distribution through an adaptive histogram equalization method, eliminating uneven illumination and shadow interference, and obtaining a bridge image after illumination normalization; for a bridge image after scattering correction, extracting multi-scale features of the bridge structure through a morphological top-hat operation, clustering the bridge surface texture using a local self-similarity metric, and obtaining a bridge structure image after scale normalization and texture segmentation; according to a bridge skeleton structure diagram, determining the anisotropic area of the bridge surface through a structural tensor analysis, extracting high-order texture features of the anisotropic area using a grayscale co-occurrence matrix, and constructing a texture feature dictionary of bridge damage.

Description

A bridge detection image recognition method and system based on artificial intelligence

Technical Field

The present invention relates to the field of information technology, and in particular to an artificial intelligence-based bridge detection image recognition method and system.

Background Art

In bridge damage detection, the high mountain bridges located in heavy industrial areas are surrounded by haze all year round, with low visibility. It is necessary to explore bridge damage detection under haze weather, which has a serious impact on the quality of image acquisition. Haze reduces image contrast and blurs edges, making it difficult to identify key damage features. At the same time, suspended particles in haze will form noise on the image, interfering with the extraction of texture and shape features. In this case, traditional damage detection methods based on gradient and texture analysis are difficult to work. On the other hand, in order to overcome the influence of haze, preprocessing to enhance image contrast and sharpness may be used. However, excessive enhancement may lead to the generation of pseudo edges and false textures, which will affect the identification of real damage features. In addition, while removing the haze effect, the dehazing algorithm may also blur or remove some minor damage features such as cracks. In practical applications, it is also necessary to consider the changes in image quality under different degrees of haze weather. Mild haze may only affect the imaging of distant bridge structures, while heavy haze will cause the overall image to be blurred. This requires the damage detection algorithm to be able to adapt to different degrees of image quality degradation, while ensuring detection accuracy without generating too many false alarms. Therefore, how to accurately extract the characteristic information of bridge damage under haze weather conditions while avoiding the negative impact of image enhancement has become a technical problem that needs to be solved urgently. This requires innovation and optimization in multiple links such as image preprocessing, feature extraction and damage identification.

Summary of the invention

The present invention provides a bridge detection image recognition method based on artificial intelligence, which mainly includes:

Obtain the original bridge image taken in haze weather, and judge whether the visibility meets the requirements of bridge recognition and detection based on the image contrast and grayscale histogram distribution. If not, use the dark channel prior theory to estimate the global atmospheric illumination value, and combine the guided filtering algorithm for adaptive defogging to obtain the bridge image with enhanced visibility.

For the bridge image after visibility enhancement, the multi-scale Retinex algorithm is used to enhance the contrast and highlight the outline of the bridge structure. The adaptive histogram equalization method is used to balance the image brightness distribution, eliminate uneven illumination and shadow interference, and obtain the bridge image after illumination normalization.

For the bridge image after illumination normalization, the weighted least squares method is used to fit the bridge surface model at multiple scales. The scattered light interference caused by haze is eliminated through surface reconstruction. The haze parameters are estimated using the degradation function. The missing information of some bands is adaptively compensated to obtain the bridge image after scattering correction.

For the bridge image after scattering correction, the multi-scale features of the bridge structure are extracted through morphological top-hat operation, and the bridge surface texture is clustered using local self-similarity measurement to obtain the bridge structure image after scale normalization and texture segmentation.

The improved Canny operator is applied to the bridge structure image to detect the edge contour, and the background interference is eliminated by edge connection and closure to obtain a complete bridge structure edge map. The multi-scale line features of the bridge structure are detected by Hough transform, and the dominant structural lines are screened by the probabilistic voting mechanism to obtain the bridge skeleton structure map.

According to the bridge skeleton structure diagram, the anisotropic area of the bridge surface is determined through structural tensor analysis, and the high-order texture features of the anisotropic area are extracted using the gray-level co-occurrence matrix to construct a texture feature dictionary of bridge damage.

The morphological skeleton extraction algorithm is applied to the edge map of the bridge structure to obtain the skeleton line of the bridge connection part. The connectivity analysis of the skeleton line is used to determine whether the bridge structure has a fracture. If there is a fracture, the damage location is determined according to the edge contour of the fracture area, and the damage size is determined according to the fracture length.

The support vector machine regression model is used to evaluate the health status of the bridge based on the structural integrity, surface texture characteristics and quantitative damage indicators of the bridge. The predicted value of the damage degree of each part of the bridge is obtained, the predicted value is mapped to the health status level, and a bridge health assessment report is generated to intuitively present the bridge damage detection results.

The present invention provides a bridge detection image recognition system based on artificial intelligence, which mainly includes:

Image preprocessing module, used to perform defogging and contrast enhancement on the original bridge image under haze weather;

Surface model fitting and scattering correction module, used to fit the bridge surface model, eliminate scattered light interference, and improve image clarity;

The structural feature extraction module is used to extract the edge contour and texture features of the bridge structure and obtain the bridge skeleton structure diagram;

The damage identification and health assessment module is used to analyze the structural integrity of the bridge, evaluate the health status of the bridge, and generate damage detection result reports.

The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

The invention discloses an artificial intelligence-based bridge detection image recognition method. Firstly, by evaluating and enhancing the visibility of the original bridge image, the problem that the original bridge image has low visibility in haze weather and is difficult to meet the bridge recognition and detection requirements is solved. Contrast enhancement, brightness balance and scattering correction are used for the bridge image after visibility enhancement, so as to solve the problems of unclear structural contour and missing information caused by insufficient contrast, uneven brightness distribution, illumination interference and scattered light interference of the bridge image. Secondly, by extracting structural features, extracting line features and detecting edges, the problems of inaccurate scale normalization, texture segmentation, edge detection and structural line detection of the bridge image are solved, and the interference caused by texture, shape features and edge blur is avoided. In addition, by analyzing the structure and extracting the anisotropic region, the problem of difficulty in determining the texture features of the anisotropic region, judging the structural integrity of the bridge connection part and determining the damage position and size in bridge damage detection is solved, so as to realize accurate extraction of bridge damage feature information.

In summary, the present invention not only significantly improves the recognition and analysis capabilities of bridge structures in extreme environments, but also greatly enhances the accuracy and reliability of bridge health assessment, achieves accurate extraction of bridge damage feature information while avoiding the negative effects of image enhancement.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of an artificial intelligence-based bridge detection image recognition method of the present invention.

FIG. 2 is a schematic diagram of an artificial intelligence-based bridge detection image recognition method and system according to the present invention.

FIG. 3 is another schematic diagram of an artificial intelligence-based bridge detection image recognition method and system according to the present invention.

DETAILED DESCRIPTION

The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

As shown in Figures 1-3, the bridge detection image recognition method and system based on artificial intelligence in this embodiment may specifically include:

Step S101, obtain the original image of the bridge taken in haze weather, and judge whether the visibility meets the requirements of bridge recognition and detection based on the image contrast and grayscale histogram distribution. If not, use the dark channel prior theory to estimate the global atmospheric illumination value, and combine the guided filtering algorithm for adaptive defogging to obtain a bridge image with enhanced visibility.

The original image of the bridge taken under haze weather is obtained, and the local contrast of the original image of the bridge is calculated, and the local contrast is obtained by using a local contrast calculation formula; the grayscale histogram distribution is analyzed according to the original image of the bridge, and the mean and variance of the grayscale histogram distribution are calculated; according to the local contrast, mean and variance, whether the visibility of the current image meets the preset bridge recognition detection requirement threshold is judged; if the visibility of the current image does not meet the preset bridge recognition detection requirement threshold, the original image of the bridge is processed by using a dark channel prior theory to obtain a defogged bridge image; the contrast of the defogged bridge image is adjusted, and a linear stretching method is used to obtain a contrast-adjusted bridge image; the contrast-adjusted bridge image is subjected to adaptive histogram equalization processing, the contrast-adjusted bridge image is divided into a plurality of small blocks, each of the small blocks is subjected to histogram equalization, and the processing results are merged by bilinear interpolation to obtain a bridge image with enhanced visibility.

Specifically, the original image of the bridge taken in haze weather is obtained, and the local contrast of the image is calculated. The local contrast calculation formula C = (I_max-I_min)/(I_max+I_min) is used, where I_max and I_min are the maximum and minimum pixel values in the local area, respectively. The grayscale histogram distribution is analyzed, and the mean μ and variance σ^2 of the grayscale value are calculated. According to the calculated contrast C, mean μ and variance σ^2, it is judged whether the visibility of the current image meets the preset thresholds T_c, T_μ and T_σ required for bridge recognition and detection, and the visibility evaluation result is obtained. If the requirements for bridge recognition and detection are not met, the dark channel prior theory is used to process the image and calculate the dark channel image J_dark(x)=min_c(min_y∈Ω(x)(J_c(y))), where J_dark(x) represents the pixel value of the dark channel image at position x, c represents the three RGB color channels, namely red R, green G, and blue B, min_c represents the minimum value among the three RGB channels, y is the pixel position within the local window Ω(x) centered on x, Ω(x) represents the local window centered on x, and J_c(y) represents the pixel value of color channel c at position y. The darkest pixels in the first 0.1% of the dark channel image are selected, and the maximum value of the RGB values corresponding to these pixels in the original image is calculated as the global atmospheric illumination value A. According to the dark channel prior theory, the transmittance t(x) = 1-ωmin_c(min_y∈Ω(x)(I_c(y)/A_c)) is estimated, where ω is a tuning parameter, usually 0.95, I_c(y) represents the pixel value of the foggy image of color channel c at position y, and A_c represents the atmospheric light value in color channel c. Based on the obtained atmospheric illumination value A and transmittance t(x), the filter kernel function K(i,j) = ((1+ε)/|ω|^2)∑(I_i-μ_k)(I_j-μ_k) of the guided filtering algorithm is constructed, where I is the guided image, I_i and I_j represent the pixel values at positions i and j in the image, respectively, μ_k is the local mean, |ω| is the window size, and ε is the regularization parameter. The image is subjected to adaptive defogging to generate the defogging bridge image J(x) = (I(x)-A)/max(t(x), t_0)+A, where t_0 is the lower threshold and I(x) represents the pixel value of the original foggy image at position x to prevent zero division errors. The contrast of the defogging bridge image is adjusted by using a linear stretching method J_out = (J_i n-J_min)(255/(J_max-J_min)), where J_i n represents the original pixel value of the input defogging bridge image, J_min is the minimum pixel value in the input image, and J_max is the maximum pixel value in the input image. Adaptive histogram equalization is performed to divide the image into multiple small blocks, and histogram equalization is performed on each small block. Then, bilinear interpolation is used to merge the processing results to improve the image detail expression and obtain the bridge image with enhanced visibility. After obtaining the original bridge image taken in haze weather, the local contrast of the image is calculated, a sliding window of 10x10 pixels is selected, and the local contrast calculation formula is applied in each window. If the maximum pixel value in a window is 200 and the minimum pixel value is 50, the local contrast of the window is C = (200-50)/(200+50) = 0.6. Calculate the grayscale histogram for the entire image, assuming that the grayscale mean μ = 128 and the variance σ^2 = 1600. Set the preset thresholds T_c = 0.4, T_μ = 100, T_σ = 50 to determine whether the visibility of the current image meets the requirements. If not, the dark channel prior theory is used to process the image, and a 15x15 minimum filter is used when calculating the dark channel image. Select the darkest 0.1% pixels in the dark channel image, assuming that the corresponding RGB maximum value in the original image is 220, that is, determine the global atmospheric illumination value A = 220. When estimating the transmittance, set the adjustment parameter ω = 0.95 to obtain a preliminary transmittance map. Construct the filter kernel function of the guided filtering algorithm, set the window size |ω| = 16x16, the regularization parameter ε = 0.001, and use the original image as the guide image. Apply the guided filter ref i ned_t = guidedFilter(I, rough_t, 16, 0.001), where I is usually the input original foggy image, rough_t is the transmittance obtained by preliminary estimation, and obtain the optimized transmittance map. Perform defogging, set the lower threshold t_0 = 0.1, and obtain the defogged bridge image. Linearly stretch the defogged image. Assuming that the pixel value range of the defogged image is [30, 220], the stretched pixel value J_out = (J_i n-30)*(255/(220-30)). Finally, adaptive histogram equalization is performed to divide the image into 8x8 small blocks, each of which is 32x32 pixels in size. Histogram equalization is performed on each small block, and then bilinear interpolation is used to merge the processing results to obtain the final bridge image with enhanced visibility.

Step S102, for the bridge image after visibility enhancement, a multi-scale Retinex algorithm is used to enhance contrast, highlight the outline of the bridge structure, balance the image brightness distribution through an adaptive histogram equalization method, eliminate uneven illumination and shadow interference, and obtain a bridge image after illumination normalization.

According to the size of the bridge image, the Gaussian filter group parameters of the multi-scale Retinex algorithm are obtained, and the parameters include three scale values, which are 5%, 25% and 75% of the number of pixels on the long side of the image respectively; the Gaussian filter group is used to perform Gaussian blur processing on the input image to obtain three blurred images of different scales; the logarithmic difference between the input image and the blurred images of each scale is calculated to obtain Retinex output images of three scales; the Retinex output images of the three scales are weighted averaged to obtain a fused Retinex output image; nonlinear mapping is performed on the fused Retinex output image, and a sigmoid function is used to perform color restoration and dynamic range compression to obtain a contrast-enhanced bridge image; the contrast-enhanced bridge image is divided into a number of sub-blocks, and the cumulative distribution function is calculated for each sub-block to obtain a bridge image after illumination normalization; the structural similarity index between the bridge image after illumination normalization and the input image is calculated; if the structural similarity index is lower than a preset threshold, the Gaussian filter group parameters, sigmoid function and the input image are adjusted. d function parameters and cumulative distribution function parameters are reprocessed until the structural similarity index meets the quality requirements.

Specifically, for the bridge image after visibility enhancement, a Gaussian filter group of the multi-scale Ret i nex algorithm is constructed, and three scale parameters are set according to the image size and bridge structure characteristics, which are 5%, 25% and 75% of the number of pixels on the long side of the image. The input image is Gaussian blurred to obtain three blurred images of different scales. The logarithmic difference between the input image and the blurred image of each scale is calculated to obtain the Ret i nex output images of three scales, and the three output images are weighted averaged. The weights are set according to the degree of retention of the bridge structure details at each scale, which are 0.3, 0.4 and 0.3 respectively, to obtain the fused Ret i nex output image. The fused Ret i nex output image is nonlinearly mapped, and the sigmoid function f(x) = 1/(1+exp(-α(x-β))) is used for color restoration and dynamic range compression, where α controls the slope of the curve, β controls the midpoint position, and x is the pixel value of the fused Ret i nex output image. The optimal parameter value is determined through iterative optimization to obtain a contrast-enhanced bridge image and highlight the outline of the bridge structure. The contrast-enhanced bridge image is divided into 8×8 sub-blocks using the adaptive histogram equalization method, and the cumulative distribution function is calculated for each sub-block. The contrast limit threshold is set to 0.01 to prevent excessive enhancement of noise. The boundaries between blocks are processed using bilinear interpolation to eliminate uneven illumination and shadow interference, and obtain a bridge image after illumination normalization. The structural similarity index SSIM of the image before and after processing is calculated. If the SSIM value is lower than the preset threshold of 0.8, the above parameters are adjusted and reprocessed until the quality requirements are met. For the bridge image after visibility enhancement, assuming that the image size is 2048x1536 pixels, a Gaussian filter group of the multi-scale Ret i nex algorithm is constructed, and the three scale parameters are set to 102, 512 and 1536 pixels respectively. The input image is Gaussian blurred to obtain three blurred images of different scales. The logarithmic difference between the input image and the blurred image of each scale is calculated to obtain the Ret i nex output images of three scales. The three output images are weighted averaged, and the weights are set to 0.3, 0.4 and 0.3 to obtain the fused Ret i nex output image. The fused Ret i nex output image is nonlinearly mapped, and the sigmoid function is used for color restoration and dynamic range compression. The initial parameters are set to α＝5, β＝0.5. The gradient descent method is used for iterative optimization, and finally the parameter values of α＝4.8, β＝0.52 are obtained. The image is nonlinearly mapped using the optimized parameters to obtain the bridge image with contrast enhancement. Adaptive histogram equalization is used to divide the contrast-enhanced bridge image into 8×8 sub-blocks, each with a size of 256x192 pixels. The cumulative distribution function is calculated for each sub-block, and the contrast limit threshold is set to 0.01. Bilinear interpolation is used to process the boundaries between blocks to eliminate uneven illumination and shadow interference, and obtain the bridge image after illumination normalization. The structural similarity index of the images before and after processing is calculated. It is assumed that the SSIM value is 0.85, which is higher than the preset threshold of 0.8, indicating that the processing effect is good.

Step S103, for the bridge image after illumination normalization, a weighted least square method is used to fit the bridge surface model at multiple scales, scattered light interference caused by haze is eliminated through surface reconstruction, haze parameters are estimated using a degradation function, missing information of some bands is adaptively compensated, and a bridge image after scattering correction is obtained.

A processing request carrying illumination-normalized bridge image information is received, the processing request being issued by an image acquisition device; the bridge image is decomposed into multiple scales according to the processing request, and the original image is downsampled using a Gaussian pyramid to obtain images of multiple scales; for the multiple images of different scales, a bridge surface model is fitted using a weighted least squares method, an error function is constructed, and the error function is iteratively optimized and minimized by a gradient descent method to obtain a bridge surface model at each scale; the scattered light intensity distribution is calculated according to the bridge surface model, an initial transmittance is estimated using an atmospheric scattering model, and a fog-free image and transmittance are solved by alternate optimization to obtain a degradation function; based on the degradation function, a scatter correction is performed on the original image, and a Wiener filter is used to adaptively compensate for missing band information, and the Wiener filter parameters dynamically adjust the filter strength according to the signal-to-noise ratio estimation of the local image block to obtain a scatter-corrected bridge image; a peak signal-to-noise ratio between the original image and the scatter-corrected bridge image is calculated, and if the peak signal-to-noise ratio is less than a preset threshold, the adjustment parameters are returned to perform scatter correction processing again.

Specifically, the input bridge image after illumination normalization is decomposed into multiple scales, and the original image is downsampled using a Gaussian pyramid. The number of pyramid layers is dynamically determined according to the minimum side length of the image. When the minimum side length is greater than 64 pixels, downsampling is continued to generate multiple images of different scales. For each scale image, the weighted least squares method is used to fit the bridge surface model and construct the error function

E(x)＝Σw_i(f_i(x)-y_i)^2, where w_i is the weight set based on pixel gradient, f_i(x) is the fitting function, and y_i is the observed value. The gradient descent method is used to iteratively optimize and minimize the error function to obtain the bridge surface model at each scale. The obtained bridge surface model is used to calculate the scattered light intensity distribution. According to the atmospheric scattering model I(x)＝J(x)t(x)+A(1-t(x)), where I(x) is the observed image, J(x) is the fog-free image, t(x) is the transmittance, and A is the atmospheric light value, the initial transmittance is estimated by dark channel priori, and then J(x) and t(x) are alternately optimized to obtain the degradation function. Based on the degradation function, the original image is scattering corrected, and the Wiener filter is used to adaptively compensate for the missing band information. The filter parameters dynamically adjust the filter strength according to the signal-to-noise ratio estimation of the local image block, and finally the bridge image after scattering correction is obtained. The peak signal-to-noise ratio (PSNR) of the image before and after processing is calculated to determine whether the scatter correction effect reaches the preset threshold. If not, the adjustment parameters are returned for reprocessing. The input 2048x1536 pixel bridge image is decomposed at multiple scales and downsampled using the Gaussian pyramid to generate three scale images of 2048x1536, 1024x768 and 512x384. For each scale image, the bridge surface model is fitted using the weighted least squares method. The weight w_i is set based on the pixel gradient calculated by the Sobel operator. The larger the gradient value, the higher the weight. The gradient descent method is used for iterative optimization. The learning rate is set to 0.01 and the maximum number of iterations is 1000. The bridge surface model at each scale is obtained. Using the obtained surface model, the scattered light intensity distribution is calculated, and the atmospheric scattering model I(x)=J(x)t(x)+A(1-t(x)) is applied. The initial transmittance is estimated by the dark channel prior, and the average value of the original image corresponding to the first 0.1% of the brightest pixels in the dark channel image is taken as the atmospheric light value A. Alternate optimization was performed 10 times to solve J(x) and t(x) to obtain the degradation function. Based on the degradation function, the original image was scatter corrected and Wiener filtering was used to compensate for the missing band information. The filter window size was set to 8x8, and the filter strength was dynamically adjusted according to the signal-to-noise ratio estimation of the local image block. The lower the signal-to-noise ratio, the greater the filter strength. Finally, the bridge image after scatter correction was obtained, and the peak signal-to-noise ratio (PSNR) of the image before and after processing was calculated. If the PSNR value is lower than 30dB, the adjustment parameters are returned and reprocessed until the PSNR reaches more than 30dB or the number of iterations exceeds 5 times.

Step S104, for the bridge image after scattering correction, multi-scale features of the bridge structure are extracted through morphological top-hat operation, and the bridge surface texture is clustered using local self-similarity measurement to obtain a bridge structure image after scale normalization and texture segmentation.

A multi-scale morphological structural element set is constructed according to the bridge image after scattering correction, and the size of the structural element is dynamically set to a preset ratio according to the minimum side length of the image; a morphological opening operation is performed on the bridge image to obtain a multi-scale top-hat transformation result; feature fusion is performed on the multi-scale top-hat transformation result, and a weighted summation method based on information entropy is adopted to obtain a fused bridge structure feature image; on the fused bridge structure feature image, the similarity between each pixel and its neighboring pixels is calculated through a sliding window to construct a multi-scale similarity matrix; based on the multi-scale similarity matrix, the K-means clustering algorithm is used to classify the bridge surface texture, the number of clusters is set to a preset value, and the K-means++ algorithm is used to select the initial center point, and iterative optimization is performed until the change of the class center is less than the preset threshold or the maximum number of iterations is reached; the classification result is post-processed by morphological opening and closing operations, the segmentation boundary is smoothed and small isolated areas are eliminated, and the bridge structure image after scale normalization and texture segmentation is obtained; the contour consistency index before and after segmentation is calculated to evaluate the segmentation quality.

Specifically, for the bridge image after scattering correction, a multi-scale morphological structure element set is constructed. The size of the structure element is dynamically set to 1%, 2% and 3% according to the minimum side length of the image. The image is morphologically opened and then subtracted from the original image to obtain the multi-scale top-hat transformation result. The multi-scale top-hat transformation result is feature fused, and the weighted summation method based on information entropy is used to calculate the entropy value of each scale feature. The higher the entropy value, the greater the weight. The fused bridge structure feature image is obtained. On the fused feature image, the local self-similarity measure is calculated, and three local windows of 5×5, 7×7 and 9×9 are selected. The similarity between each pixel and its neighboring pixels is calculated through the sliding window to construct a multi-scale similarity matrix. Based on the multi-scale similarity matrix, the K-means clustering algorithm is used to classify the bridge surface texture. The number of clusters is set to 5. The initial center point is selected using the K-means++ algorithm. It is iteratively optimized until the change of the class center is less than the preset threshold or the maximum number of iterations is 100, and the bridge structure image after scale normalization and texture segmentation is obtained. The segmentation results were post-processed by morphological opening and closing operations, with a structure element size of 3×3, to smooth the segmentation boundaries and eliminate small isolated areas. The contour consistency index before and after segmentation was calculated to evaluate the segmentation quality. For the 2048x1536 pixel scatter-corrected bridge image, a multi-scale morphological structure element set was constructed, with structure element sizes of 15x15, 31x31, and 46x46 pixels, respectively. The image was morphologically opened and then subtracted from the original image to obtain the multi-scale top-hat transformation result. The transformation results were feature fused, and the entropy value of each scale feature was calculated. The entropy value of the 15x15 scale was 6.2, the 31x31 scale was 5.8, and the 46x46 scale was 5.3. After normalization, the weights were 0.36, 0.34, and 0.30, and the weighted summation was performed to obtain the fused bridge structure feature image. On the fused feature image, three local windows of 5x5, 7x7 and 9x9 were used to calculate the self-similarity measure, and the structural similarity index SSIM was used as the similarity index to construct a multi-scale similarity matrix. Based on the multi-scale similarity matrix, the K-means clustering algorithm was used to classify the bridge surface texture, with the number of clusters being 5, and the K-means++ algorithm was used to select the initial center point. After 20 iterative optimizations, the class center change was less than 0.001, reaching the convergence condition, and the texture segmentation result was obtained. The segmentation result was post-processed with a 3x3-sized structural element for morphological opening and closing operations to smooth the segmentation boundaries and eliminate isolated areas with an area of less than 50 square pixels. The contour consistency index before and after segmentation was calculated, and a consistency score of 0.92 was obtained, indicating that the segmentation quality was good.

Step S105, applying the improved Canny operator to the bridge structure image to detect edge contours, eliminating background interference through edge connection and closure, obtaining a complete bridge structure edge map, using Hough transform to detect multi-scale line features of the bridge structure, and screening dominant structural lines through a probabilistic voting mechanism to obtain a bridge skeleton structure map.

The improved Canny operator is applied to the bridge structure image. The improved Canny operator dynamically adjusts the threshold based on the mean and standard deviation of the local area of the image, eliminates noise through Gaussian filtering, calculates the image gradient amplitude and direction, performs non-maximum suppression, and obtains a preliminary edge detection result; edge connection and closure processing are performed based on the preliminary edge detection result, and the morphological expansion operation is used to connect the disconnected edges, and the connection distance threshold is set. The least squares method is used to perform curve fitting to fill the missing edges, and the small holes are filled by the regional growing method to obtain a complete bridge structure edge map; multi-scale Hough transform is applied to the bridge structure edge map, and the image size and expected bridge structure are calculated according to the image size and expected bridge structure. The accumulator step size and polar coordinate resolution are set according to the structural dimensions, line features at multiple scales are detected, and the detection results are stored in the Hough space composed of a three-dimensional array; according to the line segment information in the Hough space, a probabilistic voting mechanism is implemented to screen the dominant structural lines, the voting weight is proportional to the line segment length and the gradient amplitude, the voting threshold and the line segment length threshold are set, the DBSCAN clustering algorithm is used to cluster the line segments with votes higher than the threshold, the similar line segments are merged, and the bridge skeleton structure diagram is obtained; the principal component analysis is used to verify whether the main direction of the skeleton structure extracted from the bridge skeleton structure diagram is in line with expectations, the angle between the principal component direction and the preset bridge direction is calculated, and the accuracy of the extraction result is judged.

Specifically, the improved Canny operator is applied to the bridge structure image, and the threshold is dynamically adjusted based on the mean and standard deviation of the local area of the image. The noise is eliminated by Gaussian filtering, the image gradient amplitude and direction are calculated, and non-maximum suppression is performed to obtain the preliminary edge detection results. The preliminary edge detection results are processed by edge connection and closure. The morphological expansion operation is used to connect the disconnected edges, and the connection distance threshold is set. The least squares method is used to perform curve fitting to complete the missing edges. The small holes are filled by the regional growing method. The seed point selection is based on the consistency of the gradient direction. The growth condition considers the pixel intensity and direction differences to obtain a complete bridge structure edge map. Multi-scale Hough transform is applied to the bridge structure edge map. The accumulator step size and polar coordinate resolution are set according to the image size and the expected bridge structure size. Line features at multiple scales are detected, and the detection results are stored in the Hough space composed of a three-dimensional array. The ρ, θ parameters and the number of votes for each line segment are recorded. Based on the line segment information in Hough space, a probabilistic voting mechanism is implemented to screen the dominant structural lines. The voting weight is proportional to the line segment length and gradient amplitude. The voting threshold and line segment length threshold are set. The DBSCAN clustering algorithm is used to cluster the line segments with votes higher than the threshold, and the similar line segments are merged to finally obtain the bridge skeleton structure diagram. The principal component analysis is used to verify whether the main direction of the extracted skeleton structure is in line with expectations. The angle between the principal component direction and the preset bridge direction is calculated to determine the accuracy of the extraction result. The improved Canny operator is applied to the 2048x1536 pixel bridge structure image, and the mean and standard deviation are calculated using a local window of 11x11 pixels. The low threshold is set to μ-σ and the high threshold is set to μ+σ. A 5x5 Gaussian kernel is used for filtering, σ=1.4, and the gradient amplitude and direction are calculated. The non-maximum suppression window is 3x3. The edge connection uses an 8x8 dilation kernel, and the connection distance threshold is set to 20 pixels. The least squares method is used to fit a third-order polynomial curve to complete the missing edges. The region growing method uses edge points with a gradient direction difference of less than 15° as seeds, and the growth conditions are that the pixel intensity difference is less than 20 and the direction difference is less than 30°. The multi-scale Hough transform sets three scales, the accumulator step size is 1, 2, and 4 pixels, and the θ resolution is 0.5°, 1°, and 2°. The Hough space is stored in a three-dimensional array of 1024x180xN, where N is the number of detected line segments. In the probability voting, the weight w=L*G, L is the line segment length, and G is the average gradient amplitude. The voting threshold is set to 5% of the total number of votes, and the line segment length threshold is 10% of the image diagonal length. The DBSCAN clustering parameters ε=20 pixels, MinPts=3. The principal component analysis calculates the eigenvector to determine whether the angle between the first principal component direction and the preset bridge direction is less than 15°.

Step S106, according to the bridge skeleton structure diagram, the anisotropic area of the bridge surface is determined by structural tensor analysis, the high-order texture features of the anisotropic area are extracted using the gray level co-occurrence matrix, and a texture feature dictionary of bridge damage is constructed.

Image data is acquired according to the bridge skeleton structure diagram, the gradient of the image data is calculated using the Sobel operator, a 2x2 structure tensor matrix is constructed, and the structure tensor of the local area is obtained; the degree of anisotropy is determined by performing eigenvalue analysis on the structure tensor; if the degree of anisotropy is greater than a preset threshold, it is determined to be a significant anisotropic area, and an anisotropic area map of the bridge surface is obtained; for the anisotropic area map, a quadtree decomposition algorithm is used for multi-scale segmentation, and if the size of the sub-area is less than 32x32 pixels or the change in the degree of anisotropy within the area is less than 0.1, the segmentation is terminated; the gray-level co-occurrence matrix is calculated for the sub-area, high-order texture features such as energy, contrast, correlation, and entropy are extracted, and a sub-area feature vector is constructed; the K-means++ algorithm is used to initialize K-means clustering, the sub-area feature vector is clustered, and the cluster center is obtained; the cluster center is used as a representative feature to construct a texture feature dictionary of bridge damage.

Specifically, according to the bridge skeleton structure diagram, the Sobel operator is used to calculate the image gradient, a 2x2 structure tensor matrix is constructed, the structure tensor of the local area is calculated, the degree of anisotropy is determined by eigenvalue analysis, and a threshold is set to screen out significant anisotropic areas to obtain the anisotropic area map of the bridge surface. The anisotropic area map is segmented at multiple scales, and the quadtree decomposition algorithm is used to adaptively divide the sub-areas according to the area size and the degree of anisotropy. The termination condition is that the sub-area size is less than 32x32 pixels or the anisotropy degree change in the area is less than 0.1, ensuring that the structural features of different scales are captured. For the divided sub-areas, the grayscale co-occurrence matrix is calculated, the distance is set to 1-5 pixels, the direction is 0°, 45°, 90° and 135°, and the grayscale level is 32. High-order texture features such as energy, contrast, correlation and entropy are extracted to construct the feature vector of each sub-area. The K-means clustering algorithm was used to cluster the feature vectors of all sub-regions. The K-means++ algorithm was used for initialization. The number of clusters was set to 10. The iterative optimization was performed until the intra-class distance was minimized or the maximum number of iterations reached 100. The cluster center was used as the representative feature to construct the texture feature dictionary of bridge damage. The constructed feature dictionary was optimized by principal component analysis, retaining 95% of the variance information, reducing redundancy and improving the representation ability of the dictionary, and the final bridge damage texture feature dictionary was obtained. For the 2048x1536 pixel bridge skeleton structure image, the 3x3Sobel operator was used to calculate the x and y direction gradients and construct a 2x2 structure tensor matrix. The local structure tensor was calculated in a sliding window of 16x16 pixels, and the anisotropy was determined by eigenvalue analysis. The threshold was set to 0.7 to screen out the significant anisotropic areas. The quadtree decomposition algorithm was used for multi-scale segmentation. The initial block size was 512x512 pixels, and the recursive partitioning was performed until the sub-region was smaller than 32x32 pixels or the anisotropy change was less than 0.1. The grayscale co-occurrence matrix is calculated for each sub-region, and the distance is set to 1-5 pixels, the direction is 0°, 45°, 90°, 135°, and 32 levels of grayscale quantization are used to extract 20 high-order texture features such as energy, contrast, correlation, and entropy to form a feature vector. The K-means++ algorithm is used to initialize 10 cluster centers, and K-means clustering is performed on all sub-region feature vectors. The maximum number of iterations is set to 100, and the convergence threshold is 1e-4. The 10 cluster centers obtained are used as representative features to construct the initial texture feature dictionary. The principal component analysis is applied to optimize the feature dictionary, calculate the feature covariance matrix, solve the eigenvalues and eigenvectors, and select the first N principal components with a cumulative variance contribution rate of 95%, usually N is 6-8. The original 20-dimensional features are projected into the N-dimensional principal component space to obtain the final bridge damage texture feature dictionary.

Step S107, applying a morphological skeleton extraction algorithm to the bridge structure edge map to obtain the skeleton line of the bridge connection part, and judging whether there is a fracture in the bridge structure through the connectivity analysis of the skeleton line. If there is a fracture, the damage location is determined according to the edge contour of the fracture area, and the damage size is determined according to the fracture length.

The Zhang-Suen thinning algorithm is applied to the bridge structure edge map for morphological skeleton extraction to obtain the initial skeleton line image of the bridge connection part; post-processing is performed based on the initial skeleton line image, and pixel connectivity analysis is used to remove branches and isolated points whose lengths are less than a preset threshold, and morphological opening and closing operations are performed through predefined structural elements to obtain an optimized bridge skeleton line image; connectivity analysis is performed on the optimized skeleton line image, and an 8-neighborhood connected area labeling algorithm is used to obtain the number of connected areas and the number of pixels in each area; if the distance between the connected areas exceeds the preset threshold, it is determined that there is a fracture, and the fracture area is located and measured, and the coordinates of the damage position are determined by the Moore neighborhood boundary tracking algorithm, and the shortest path length between the two end points of the skeleton line fracture is calculated as the damage size; according to the fracture position and damage size, a comparison is performed with the preset bridge structure model to determine whether the fracture position and damage size are within a reasonable range. If they exceed the preset range, the parameters are adjusted and reprocessed.

Specifically, the Zhang-Suen thinning algorithm is applied to the edge map of the bridge structure for morphological skeleton extraction. The pixel change number is set to be less than 10 or reach 100 iterations as the termination condition to obtain the initial skeleton line image of the bridge connection. The initial skeleton line image is post-processed, and branches and isolated points with a length of less than 20 pixels are removed based on pixel connectivity analysis. The skeleton line is smoothed by morphological opening and closing operations of 3x3 circular structural elements to obtain the optimized bridge skeleton line image. The optimized skeleton line image is subjected to connectivity analysis, and the 8-neighborhood connected region labeling algorithm is used to count the number of connected regions and the number of pixels in each region. If the distance between connected regions exceeds 50 pixels, it is judged that there is a fracture. If there is a fracture, the fracture area is located and measured, and the coordinates of the damage position are determined by the Moore neighborhood boundary tracking algorithm. The shortest path length between the two end points of the skeleton line fracture is calculated as the damage size. By comparing with the preset bridge structure model, it is verified whether the detected fracture position and size are within a reasonable range. If it exceeds the preset range, the aforementioned parameters are adjusted and reprocessed. The key intermediate results in the processing process are visualized, including the initial skeleton line, the optimized skeleton line, the connectivity analysis results, and the fracture location results. The Zhang-Suen thinning algorithm is applied to the 2048x1536 pixel bridge structure edge map. After 98 iterations, the pixel change number is less than 10, and the initial skeleton line image is obtained. The initial skeleton line is post-processed to remove 175 branches with a length of less than 20 pixels and 23 isolated points. The 3x3 circular structure element is used to perform 3 opening operations and 2 closing operations to obtain the optimized bridge skeleton line image. The 8-neighborhood connected region marking algorithm is used to analyze connectivity, and 7 connected regions are marked. The maximum number of pixels in the region is 12453 and the minimum number of pixels in the region is 89. It is found that the distances between the two connected regions are 62 pixels and 78 pixels, respectively, which exceeds the 50-pixel threshold, and it is determined that there is a fracture. The Moore neighborhood boundary tracking algorithm is used to determine the coordinates of the damage location. The first fracture is located at (1024,768) and the second is located at (1536,1152). The A* algorithm was used to calculate the shortest path length between the two end points of the skeleton line fracture, and the damage sizes were 85 pixels and 103 pixels respectively. The detection results were compared with the preset bridge structure model. The first fracture position deviation was 3.2%, the size deviation was 2.8%, and the second fracture position deviation was 4.1%, and the size deviation was 3.5%, all within the reasonable range of 5%. Visual results were generated, including the initial skeleton line graph, the optimized skeleton line graph, the connected area labeling graph, and the fracture location graph.

Step S108, comprehensively consider the structural integrity, surface texture characteristics and quantitative damage indicators of the bridge, use a support vector machine regression model to evaluate the health status of the bridge, obtain the predicted value of the damage degree of each part of the bridge, map the predicted value to the health status level, generate a bridge health assessment report, and intuitively present the bridge damage detection results.

The structural integrity, surface texture characteristics and quantitative damage indicators of the bridge are obtained; based on the structural integrity, surface texture characteristics and quantitative damage indicators, the three types of feature vectors are spliced by the feature cascade method to obtain the feature vector for bridge health status assessment; a sample data set is obtained from historical data, and a support vector machine regression model is constructed based on the sample data set; the kernel function parameters and penalty factors of the support vector machine regression model are determined by the grid search method; the feature vectors of various parts of the bridge are predicted by the support vector machine regression model to obtain the predicted value of the degree of damage; the measured degree of damage value is obtained through field testing; the root mean square error is calculated based on the measured degree of damage value and the predicted value of the degree of damage; if the root mean square error exceeds the preset threshold, the parameters of the support vector machine regression model are returned to adjust.

Specifically, the structural integrity, surface texture characteristics and quantitative damage indicators of the bridge are fused, and the three types of feature vectors are spliced using the feature cascade method. The weights are determined by data-driven correlation analysis, and then the principal component analysis method is used to reduce the dimension, retaining the principal components with a cumulative contribution rate of 95%, and constructing the feature vector for bridge health status assessment. The support vector machine regression model is constructed based on historical data, and the radial basis kernel function is selected. The kernel function parameters and penalty factors are optimized by the grid search method. The search range of parameter C is from 2^-5 to 2^15, and gamma is from 2^-15 to 2^3. The model performance is evaluated by 10-fold cross validation. Using the trained support vector machine regression model, the feature vectors of various parts of the bridge are predicted to obtain the predicted value of the degree of damage, and the delta method based on the prediction variance is used to calculate the 95% confidence interval of the prediction result. The health status classification standard is set, and the predicted value and its confidence interval are mapped to the corresponding health status level. The bridge status is divided into five levels: excellent 0-0.2, good 0.2-0.4, general 0.4-0.6, poor 0.6-0.8 and dangerous 0.8-1.0. An assessment report containing damage degree, health level and confidence level is generated, and the damage detection results of various parts of the bridge are intuitively presented in the form of a heat map. The accuracy of the assessment report is verified through field testing, and the root mean square error between the predicted value and the measured value is calculated. If the error exceeds the preset threshold, the model parameters are adjusted to retrain and predict. The structural integrity, surface texture characteristics and quantitative damage indicators of the bridge are fused, and the three types of feature vectors are spliced using the feature cascade method to obtain a 300-dimensional original feature vector. Through correlation analysis, weights of 0.4, 0.3 and 0.3 are assigned to the structural integrity, surface texture and damage indicators respectively. The principal component analysis method is used for dimensionality reduction, and the first 20 principal components are retained, with a cumulative contribution rate of 96.5%. The support vector machine regression model was constructed, and the RBF kernel function was selected. The parameters were optimized in the range of C = [2^-5, 2^-3, ..., 2^15] and gamma = [2^-15, 2^-13, ..., 2^3] by grid search method. The optimal parameters C = 2^7 and gamma = 2^-5 were finally determined by 10-fold cross validation. The optimized model was used to predict 1000 bridge parts, and the damage degree prediction value was obtained. The 95% confidence interval was calculated. The average prediction value was 0.45, and the confidence interval width was

±0.08. The prediction results are mapped to health status levels, of which 150 parts are excellent 0-0.2, 30 are good 0.2-0.4, 350 are average 0.4-0.6, 150 are poor 0.6-0.8, and 50 are dangerous 0.8-1.0. A heat map visualization evaluation report is generated to intuitively display the damage of each part. Through field testing, 100 randomly selected parts are verified, and the calculated root mean square error is 0.062, which is lower than the preset threshold of 0.1, verifying that the prediction accuracy of the model meets the requirements.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. A bridge detection image recognition method based on artificial intelligence, characterized in that the method comprises: Obtain the original bridge image taken in haze weather, and judge whether the visibility meets the requirements of bridge recognition and detection based on the image contrast and grayscale histogram distribution. If not, use the dark channel prior theory to estimate the global atmospheric illumination value, and combine the guided filtering algorithm for adaptive defogging to obtain the bridge image with enhanced visibility. For the bridge image after visibility enhancement, the multi-scale Retinex algorithm is used to enhance the contrast and highlight the outline of the bridge structure. The adaptive histogram equalization method is used to balance the image brightness distribution, eliminate uneven illumination and shadow interference, and obtain the bridge image after illumination normalization. For the bridge image after illumination normalization, the weighted least squares method is used to fit the bridge surface model at multiple scales. The scattered light interference caused by haze is eliminated through surface reconstruction. The haze parameters are estimated using the degradation function. The missing information of some bands is adaptively compensated to obtain the bridge image after scattering correction. For the bridge image after scattering correction, the multi-scale features of the bridge structure are extracted through morphological top-hat operation, and the bridge surface texture is clustered using local self-similarity measurement to obtain the bridge structure image after scale normalization and texture segmentation. The improved Canny operator is applied to the bridge structure image to detect the edge contour, and the background interference is eliminated by edge connection and closure to obtain a complete bridge structure edge map. The multi-scale line features of the bridge structure are detected by Hough transform, and the dominant structural lines are screened by the probabilistic voting mechanism to obtain the bridge skeleton structure map. According to the bridge skeleton structure diagram, the anisotropic area of the bridge surface is determined through structural tensor analysis, and the high-order texture features of the anisotropic area are extracted using the gray-level co-occurrence matrix to construct a texture feature dictionary of bridge damage. The morphological skeleton extraction algorithm is applied to the edge map of the bridge structure to obtain the skeleton line of the bridge connection part. The connectivity analysis of the skeleton line is used to determine whether the bridge structure has a fracture. If there is a fracture, the damage location is determined according to the edge contour of the fracture area, and the damage size is determined according to the fracture length. The support vector machine regression model is used to evaluate the health status of the bridge based on the structural integrity, surface texture characteristics and quantitative damage indicators of the bridge. The predicted value of the damage degree of each part of the bridge is obtained, the predicted value is mapped to the health status level, and a bridge health assessment report is generated to intuitively present the bridge damage detection results. 2. The method according to claim 1, wherein the original image of the bridge taken in haze weather is obtained, and whether the visibility meets the bridge recognition and detection requirements is judged according to the image contrast and grayscale histogram distribution. If not, the global atmospheric illumination value is estimated by using the dark channel prior theory, and the guided filtering algorithm is combined with adaptive defogging to obtain the bridge image with enhanced visibility, comprising: Obtaining an original image of a bridge taken in haze weather, and calculating a local contrast of the original image of the bridge, wherein the local contrast is obtained using a local contrast calculation formula; Analyzing the grayscale histogram distribution according to the original image of the bridge, and calculating the mean and variance of the grayscale histogram distribution; According to the local contrast, mean and variance, determine whether the current image visibility meets the preset bridge recognition and detection threshold; If the visibility of the current image does not meet the preset bridge recognition detection requirement threshold, the original bridge image is processed using the dark channel prior theory to obtain a defogged bridge image; Performing contrast adjustment on the defogged bridge image, and obtaining a contrast-adjusted bridge image using a linear stretching method; The contrast-adjusted bridge image is subjected to adaptive histogram equalization processing, the contrast-adjusted bridge image is divided into a plurality of small blocks, each of the small blocks is subjected to histogram equalization, and the processing results are combined by bilinear interpolation to obtain a bridge image with enhanced visibility. 3. The method according to claim 1, wherein the bridge image after visibility enhancement is subjected to contrast enhancement by using a multi-scale Retinex algorithm to highlight the outline of the bridge structure, and the image brightness distribution is balanced by an adaptive histogram equalization method to eliminate uneven illumination and shadow interference, so as to obtain a bridge image after illumination normalization, comprising: According to the size of the bridge image, the Gaussian filter group parameters of the multi-scale Retinex algorithm are obtained. The parameters include three scale values, which are 5%, 25% and 75% of the number of pixels on the long side of the image. The Gaussian filter bank is used to perform Gaussian blur processing on the input image to obtain three blurred images of different scales; Calculate the logarithmic difference between the input image and the blurred image at each scale to obtain the Retinex output images at three scales; Performing weighted averaging on the Retinex output images of the three scales to obtain a fused Retinex output image; Performing nonlinear mapping on the fused Retinex output image, using a sigmoid function to perform color restoration and dynamic range compression, and obtaining a contrast-enhanced bridge image; Dividing the contrast-enhanced bridge image into a plurality of sub-blocks, calculating a cumulative distribution function for each sub-block, and obtaining a bridge image after illumination normalization; Calculating a structural similarity index between the illumination-normalized bridge image and the input image; If the structural similarity index is lower than a preset threshold, the Gaussian filter group parameters, sigmoid function parameters and cumulative distribution function parameters are adjusted and reprocessed until the structural similarity index meets the quality requirements. 4. The method according to claim 1, wherein the bridge image after illumination normalization is used to fit the bridge surface model at multiple scales by using the weighted least square method, the scattered light interference caused by haze is eliminated by surface reconstruction, the haze parameters are estimated by using the degradation function, and the missing information of some bands is adaptively compensated to obtain the bridge image after scattering correction, comprising: receiving a processing request carrying bridge image information after illumination normalization, wherein the processing request is issued by an image acquisition device; Perform multi-scale decomposition on the bridge image according to the processing request, and downsample the original image using a Gaussian pyramid to obtain multiple images of different scales; For the images of the plurality of different scales, a bridge surface model is fitted by using a weighted least square method, an error function is constructed, and the error function is iteratively optimized and minimized by using a gradient descent method to obtain a bridge surface model at each scale; The scattered light intensity distribution is calculated according to the bridge surface model, the initial transmittance is estimated using the atmospheric scattering model, and the fog-free image and transmittance are solved by alternating optimization to obtain a degradation function; Based on the degradation function, the original image is subjected to scattering correction, and the missing band information is adaptively compensated by using Wiener filtering, wherein the Wiener filter parameter dynamically adjusts the filter strength according to the signal-to-noise ratio estimation of the local image block to obtain a bridge image after scattering correction; The peak signal-to-noise ratio of the original image and the bridge image after scatter correction is calculated. If the peak signal-to-noise ratio is less than a preset threshold, the adjustment parameters are returned to perform scatter correction again. 5. The method according to claim 1, wherein the multi-scale features of the bridge structure are extracted by morphological top-hat operation for the scatter-corrected bridge image, the bridge surface texture is clustered by local self-similarity metric, and the bridge structure image after scale normalization and texture segmentation is obtained, comprising: Constructing a set of multi-scale morphological structure elements according to the scatter-corrected bridge image, wherein the size of the structure elements is dynamically set to a preset ratio according to the minimum side length of the image; Performing a morphological opening operation on the bridge image to obtain a multi-scale top-hat transformation result; Performing feature fusion on the multi-scale top-hat transformation results, using a weighted summation method based on information entropy to obtain a fused bridge structure feature image; On the fused bridge structure feature image, the similarity between each pixel and its neighboring pixels is calculated through a sliding window to construct a multi-scale similarity matrix; Based on the multi-scale similarity matrix, the K-means clustering algorithm is used to classify the bridge surface texture, the number of clusters is set to a preset value, the K-means++ algorithm is used to select the initial center point, and iterative optimization is performed until the change of the class center is less than a preset threshold or the maximum number of iterations is reached; Performing morphological opening and closing operation post-processing on the classification results, smoothing the segmentation boundaries and eliminating small isolated areas, and obtaining a bridge structure image after scale normalization and texture segmentation; Calculate the contour consistency index before and after segmentation to evaluate the segmentation quality. 6. The method according to claim 1, wherein the improved Canny operator is applied to the bridge structure image to detect edge contours, background interference is eliminated by edge connection and closure to obtain a complete bridge structure edge map, the multi-scale line features of the bridge structure are detected by Hough transform, and the dominant structural lines are screened by a probabilistic voting mechanism to obtain a bridge skeleton structure map, including: Applying the improved Canny operator to the bridge structure image, the improved Canny operator dynamically adjusts the threshold based on the mean and standard deviation of the local area of the image, eliminates noise through Gaussian filtering, calculates the image gradient amplitude and direction, performs non-maximum suppression, and obtains preliminary edge detection results; Perform edge connection and closure processing according to the preliminary edge detection results, use morphological dilation operation to connect the disconnected edges, set the connection distance threshold, use the least squares method to perform curve fitting to complete the missing edges, and use the region growing method to fill small holes to obtain a complete bridge structure edge map; Applying a multi-scale Hough transform on the bridge structure edge map, setting an accumulator step size and a polar coordinate resolution according to an image size and an expected bridge structure size, detecting line features at multiple scales, and storing the detection results in a Hough space consisting of a three-dimensional array; According to the line segment information in the Hough space, a probabilistic voting mechanism is implemented to screen the dominant structural lines. The voting weight is proportional to the line segment length and the gradient amplitude. The voting threshold and the line segment length threshold are set. The line segments with votes higher than the threshold are clustered using the DBSCAN clustering algorithm, and similar line segments are merged to obtain the bridge skeleton structure diagram. The principal component analysis is used to verify whether the main direction of the skeleton structure extracted from the bridge skeleton structure diagram is consistent with expectations, and the angle between the principal component direction and the preset bridge direction is calculated to determine the accuracy of the extraction result. 7. The method according to claim 1, wherein the anisotropic region of the bridge surface is determined by structural tensor analysis based on the bridge skeleton structure diagram, and the high-order texture features of the anisotropic region are extracted using the gray level co-occurrence matrix to construct a texture feature dictionary of bridge damage, comprising: Acquire image data according to the bridge skeleton structure diagram, calculate the gradient of the image data using the Sobel operator, construct a 2x2 structure tensor matrix, and obtain the structure tensor of the local area; determining the degree of anisotropy by performing eigenvalue analysis on the structural tensor; If the anisotropy degree is greater than a preset threshold, it is determined to be a significant anisotropic region, and an anisotropic region map of the bridge surface is obtained; For the anisotropic region map, a quadtree decomposition algorithm is used to perform multi-scale segmentation. If the size of the sub-region is less than 32x32 pixels or the anisotropy degree change in the region is less than 0.1, the segmentation is terminated; Calculating the gray level co-occurrence matrix for the sub-region, extracting high-order texture features such as energy, contrast, correlation and entropy, and constructing a sub-region feature vector; Initialize K-means clustering using K-means++ algorithm, cluster the sub-region feature vectors, and obtain cluster centers; The cluster centers are used as representative features to construct a texture feature dictionary of bridge damage. 8. The method according to claim 1, wherein the morphological skeleton extraction algorithm is applied to the bridge structure edge map to obtain the skeleton line of the bridge connection part, and the connectivity analysis of the skeleton line is used to determine whether the bridge structure has a fracture. If a fracture exists, the damage position is determined according to the edge contour of the fracture area, and the damage size is determined according to the fracture length, including: The Zhang-Suen thinning algorithm is applied to the edge map of the bridge structure to extract the morphological skeleton and obtain the initial skeleton line image of the bridge connection part. Post-processing is performed according to the initial skeleton line image, branches and isolated points whose lengths are less than a preset threshold are removed by pixel connectivity analysis, and morphological opening and closing operations are performed by predefined structural elements to obtain an optimized bridge skeleton line image; Performing connectivity analysis on the optimized skeleton line image, using an 8-neighborhood connected region labeling algorithm to obtain the number of connected regions and the number of pixels in each region; If the distance between the connected areas exceeds a preset threshold, it is determined that there is a fracture, the fracture area is located and measured, the coordinates of the damage position are determined by the Moore neighborhood boundary tracking algorithm, and the shortest path length between the two end points of the skeleton line fracture is calculated as the damage size; The fracture position and damage size are compared with a preset bridge structure model to determine whether the fracture position and damage size are within a reasonable range. If they exceed the preset range, the parameters are adjusted and reprocessed. 9. The method according to claim 1, wherein the bridge structural integrity, surface texture characteristics and quantitative damage indicators are integrated to evaluate the health status of the bridge using a support vector machine regression model, obtain the predicted value of the damage degree of each part of the bridge, map the predicted value to the health status level, generate a bridge health assessment report, and intuitively present the bridge damage detection results, including: Obtain bridge structural integrity, surface texture characteristics and quantitative damage indicators; According to the structural integrity, surface texture characteristics and damage quantitative index, the three types of feature vectors are spliced by using the feature cascade method to obtain the feature vector for bridge health status assessment; Acquire a sample data set from historical data, and construct a support vector machine regression model according to the sample data set; Determine the kernel function parameters and penalty factor of the support vector machine regression model by grid search method; The support vector machine regression model is used to predict the characteristic vectors of various parts of the bridge to obtain a predicted value of the damage degree; Obtain the actual damage degree value through on-site testing; Calculating a root mean square error based on the measured damage degree value and the damage degree prediction value; If the root mean square error exceeds a preset threshold, the parameters of the support vector machine regression model are returned for adjustment. 10. A bridge detection image recognition system based on artificial intelligence, characterized in that the system comprises: Image preprocessing module, used to perform defogging and contrast enhancement on the original bridge image under haze weather; Surface model fitting and scattering correction module, used to fit the bridge surface model, eliminate scattered light interference, and improve image clarity; The structural feature extraction module is used to extract the edge contour and texture features of the bridge structure and obtain the bridge skeleton structure diagram; The damage identification and health assessment module is used to analyze the structural integrity of the bridge, evaluate the health status of the bridge, and generate damage detection result reports.