Method for detecting the compaction degree of prestressed duct grouting based on visual detection

Abstract

The application discloses a prestressed hole grouting compactness detection method based on visual detection and belongs to the technical field of prestressed hole grouting quality detection. The method comprises the following steps: acquiring an initial panoramic image sequence of the inner wall of a prestressed hole and performing uneven illumination correction; extracting local texture feature points from the uniform illumination panoramic image sequence, identifying a grout filling area and a gap area based on distribution density; calculating a grout interface transition zone width value according to a boundary shape, marking a to-be-rechecked section when the width value exceeds a preset width threshold value, and acquiring multi-angle local images; performing gradient direction histogram feature extraction on the multi-angle local images to obtain a gradient direction feature vector; determining a grout void boundary line according to a similarity matching result between the gradient direction feature vectors; and obtaining a prestressed hole grouting compactness detection value according to a ratio of an area of a region enclosed by the grout void boundary line to a hole cross-sectional area.

Description

Technical Field

[0001] This invention relates to the field of prestressed duct grouting quality inspection technology, specifically a method for detecting the compactness of prestressed duct grouting based on visual inspection. Background Technology

[0002] The grouting density of prestressed ducts directly affects the structural safety and durability of prestressed components. Existing grouting density testing technologies include non-destructive testing methods such as impact echo, ultrasonic transmission, and ground-penetrating radar. The impact echo method identifies defects by analyzing the reflected signals of elastic waves propagating within the duct, while the ultrasonic transmission method relies on the changes in wave velocity and attenuation characteristics of sound waves passing through the grout to locate voids. Infrared thermal imaging technology uses the difference in thermal conduction between the grout and the void region to detect defect boundaries, and X-ray inspection obtains images of the density distribution inside the duct through transmission imaging.

[0003] Existing technologies have significant shortcomings in terms of accuracy and boundary location capabilities for identifying grouting defects. Impact echo and ultrasonic methods lack sensitivity to micro-void regions; when the void size is smaller than the sensor wavelength, the reflected signal is difficult to separate from background noise, leading to missed detection of local voids. Infrared thermal imaging is severely affected by ambient temperature gradients; the non-uniformity of the temperature field in the concrete surrounding the duct results in the absence of a clearly identifiable grayscale transition interface between the grout and voids in the thermal image. While X-ray imaging can directly image, the two-dimensional projection image loses circumferential depth information, making it impossible to distinguish the overlap of voids on the front and back sides of the duct, leading to misjudgments of void surface locations. Panoramic camera images of the duct's inner wall provide complete circumferential texture information, but existing image analysis methods only use a global grayscale threshold to divide the grout and void regions, ignoring the unevenly wide transition zone between the grout and voids. Single threshold segmentation forces the transition zone to be classified as either grout or void, resulting in a systematic overestimation or underestimation of the void area near the boundary. Meanwhile, for suspected voided sections with insignificant grayscale differences, the global thresholding method lacks an independent multi-angle feature verification mechanism, making it impossible to accurately pinpoint the true circumferential boundary of the slurry void. Therefore, how to quantify the width of the slurry void transition zone and trigger a targeted verification process based on this width, and how to use multi-angle texture feature matching to determine the void boundary line in grayscale-blurred areas, are urgent problems to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a visual inspection-based method for detecting the compactness of prestressed duct grouting. This method extracts local texture feature points from the inner wall of the duct and identifies the grout-filled and void regions based on their distribution density. Then, it calculates the width of the grout interface transition zone based on the boundary morphology of the two regions, quantifying the transition range of the non-ideal abrupt interface between the grout and voids. When the width of the grout interface transition zone exceeds a preset width threshold, the section to be checked is marked. Multi-angle local images of the section to be checked are acquired, and gradient direction histogram feature vectors are extracted. The similarity matching results between feature vectors at different axial positions are used to determine the grout void boundary line. Finally, the grouting compactness detection value is calculated based on the area enclosed by the void boundary line.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a visual detection-based method for detecting the grouting density of prestressed ducts, the method comprising: acquiring an initial panoramic image sequence of the inner wall of the prestressed duct; performing illumination unevenness correction on the initial panoramic image sequence to obtain a uniform illumination panoramic image sequence; extracting multiple local texture feature points of the inner wall of the duct based on the uniform illumination panoramic image sequence; identifying multiple grout-filled regions and multiple void regions of the inner wall of the duct based on the distribution density of the multiple local texture feature points in the uniform illumination panoramic image sequence; and identifying multiple grout-filled regions and multiple void regions of the inner wall based on the edge distribution between the multiple grout-filled regions and the multiple void regions. The grout interface transition zone width of the duct inner wall is calculated. When the grout interface transition zone width exceeds a preset width threshold, the corresponding duct segment is marked as a segment to be verified, and multi-angle local images of the segment to be verified are obtained in the initial panoramic image sequence. Gradient direction histogram features are extracted from the multi-angle local images to obtain multiple gradient direction feature vectors. Based on the similarity matching results between the multiple gradient direction feature vectors, the grout void boundary line of the segment to be verified is determined. The grout compaction test value is obtained based on the ratio of the area enclosed by the grout void boundary line to the duct cross-sectional area of ​​the segment to be verified. This solution comprehensively collects and intelligently analyzes the condition of the inner wall of the duct using panoramic visual inspection. Through illumination correction, texture feature extraction, region identification, and interface transition zone width calculation, it can more accurately reflect the transition state of slurry filling and void distribution. When anomalies occur in the interface transition zone, multi-angle verification is further initiated. Gradient direction features and similarity matching are used to accurately locate the void boundary. Finally, the density is quantified by area ratio, which effectively improves the accuracy and reliability of the detection results.

[0006] As a preferred embodiment of the present invention, the step of obtaining an initial panoramic image sequence of the inner wall of the prestressed duct includes: pushing a panoramic camera device into the duct at a constant speed along the axial direction of the prestressed duct; controlling the panoramic camera device to rotate continuously around the axial direction at a preset angular velocity during the pushing process, acquiring multiple circumferential images; stitching each circumferential image according to the order of acquisition time to form a single-circumferential panoramic ring image; arranging the multiple panoramic ring images sequentially according to the pushing direction, and performing pixel-level registration and fusion of the overlapping areas of adjacent panoramic ring images to obtain the initial panoramic image sequence. This preferred embodiment, through the coordinated control of uniform speed advancement and constant angular velocity rotation, combined with the registration and fusion of overlapping areas, can obtain a continuous, complete, and geometrically consistent panoramic image sequence of the inner wall within a narrow duct, providing high-quality basic image data for subsequent analysis and processing.

[0007] As a preferred embodiment of the present invention, the step of performing illumination unevenness correction on the initial panoramic image sequence to obtain a uniformly illuminated panoramic image sequence includes: dividing each frame of the initial panoramic image sequence into multiple image sub-blocks of equal size; calculating the average gray value within each image sub-block, and constructing an illumination background surface based on the average gray value of all image sub-blocks; performing smooth interpolation processing on the illumination background surface to obtain an estimated background illumination value corresponding to each pixel position; dividing the original gray value of each pixel in each frame of the panoramic image by the estimated background illumination value corresponding to that pixel position to obtain a corrected gray value for each pixel; rearranging all the corrected gray values ​​according to the original pixel positions to generate a corrected image for each frame of the panoramic image, and arranging all the corrected images in the original sequence order to obtain the uniformly illuminated panoramic image sequence. This preferred embodiment can adaptively estimate the background distribution of uneven illumination within the aperture, and effectively eliminate the influence of illumination gradients and local shadows through pixel-by-pixel correction, making the texture of each region in the corrected image sequence consistent, facilitating stable and reliable feature extraction.

[0008] As a preferred embodiment of the present invention, the step of extracting multiple local texture feature points of the inner wall of the channel based on the uniformly illuminated panoramic image sequence includes: performing multi-scale Gaussian filtering on each frame of the uniformly illuminated panoramic image sequence to construct an image Gaussian pyramid for that frame; calculating the Hessian matrix response value of each pixel in each layer of the image Gaussian pyramid; marking pixels in the same layer whose Hessian matrix response value is greater than a preset response threshold as candidate feature points of that layer; performing spatial position matching on the candidate feature points between different layers, and grouping multiple candidate feature points whose spatial position deviation is less than a preset deviation threshold into the same group; and taking the candidate feature point with the largest Hessian matrix response value in the same group as the local texture feature point of that group, thereby obtaining multiple local texture feature points. This preferred embodiment, through multi-scale spatial analysis and Hessian matrix response filtering, can robustly extract robust texture feature points that repeatedly appear on the inner wall of the channel at different scales, eliminate noise and weak texture interference, and the extraction result has scale invariance and good repeatability.

[0009] As a preferred embodiment of the present invention, the step of identifying multiple slurry-filled regions and multiple void regions on the inner wall of a channel based on the distribution density of multiple local texture feature points in the uniformly illuminated panoramic image sequence includes: dividing each frame of the uniformly illuminated panoramic image sequence into multiple statistical grids of equal area; counting the number of local texture feature points contained in each statistical grid to obtain a feature point count for each statistical grid; marking statistical grids with feature point counts greater than a preset high-density threshold as slurry-filled grids, and marking statistical grids with feature point counts less than a preset low-density threshold as void grids; merging all slurry-filled grids into four-connected components to obtain multiple slurry-filled regions; and merging all void grids into four-connected components to obtain multiple void regions. This preferred embodiment transforms feature point distribution density into a regional attribute criterion, enabling efficient automatic segmentation of slurry-covered and voided areas, avoiding misjudgments caused by relying on a single grayscale threshold, and improving the sensitivity and robustness of region recognition to texture differences.

[0010] As a preferred embodiment of the present invention, the step of calculating the width of the slurry interface transition zone on the inner wall of the duct based on the boundary morphology between the plurality of slurry-filled regions and the plurality of void regions includes: extracting the outer boundary contour line of the slurry body of each slurry-filled region and extracting the outer boundary contour line of the voids of each void region; using the shortest distance between the outer boundary contour line of the slurry body and the outer boundary contour line of the voids as the initial boundary spacing between the slurry-filled region and the void region; filtering out all void regions adjacent to each slurry-filled region based on the initial boundary spacing; selecting the void region with the smallest initial boundary spacing from all void regions adjacent to each slurry-filled region as the paired void region of the slurry-filled region; and using the initial boundary spacing between the slurry-filled region and the paired void region as the width value of the slurry interface transition zone at that location. This preferred embodiment, through the calculation of the shortest contour distance and the pairing of the smallest spacing, can accurately measure the width of the transition zone between the slurry and the voids, objectively reflect the change in the density of the slurry edge, and provide a quantitative basis for judging whether there are potential weak points in the voids.

[0011] As a preferred embodiment of the present invention, the step of marking the corresponding duct segment as a segment to be verified when the width value of the slurry interface transition zone exceeds a preset width threshold, and obtaining multi-angle local images of the segment to be verified in the initial panoramic image sequence, includes: obtaining the image position coordinates of the slurry filling area corresponding to the width value of the slurry interface transition zone in the uniformly illuminated panoramic image sequence; determining the axial position of the duct corresponding to the width value of the slurry interface transition zone based on the image position coordinates, and marking the duct segment between a preset distance before and a preset distance after the axial position of the duct as the segment to be verified; extracting multiple frames of panoramic images covering the segment to be verified from the initial panoramic image sequence; cutting each frame of the panoramic image according to the circumferential angle to obtain multiple circumferential angle sub-images; arranging the circumferential angle sub-images with the same circumferential angle in axial order according to the duct axis to form the multi-angle local image corresponding to each circumferential angle. This preferred approach utilizes the initial image sequence for multi-angle reconstruction, which can obtain unfolded images of suspected gapped sections from different circumferential perspectives. This facilitates independent observation and confirmation of the existence and extent of gapping from multiple directions, reducing the risk of missed detections and false detections.

[0012] As a preferred embodiment of the present invention, the step of extracting gradient direction histogram features from the multi-angle local images to obtain multiple gradient direction feature vectors includes: calculating the gradient magnitude and gradient direction of each pixel in each frame of the multi-angle local images at each angle; dividing each frame of the sub-image into multiple cells of equal size; statistically analyzing the gradient direction distribution of all pixels in each cell to generate the gradient direction histogram of that cell; concatenating the gradient direction histograms of all cells in the same frame of the sub-image according to the cell position order to generate the sub-image feature vector of that frame of the sub-image; and extracting the gradient direction histograms of each pixel in the same circumferential angle index. The feature vector of a sub-image corresponding to a frame is used as the gradient direction feature vector of that frame. Multiple gradient direction feature vectors are obtained by arranging them in axial order. The gradient direction feature vectors of each frame under the same circumferential angle index have the same dimension, which is the total number of cells multiplied by 9. For the same multi-angle local image, the feature vector of each sub-image corresponding to the frame arranged in axial order in the multi-angle local image is used as the gradient direction feature vector of that frame. They are arranged in axial order to form the gradient direction feature vector sequence corresponding to the circumferential angle index. This operation is performed on all circumferential angle indices one by one to obtain multiple gradient direction feature vector sequences corresponding to each circumferential angle index.

[0013] As a preferred embodiment of the present invention, the step of determining the slurry voiding boundary line of the section to be verified based on the similarity matching results among the plurality of gradient direction feature vectors includes: selecting a first feature vector located at the beginning of the section to be verified and a second feature vector located at the end of the section to be verified from the plurality of gradient direction feature vectors; calculating a first cosine similarity between the first feature vector and each of the remaining gradient direction feature vectors, and calculating a second cosine similarity between the second feature vector and each of the remaining gradient direction feature vectors; weighting and summing the first cosine similarity and the second cosine similarity corresponding to each gradient direction feature vector to obtain a comprehensive similarity of the gradient direction feature vector; marking the circumferential angle positions corresponding to the gradient direction feature vectors with a comprehensive similarity lower than a preset similarity threshold as slurry voiding candidate boundaries; and connecting all the slurry voiding candidate boundaries in circumferential angle order to form the slurry voiding boundary line. This preferred scheme uses the combined similarity between the first and last reference vectors and the remaining angle vectors to determine the boundary. It can automatically find the start and end angles of the non-slurry-filled area in the circumferential direction, thereby accurately defining the complete outline of the void boundary. It has strong anti-interference ability and recognition accuracy.

[0014] As a preferred embodiment of the present invention, the step of obtaining the grouting density test value based on the ratio of the area enclosed by the grout void boundary line to the cross-sectional area of ​​the duct in the section to be verified includes: projecting the grout void boundary line onto the cross-section of the section to be verified to obtain the projected outline of the void area; calculating the first area value enclosed by the projected outline of the void area; obtaining the duct inner diameter value corresponding to the section to be verified in the initial panoramic image sequence, and calculating the second area value of the duct cross-section of the section to be verified based on the duct inner diameter value; dividing the first area value by the second area value to obtain the void area ratio of the section to be verified; and subtracting the void area ratio from the first area value to obtain the grouting density test value. This preferred embodiment normalizes the area enclosed by the void boundary line and the theoretical cross-sectional area of ​​the duct, outputting the final density test value in the form of a simple area ratio, which intuitively reflects the density of the grouting filling and facilitates project acceptance and quality assessment.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By extracting local texture feature points and identifying slurry-filled and void regions based on their distribution density, the rough texture of the slurry surface and the smooth inner walls of the void regions show significant differences in feature point density, allowing for the division of these two types of regions without relying on a single grayscale threshold. Furthermore, the outer boundary contours of each region are extracted, and the initial boundary distance between the slurry and paired void regions is calculated. This distance directly corresponds to the width of the transition zone at the slurry interface. Incorporating the width of the transition zone as an independent parameter into the analysis process allows for the quantitative representation of the width of the grayscale gradient region, rather than its forced classification into slurry or void categories. This eliminates the calculation bias in void area caused by fuzzy transition zone identification.

[0016] For pore sections where the transition zone width exceeds a preset threshold, these sections are marked as verification sections, and multi-angle local images covering these sections are acquired. Gradient direction histogram feature vectors are extracted frame-by-frame from each angle image. Gradient direction histogram features are insensitive to illumination changes and local grayscale shifts, and can stably describe the differences in texture structure between the slurry-filled and voided regions at different circumferential angles. By calculating the cosine similarity between the feature vectors at the beginning and end of the verification section and the feature vectors at each circumferential angle position, and performing weighted fusion, a comprehensive similarity is obtained. Circumferential angle positions with comprehensive similarity below the threshold are marked as candidate slurry voided boundaries and connected to form slurry voided boundary lines. This process transforms the problem of voided boundary discrimination within grayscale-blurred regions into a consistency matching problem of multi-angle texture features. Circumferential boundary localization no longer depends on the absolute value of pixel grayscale, avoiding boundary jitter and false contours caused by grayscale gradients in the transition zone, making the spatial morphology of the voided boundary lines closer to the actual slurry voided distribution. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of a visual inspection-based method for detecting the compactness of prestressed duct grouting. Figure 2 This is a flowchart of the process for obtaining the initial panoramic image sequence of the inner wall of the prestressed duct; Figure 3 This is a flowchart of the panoramic image illumination unevenness correction process; Figure 4 This is a flowchart of a method for extracting local texture feature points on the inner wall of a channel; Figure 5 This is a flowchart for calculating the width of the transition zone between the slurry interface and the pore wall. Figure 6 This is a flowchart of the grouting compaction detection process based on gradient direction feature similarity matching. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1This invention provides a vision-based method for detecting the grouting density of prestressed ducts, comprising: acquiring an initial panoramic image sequence of the inner wall of the prestressed duct; performing illumination unevenness correction on the initial panoramic image sequence to obtain a uniform illumination panoramic image sequence; extracting multiple local texture feature points of the inner wall of the duct based on the uniform illumination panoramic image sequence; identifying multiple grout-filled regions and multiple void regions of the inner wall of the duct based on the distribution density of the multiple local texture feature points in the uniform illumination panoramic image sequence; and calculating the grout density of the inner wall of the duct based on the boundary morphology between the multiple grout-filled regions and the multiple void regions. The width value of the interface transition zone; when the width value of the slurry interface transition zone exceeds a preset width threshold, the corresponding channel segment is marked as the segment to be verified, and multi-angle local images of the segment to be verified are obtained in the initial panoramic image sequence; gradient direction histogram features are extracted from the multi-angle local images to obtain multiple gradient direction feature vectors; based on the similarity matching results between the multiple gradient direction feature vectors, the slurry void boundary line of the segment to be verified is determined; based on the ratio of the area enclosed by the slurry void boundary line to the cross-sectional area of ​​the channel of the segment to be verified, the grouting density detection value of the channel is obtained.

[0021] Example 1:

[0022] In specific implementation, please refer to Figure 2 The process of obtaining the initial panoramic image sequence of the inner wall of the prestressed duct includes: The panoramic camera device is pushed into the prestressed duct at a constant speed along its axial direction. This pushing can be achieved using an electric push rod or an electric winch in conjunction with guide wheels. The constant pushing speed is pre-set based on the image acquisition frame rate of the panoramic camera device and the required axial overlap between two adjacent panoramic ring images, ensuring that speed fluctuations do not exceed a predetermined range during the pushing process. The panoramic camera device includes at least one optical imaging lens and an illumination source. The optical axis of the imaging lens is perpendicular to the axial direction of the prestressed duct to capture circumferential images of the inner wall of the prestressed duct.

[0023] During the pushing process, the panoramic camera device is controlled to rotate continuously around the axis of the prestressed duct at a preset angular velocity. The value of the preset angular velocity is coordinated with the constant speed, the single-frame exposure time of the panoramic camera device, and the frame rate, so that the number of circumferential image frames acquired by the panoramic camera device in one rotation can cover the entire circumference of the inner wall of the prestressed duct, and there is an overlapping area between adjacent circumferential image frames. The panoramic camera device is driven by a rotary motor, which is mounted on the moving bracket of the pushing mechanism. The output shaft of the rotary motor is fixedly connected to the panoramic camera device. The rotary motor rotates continuously at the preset angular velocity, while the panoramic camera device continuously acquires images at a fixed frame rate. Each trigger acquisition obtains one circumferential image, and the multiple circumferential image frames acquired in one rotation constitute a complete circumferential image.

[0024] Each circumferential image is stitched together in the order of acquisition time to form a single-circumferential panoramic image. For several frames of circumferential images constituting the same circle, feature points are first extracted from each frame. Based on the feature point matching relationship between adjacent frames, the image transformation matrix between adjacent frames is calculated. The image transformation matrix is ​​used to transform each frame to the same reference plane. Then, the pixel grayscale values ​​of the overlapping areas of adjacent frames are weighted and averaged to eliminate seams, forming a single-circumferential panoramic image covering the entire inner wall of the prestressed duct. In the weighted fusion, the weight of a pixel in the overlapping area is related to the distance of that pixel from the boundary of the overlapping area; the farther the distance from the boundary, the greater the weight of that frame.

[0025] The multi-ring panoramic images are arranged sequentially according to the pushing direction, and the overlapping areas of adjacent ring images are pixel-level registered and fused to obtain an initial panoramic image sequence. For adjacent ring images, multiple matching blocks are selected in the overlapping area, and the phase correlation spectrum between adjacent ring images is calculated. The axial displacement between adjacent ring images is determined by the peak position of the phase correlation spectrum. The adjacent ring images are translated and aligned according to the axial displacement, and the corresponding pixel gray values ​​in the overlapping area of ​​adjacent ring images are fused using a linear gradient weight. Pixels far from the boundary of the overlapping area have a higher fusion weight. All fused single-ring panoramic images are arranged sequentially according to the axial order of the prestressed duct to form an initial panoramic image sequence that is continuous in both the axial and circumferential directions.

[0026] Example 2:

[0027] In specific implementation, please refer to Figure 3 The process of performing illumination unevenness correction on the initial panoramic image sequence to obtain a uniformly illuminated panoramic image sequence includes: Each frame of the initial panoramic image sequence is divided into multiple equally sized image sub-blocks. The size of the image sub-blocks is set according to the resolution of the panoramic image and the actual texture scale of the inner walls of the apertures, ensuring that the grayscale variations within a single image sub-block primarily originate from uneven illumination rather than scene structure. For each frame of the panoramic image, the pixel matrix is ​​divided into non-overlapping rectangular regions according to the image sub-block size parameter; each rectangular region constitutes an image sub-block. When the number of rows and columns of pixels in the panoramic image is not divisible by the image sub-block size, the boundary pixels of the panoramic image are padded to integer multiples using a mirror expansion method. The expanded regions are not included in the generation of the final corrected image.

[0028] The average grayscale value within each image sub-block is calculated, and an illumination background surface is constructed based on the average grayscale values ​​of all image sub-blocks. For each frame of the panoramic image, the original grayscale values ​​of all pixels within each image sub-block are statistically analyzed, summed, and divided by the total number of pixels in the sub-block to obtain the average grayscale value corresponding to each sub-block. The center pixel coordinates of each sub-block are used as the spatial position coordinates of that sub-block. The x and y coordinates of the spatial position coordinates are used as two independent variables, and the average grayscale value corresponding to each sub-block is used as the dependent variable to form a set of discrete spatial sampling points. A two-dimensional polynomial surface fitting method is used to construct the illumination background surface from the discrete spatial sampling points. The degree of the two-dimensional polynomial is determined by the complexity of the illumination changes within the aperture. The polynomial used is quadratic. The calculation process of the two-dimensional quadratic polynomial fitting surface uses the least squares method to solve for the polynomial coefficients, minimizing the sum of squared residuals between the fitted values ​​and the actual average grayscale values ​​at all discrete spatial sampling points. The fitted illumination background surface continuously outputs the initial value of the background illumination estimate with the pixel coordinate position of the entire panoramic image as the independent variable.

[0029] The background illumination surface is smoothed and interpolated to obtain the background illumination estimate for each pixel location. The initial value of the background illumination estimate output from the fitted background illumination surface has a definite value only at the center pixel coordinates of the image sub-block. To obtain the background illumination estimate for each pixel location in the panoramic image, bicubic interpolation is performed on the initial value of the background illumination estimate obtained from the two-dimensional polynomial surface. Centered on each pixel location to be interpolated in the panoramic image, the interpolation kernel function is convolved with discrete points of the surrounding initial background illumination estimates to obtain the background illumination estimate for each pixel location in the panoramic image. During the interpolation process, symmetrical boundary extension is used for pixels at the boundaries of the panoramic image to obtain the additional data points required for interpolation.

[0030] The corrected grayscale value of each pixel in each frame of the panoramic image is obtained by dividing the original grayscale value of each pixel by the estimated background illumination value corresponding to that pixel location. The expression for calculating the corrected grayscale value is:

[0031] in, This represents the row and column indexes of pixels in a panoramic image. The value is an integer ranging from 1 to the total number of rows in the panoramic image. The value is an integer ranging from 1 to the total number of columns in the panoramic image. Indicating panoramic image The original grayscale value of the pixel at the location, which is directly read from the corresponding frame of the initial panoramic image sequence; Indicating panoramic image The estimated background illumination value corresponding to the pixel at the location is obtained after constructing the illumination-background surface and performing bicubic interpolation. Indicating panoramic image The corrected grayscale value is calculated from the pixel at the specified location. This corrected grayscale value is stored in floating-point format to ensure computational accuracy. When... When it approaches zero, The value is assigned to a preset minimum positive value to avoid abnormal division operations.

[0032] All corrected grayscale values ​​are rearranged according to their original pixel positions to generate a corrected image for each frame of the panoramic image. The calculated values ​​are then processed according to the row and column coordinate indices of the pixels within the frame. The values ​​are sequentially filled into a two-dimensional matrix of the same size as the original panoramic image. The matrix is ​​then normalized to normalize its numerical range, linearly mapping the corrected grayscale values ​​to a preset grayscale display range to generate the corrected image. The final grayscale values ​​of the corrected image pixels are obtained by rounding the floating-point corrected grayscale values ​​to integer values.

[0033] Arrange all corrected images in the original sequence order to obtain a uniformly illuminated panoramic image sequence. Keeping the temporal and spatial axial order of the corrected images in the initial panoramic image sequence unchanged, store each frame of the corrected image into the image sequence container in turn to generate a uniformly illuminated panoramic image sequence with the same number of frames as the initial panoramic image sequence and with corresponding frames in a one-to-one correspondence.

[0034] Example 3:

[0035] In specific implementation, please refer to Figure 4 The process of extracting multiple local texture feature points of the inner wall of the channel from a panoramic image sequence under uniform illumination includes: For each frame of a uniformly illuminated panoramic image sequence, multi-scale Gaussian filtering is performed to construct an image Gaussian pyramid for that frame. For a single frame of uniformly illuminated panoramic image, a set of scale-increasing two-dimensional Gaussian convolution kernels is used to perform multiple filtering processes, with each filtering smoothing producing one layer of the image Gaussian pyramid. The image Gaussian pyramid consists of multiple layers. The bottom layer is the original uniformly illuminated panoramic image. Each upper layer is obtained by convolving the next layer with a two-dimensional Gaussian convolution kernel, followed by interleaved downsampling of the convolution result, resulting in a progressively smaller image size. The standard deviation parameter of the two-dimensional Gaussian convolution kernel increases in a geometric progression, with the ratio of the standard deviations between adjacent layers being a fixed constant. The number of layers in the image Gaussian pyramid and the initial standard deviation of the Gaussian convolution kernels are set according to the spatial scale range of the texture features of the prestressed duct inner wall. The number of layers in the image Gaussian pyramid is set such that the spatial wavelength corresponding to the smallest resolvable texture detail in the top layer of the image Gaussian pyramid is greater than the expected scale of the coarsest texture on the duct inner wall. Optionally, the image Gaussian pyramid contains four layers.

[0036] In each layer of the Gaussian pyramid of the image, the Hessian matrix response value for each pixel is calculated. For each layer of the Gaussian pyramid, the second-order partial derivatives of that layer with respect to the horizontal and vertical directions are first calculated at each pixel. These second-order partial derivatives are obtained by performing a discrete convolution operation between the layer and the corresponding Gaussian second-order partial derivative convolution kernel. The standard deviation of the Gaussian second-order partial derivative convolution kernel is consistent with the standard deviation of the corresponding Gaussian convolution kernel for that layer. At each pixel, a second-order partial derivative matrix is ​​constructed using the horizontal second-order partial derivative value, the mixed second-order partial derivative value involving both the horizontal and vertical directions, and the vertical second-order partial derivative value. The Hessian matrix response value is calculated from the elements of the second-order partial derivative matrix according to the following expression:

[0037] in, This represents the layer index number of the Gaussian pyramid of the image, and the layer corresponding to the bottom layer of the Gaussian pyramid. The value is 1. The values ​​increase progressively upwards, ranging from 1 to the total number of layers in the Gaussian pyramid of the image. The image represents the Gaussian pyramid of the first generation. The row and column indices of pixels in the layer image. The value ranges from 1 to the An integer representing the total number of rows in the layer image. The value ranges from 1 to the An integer representing the total number of columns in the layer image; The image represents the Gaussian pyramid of the first generation. Layer image at position The value of the second partial derivative along the direction of travel, which is obtained through the first... The layer image is obtained by convolving it with a Gaussian second-order partial derivative convolution kernel along the row direction; The image represents the Gaussian pyramid of the first generation. Layer image at position The value of the second partial derivative along the column direction at that point, this value is obtained through the first... The layer image is obtained by convolving it with a Gaussian second-order partial derivative convolution kernel along the column direction; The image represents the Gaussian pyramid of the first generation. Layer image at position The mixed second-order partial derivative value at the point is obtained through the first... The layer image is obtained by first convolving it with a Gaussian first-order partial derivative convolution kernel along the row direction, and then convolving it with a Gaussian first-order partial derivative convolution kernel along the column direction. The adjustment factor representing the response value of the Hessian matrix. The value is a constant, set The value is 0.9, set The basis for the value of 0.9 is to achieve a trade-off in response intensity when detecting elongated and speckled texture features in the texture features of the inner wall of the pore, so that both elongated and speckled texture features can obtain sufficiently high Hessian matrix response values. The image represents the Gaussian pyramid of the first generation. Layer image at position The Hessian matrix response value is calculated at [location].

[0038] Pixels in the same layer whose Hessian matrix response value is greater than a preset response threshold for that layer are marked as candidate feature points for that layer. For the Gaussian pyramid of an image... Layer, preset response threshold according to the first The statistical distribution of the Hessian matrix response values ​​of all pixels in the layer is determined, and the preset response threshold is set to the first... A fixed proportion of the maximum value of the Hessian matrix response in the layer, which is 0.001. Traverse the... For each pixel in the layer image, if the Hessian matrix response value of the pixel is greater than a preset response threshold, then the pixel is marked as the [number]th [pixel]. Candidate feature points for each layer; if the Hessian matrix response value of a pixel is less than or equal to a preset response threshold, the pixel is ignored. This thresholding process is performed on all layers of the Gaussian pyramid to obtain the candidate feature point set for each layer.

[0039] Spatial location matching is performed on candidate feature points across different layers, and multiple candidate feature points with spatial location deviations less than a preset deviation threshold are grouped into the same group. Candidate feature points between adjacent layers in the Gaussian pyramid of the image are compared, and for the ... Each candidate feature point in the layer is in the Gaussian pyramid of the image at the th . The pixel coordinates in the layer image are mapped to the next layer according to the scaling factor between the adjacent layers. In the pixel coordinate space of the layer image, the mapped coordinate position is obtained. In the first... In the layer image, candidate feature points are searched within a preset radius neighborhood of the mapped coordinate position. If a candidate feature point exists whose spatial position deviation between its actual coordinates and the mapped coordinate position is less than a preset deviation threshold, then the layer is selected as the next layer. Candidate feature points in the layer and the first The candidate feature points found in the layer are grouped into the same group; if there are no candidate feature points in the neighborhood that satisfy the condition that the spatial position deviation is less than a preset deviation threshold, then the layer... Candidate feature points in each layer are grouped separately. The pixel distance value of the preset deviation threshold is determined by a preset radius of 3 pixels. The spatial location matching process is performed layer by layer from the bottom to the top of the image Gaussian pyramid, and cross-layer associations of candidate feature points are established between layers to complete the grouping of candidate feature points in all layers.

[0040] Candidate feature points with the largest Hessian matrix response values ​​within the same group are selected as local texture feature points for that group, resulting in multiple local texture feature points. For each group of candidate feature points after merging, the candidate feature point with the largest absolute value of the Hessian matrix response value is selected from all candidate feature points from different layers within the group, and this candidate feature point is selected as the unique local texture feature point corresponding to that group. The spatial coordinates of the local texture feature point are represented by the pixel coordinates in the Gaussian pyramid layer of the image where the candidate feature point is located, and the pixel coordinates are converted back to the corresponding pixel position in the original uniformly illuminated panoramic image to obtain the final position of the local texture feature point in the original uniformly illuminated panoramic image. The selection operation is repeated for all candidate feature point groups to obtain multiple local texture feature points, each corresponding to a stable texture change position on the inner wall of the channel.

[0041] Example 4:

[0042] In specific implementation, please refer to Figure 5 Based on the distribution density of multiple local texture feature points in a uniformly illuminated panoramic image sequence, multiple slurry-filled regions and multiple void regions on the inner wall of the pore are identified. The process of calculating the width of the slurry interface transition zone on the inner wall of the pore, based on the boundary morphology between these regions, includes: Each frame of the uniformly illuminated panoramic image sequence is divided into multiple statistical grids of equal area. The side length of the statistical grid is set according to the expected texture graininess of the prestressed duct inner wall and the spatial resolution of the panoramic image, ensuring that each statistical grid contains a sufficient number of local texture feature points to reliably distinguish between high-density and low-density areas. Simultaneously, the side length of the statistical grid is smaller than the minimum scale of a typical slurry filling defect. The side length of the statistical grid is 32 pixels. The division begins from the upper left corner of each frame of the uniformly illuminated panoramic image, dividing it into a non-overlapping array of squares that seamlessly cover the entire frame according to the side length of the statistical grid. Each frame of the uniformly illuminated panoramic image consists of several complete statistical grids. When a complete statistical grid cannot be formed at the boundary of the uniformly illuminated panoramic image, the residual pixels at the boundary are not included in subsequent statistics.

[0043] The number of local texture feature points contained within each statistical grid is counted to obtain the feature point count for each statistical grid. For each frame of uniformly illuminated panoramic image, all statistical grids are traversed. For each statistical grid, all local texture feature points falling within the spatial range of that statistical grid in the uniformly illuminated panoramic image of that frame are obtained. The count of the falling local texture feature points is accumulated, and the integer value obtained is the feature point count for that statistical grid. The feature point count reflects the texture richness and slurry coverage of the inner wall of the prestressed duct at the corresponding location of the statistical grid.

[0044] Statistical grids with feature point counts greater than a preset high-density threshold are marked as grout-filled grids, and statistical grids with feature point counts less than a preset low-density threshold are marked as void grids. The preset high-density and preset low-density thresholds are determined by the statistical distribution of local texture feature point densities on the inner wall of the prestressed duct under a known standard compacted grouting state. Under standard compacted grouting, the grout fully fills all the recessed textures on the inner wall of the duct, and the inner wall surface exhibits a uniform random texture, corresponding to a higher value range for feature point counts. Under voided grouting, the inner wall surface exposes a smooth corrugated metal pipe wall or concrete duct wall, the texture is significantly weakened, and the corresponding feature point count distribution is in a lower value range. Optionally, the preset high-density threshold is taken as the lower quartile of the feature point count distribution under standard compacted grouting, and the preset low-density threshold is taken as the upper quartile of the feature point count distribution under standard ungrouted exposed duct wall conditions. For each frame of uniformly illuminated panoramic image, all statistical grids are evaluated one by one. Statistical grids with feature point counts greater than a preset high-density threshold are marked as slurry-filled grids, statistical grids with feature point counts less than a preset low-density threshold are marked as gap grids, and statistical grids with feature point counts between the preset low-density threshold and the preset high-density threshold are ignored and do not participate in subsequent connected component merging operations.

[0045] All slurry-filled grids are merged using four-connected domains to obtain multiple slurry-filled regions. All statistical grids with slurry-filled grid markers are considered foreground pixels in the binary image. The adjacency relationship between statistical grids is determined using four-connectivity adjacency; that is, two statistical grids with slurry-filled grid markers are considered connected if they are adjacent in the horizontal or vertical direction. Starting from any unvisited statistical grid with slurry-filled grid marker, the process recursively expands through four-connectivity adjacency relationships, grouping all four-connected reachable statistical grids with slurry-filled grid markers into the same connected domain, until no unmerged statistical grids with slurry-filled grid markers exist within the four-neighborhood of all statistical grids in the current connected domain. This connected domain corresponds to a slurry-filled region. The above operation is repeated for all slurry-filled grids in all frames of the uniformly illuminated panoramic image sequence, and slurry-filled grids between adjacent frames along the axial direction are also merged while maintaining four-connectivity adjacency, ultimately resulting in multiple slurry-filled regions spanning multiple frames of uniformly illuminated panoramic images.

[0046] All gapped meshes are merged using four-connected domain merging to obtain multiple gapped regions. The four-connected domain merging operation for gapped meshes is based on the same principle as the four-connected domain merging operation for slurry-filled meshes. All statistical meshes with gapped mesh labels are considered as foreground pixels in the binary image. Using the four-connected adjacency determination rule, four-connected reachable gapped meshes are merged into the same gapped region. Cross-frame connected domain merging is also performed between axially adjacent frames to obtain multiple gapped regions spanning multiple frames of uniformly illuminated panoramic images.

[0047] Extract the outer boundary contour of the slurry body for each slurry-filled region and the outer boundary contour of the void region for each void region. For each slurry-filled region, use the set of outer edge coordinates of all statistical grids contained within the region in image space as input, and employ a boundary tracking algorithm to extract a closed and non-self-intersecting outer boundary contour of the slurry body, consisting of a sequence of pixel coordinates. The outer boundary contour of the slurry body encloses the minimum bounding polygon of the slurry-filled region. For each void region, use the same boundary tracking algorithm to extract a closed and non-self-intersecting outer boundary contour of the void region, consisting of a sequence of pixel coordinates.

[0048] The shortest distance between the outer boundary contour line of the slurry body and the outer boundary contour line of the void is taken as the initial boundary spacing between the slurry-filled region and the void region. The formula for calculating the initial boundary spacing is:

[0049] in, The index number of the slurry filling area is indicated as The slurry-filled area, The value is an integer ranging from 1 to the total number of slurry-filled regions; The index number of the void region is The gap area The value can be an integer ranging from 1 to the total number of empty regions; Indicates the first The set of coordinates of all pixels on the outer boundary contour line of the slurry filling region; Indicates the first The set of coordinates of all pixels on the outer boundary contour line of a gap region; Indicates that it is taken from A pixel in the set; Indicates that it is taken from A pixel in the set; Represents pixels Planar coordinates in a uniformly illuminated panoramic image sequence For row coordinates, Column coordinates; Represents pixels Planar coordinates in a uniformly illuminated panoramic image sequence For row coordinates, Column coordinates; Indicates the first The first slurry filling area and the first The initial boundary spacing between the gap regions is a value that applies to all regions. and The minimum value obtained after calculating the Euclidean distance between pixel pairs, in pixels.

[0050] Based on the initial boundary spacing, all void regions adjacent to each grout-filled area are selected. An upper limit for the adjacency determination distance is set, which is determined by several times the maximum possible width of the transition zone between the grout interface and the prestressed duct inner wall. When the... The first slurry filling area and the first Initial boundary spacing between the void regions When the distance is less than or equal to the upper limit of the adjacency determination, the first decision is made. The gap region is related to the first... Each slurry-filled region is adjacent to a void region. Adjacency checks are performed on each pair of slurry-filled regions and each void region to obtain the set of adjacent void regions corresponding to each slurry-filled region.

[0051] From all void regions adjacent to each slurry-filled region, select the void region with the smallest initial boundary spacing as the paired void region for that slurry-filled region. For any given... For each slurry-filled region, find the initial boundary spacing from the set of adjacent void regions of that slurry-filled region. The gap region that yields the minimum value is defined as the first gap region. Each slurry-filled region has a paired void region. The paired void region and the slurry-filled region have the shortest spatial separation distance in this local region.

[0052] The initial boundary distance between the slurry-filled region and the paired void region is taken as the width value of the slurry interface transition zone at that location. For the first... The width of the slurry interface transition zone is taken as the initial boundary distance between the two regions. This initial boundary distance, in pixels, is converted into the actual physical length according to the spatial resolution of the uniformly illuminated panoramic image sequence, thus obtaining the width of the slurry interface transition zone at this location on the inner wall of the duct. After using the initial boundary spacing between the slurry-filled region and the paired void region as the width value of the slurry interface transition zone at that location, the width value of the slurry interface transition zone at that location is redefined as follows to characterize the actual physical width of the grayscale gradient band at the interface between the slurry and the void: In the gap region between the outer boundary contour line of the slurry body and the outer boundary contour line of the void, grayscale values ​​are scanned pixel by pixel along a direction perpendicular to the outer boundary contour line of the slurry body to obtain a grayscale gradient sequence from the outer boundary contour line of the slurry body to the outer boundary contour line of the void; in the grayscale gradient sequence, the pixel distance traversed by the grayscale value decreasing from the average grayscale value of the slurry body region to the average grayscale value of the void region is taken as the width value of the slurry interface transition zone at that location; the average grayscale value of the slurry body region is obtained by averaging the grayscale values ​​of all pixels in the slurry-filled region, and the average grayscale value of the void region is obtained by averaging the grayscale values ​​of all pixels in the void region; when there is no monotonically decreasing trend in the grayscale gradient sequence, the width value of the slurry interface transition zone at that location is set to zero. Cylindrical projection distortion correction is performed on the aforementioned slurry interface transition zone width value to obtain the corrected slurry interface transition zone width value. The focal length parameters of the optical imaging lens of the panoramic camera device and the total number of circumferential pixels in the panoramic image are obtained. Based on the focal length parameters and the total number of circumferential pixels, the circumferential angle increment corresponding to each pixel in the panoramic image is calculated. The circumferential angle increment is equal to twice pi divided by the total number of circumferential pixels. Based on the pixel coordinates of the two endpoint pixels of this slurry interface transition zone width value in the uniformly illuminated panoramic image sequence, the circumferential angle positions corresponding to the two endpoint pixels are calculated respectively. The circumferential angle position corresponding to the endpoint pixel is equal to the column coordinate of that endpoint pixel multiplied by the circumferential angle increment. The difference between the circumferential angle positions corresponding to the two endpoint pixels is taken as the circumferential angle span corresponding to the width value of the slurry interface transition zone. The circumferential angle span is multiplied by the radius value of the inner wall of the duct to obtain the arc length of the inner wall surface of the duct corresponding to the width value of the slurry interface transition zone. This arc length is taken as the corrected width value of the slurry interface transition zone. The radius value of the inner wall of the duct is obtained through the calibration parameters of the panoramic camera device, which include the radial distance from the panoramic camera device to the inner wall of the prestressed duct. The method for determining the preset width threshold is as follows: the width value of the slurry interface transition zone is measured on the inner wall of the prestressed duct under a known standard dense grouting state. The width values ​​of the slurry interface transition zone at multiple measurement positions under the standard dense grouting state are obtained. The average value and standard deviation of the multiple slurry interface transition zone width values ​​are calculated. The average value plus three times the standard deviation is taken as the preset width threshold.

[0053] Example 5:

[0054] In specific implementation, please refer to Figure 6 When the width of the transition zone at the slurry interface exceeds a preset width threshold, the corresponding duct segment is marked as the segment to be verified. Multi-angle local images of the segment to be verified are acquired in the initial panoramic image sequence. Gradient direction histogram feature extraction is performed on the multi-angle local images to obtain multiple gradient direction feature vectors. The slurry void boundary line of the segment to be verified is determined based on the similarity matching results between the multiple gradient direction feature vectors. The process of obtaining the duct grouting density detection value based on the ratio of the area enclosed by the slurry void boundary line to the cross-sectional area of ​​the duct in the segment to be verified includes: Obtain the image position coordinates of the slurry-filled region corresponding to the width value of the slurry interface transition zone in a uniformly illuminated panoramic image sequence. During the calculation of the slurry interface transition zone width value, the slurry-filled region that generates this width value and its corresponding paired void region have been determined. Extract the center pixel coordinates of this slurry-filled region in the uniformly illuminated panoramic image sequence, or extract the geometric center coordinates of the outer boundary contour of the slurry-filled region. Use these coordinates as the image position coordinates of the slurry-filled region corresponding to the width value of the slurry interface transition zone in the uniformly illuminated panoramic image sequence. The image position coordinates include the axial frame index and the intra-frame pixel row and column coordinates.

[0055] The axial position of the duct corresponding to the width of the transition zone at the slurry interface is determined based on the image position coordinates. The duct segment between the previous and next preset distances from this axial position is marked as the segment to be reviewed. Utilizing the correspondence between the frame index of the uniformly illuminated panoramic image sequence and the actual axial position of the prestressed duct, the axial frame index in the image position coordinates is converted into physical length coordinates along the prestressed duct axis, thus obtaining the duct axial position. The preset distance is set based on the prestressed duct diameter and the possible axial range of slurry delamination, and is set to half the inner diameter of the prestressed duct. The starting axial position of the segment to be reviewed is obtained by tracing back one preset distance from the duct axial position, and the ending axial position of the segment to be reviewed is obtained by extending backward one preset distance. The duct segment between the starting and ending axial positions is marked as the segment to be reviewed.

[0056] Extract multiple panoramic images covering the area to be verified from the initial panoramic image sequence. Based on the start and end axial positions of the area to be verified, find the corresponding start and end frame indices in the initial panoramic image sequence, and extract all panoramic image frames between the start and end frame indices to form a multi-frame panoramic image covering the area to be verified.

[0057] Each frame of the panoramic image is segmented according to the circumferential angle, resulting in multiple circumferential angle sub-images. A number of circumferential angle segments is set to ensure that each sub-image has the same angular span; the number of segments is 72, corresponding to an angular span of 5 degrees. For each frame of the panoramic image, the panoramic image is divided into 72 equal-angle sectors along the circumferential direction. Each sector corresponds to an image region that is a circumferential angle sub-image. The width of the circumferential angle sub-image is obtained by dividing the total number of pixels in the panoramic image along the circumferential direction by the number of circumferential angle segments, and the height is the row height of the panoramic image.

[0058] The circumferential angle sub-images with the same circumferential angle are arranged in axial order along the channel to form a multi-angle local image corresponding to each circumferential angle. For each circumferential angle index, the corresponding circumferential angle sub-image is extracted sequentially from multiple frames of panoramic images of the section to be verified. These circumferential angle sub-images are arranged sequentially from the start frame to the end frame along the axial order of the prestressed channel, forming a strip-shaped image that expands along the axial direction. This strip-shaped image is the multi-angle local image for the corresponding circumferential angle. The arrangement operation is repeated for all 72 circumferential angle indices to obtain 72 multi-angle local images.

[0059] For each frame of the multi-angle local image at each angle, the gradient magnitude and gradient direction of each pixel are calculated. The horizontal and vertical gradient operators are convolved with the sub-images respectively to obtain the horizontal and vertical gradient values ​​for each pixel. The gradient magnitude is calculated as the square root of the sum of the squares of the horizontal and vertical gradient values. The gradient direction is calculated as the arctangent function of the vertical and horizontal gradient values, with a value ranging from 0 to 180 degrees. The sign of the gradient direction is ignored to suit HOG features.

[0060] Each frame of sub-image is divided into multiple cells of equal size. The cell size is set based on the pixel resolution of the circumferential angle sub-image and the typical texture scale of the inner wall of the channel, with a value of 8 pixels by 8 pixels. The division starts from the top left corner of the sub-image, dividing the sub-image into a neat array of non-overlapping cells. If there are incomplete cells at the boundaries of the sub-image, the incomplete cells are ignored.

[0061] The gradient direction distribution of all pixels within each cell is statistically analyzed to generate a gradient direction histogram for that cell. The gradient direction histogram consists of nine direction intervals, each covering a 20-degree gradient direction range. For each pixel within a cell, its gradient magnitude is used as a weight, and the gradients are weighted and accumulated according to the direction interval to which the pixel's gradient direction belongs, resulting in a weighted histogram of gradient directions for that cell. This histogram is a vector containing nine elements.

[0062] The gradient direction histograms of all cells in the same sub-image frame are concatenated according to their cell positions to generate the sub-image feature vector for that frame. The cell positions are traversed from left to right by row and from top to bottom between rows. The gradient direction histogram vectors of all cells in the sub-image are concatenated end to end to form a sub-image feature vector with dimensions equal to 9 times the total number of cells.

[0063] The sub-image feature vector corresponding to each frame of the same circumferential angle index is used as the gradient direction feature vector of that frame. These are arranged in axial order to obtain multiple gradient direction feature vectors. The gradient direction feature vectors of each frame under the same circumferential angle index have the same dimension, which is the total number of cells multiplied by 9. For the same multi-angle local image, the sub-image feature vector corresponding to each frame of the multi-angle local image arranged in axial order is used as the gradient direction feature vector of that frame. These are arranged in axial order to form the gradient direction feature vector sequence corresponding to that circumferential angle index. This operation is performed on all circumferential angle indices one by one to obtain multiple gradient direction feature vector sequences corresponding to each circumferential angle index.

[0064] From the gradient direction feature vector sequence corresponding to the circumferential angle index, select the first feature vector located at the beginning of the section to be verified and the second feature vector located at the end of the section to be verified. At the circumferential angle position corresponding to the beginning of the section to be verified, there is no grout void; the gradient direction feature vector at this position is considered a reference representing the dense grouting state. Similarly, at the circumferential angle position corresponding to the end of the section to be verified, there is also no grout void; the gradient direction feature vector at this position is considered another reference representing the dense grouting state. Both the first and second feature vectors have the same dimension as the gradient direction feature vectors of the remaining frames in the sequence.

[0065] Calculate the first cosine similarity between the first feature vector and the gradient direction feature vectors of all other frames, and calculate the second cosine similarity between the second feature vector and the gradient direction feature vectors of all other frames. The cosine similarity is obtained by dividing the inner product of the two vectors by the product of their magnitudes. The weighted sum of the first and second cosine similarities for each gradient direction feature vector yields the comprehensive similarity for that gradient direction feature vector. The expression for the comprehensive similarity is:

[0066] in, The index number of the eigenvector representing the gradient direction. The value can be an integer from 1 to 72; Represents the relationship between the first eigenvector and the second eigenvector. First cosine similarity between eigenvectors in each gradient direction; Indicates the relationship between the second eigenvector and the first eigenvector. Second cosine similarity between eigenvectors in each gradient direction; The weighting coefficients represent the first cosine similarity. The weighting coefficients represent the second cosine similarity. and The value of satisfies ; The value is 0.5. The value is set to 0.5, based on the fact that the reference value of the beginning and end of the section to be reviewed is equal, and the cosine similarity of the two reference ends is given the same trust weight. Indicates the first The combined similarity of eigenvectors in each gradient direction.

[0067] The circumferential angle positions corresponding to gradient direction feature vectors with a comprehensive similarity below a preset similarity threshold are marked as candidate boundaries for slurry vacuolation. All gradient direction feature vectors under all circumferential angle indices are traversed to find those that satisfy... The circumferential angle positions corresponding to gradient direction feature vectors with similarity values ​​less than a preset similarity threshold are marked as candidate boundaries for grout voids. The preset similarity threshold is determined as follows: within a prestressed duct section in a known compacted grouting state, multiple gradient direction feature vector sequences under circumferential angle indices are obtained using the same circumferential angle segmentation method and axial frame extraction method as the section to be reviewed. The comprehensive similarity of each frame's gradient direction feature vector relative to the first and second feature vectors in the sequence is calculated, and the minimum value of all comprehensive similarities is taken as the minimum comprehensive similarity value for the compacted state; or, the same calculation is performed within a prestressed duct section in a known voided state, and the maximum value of the comprehensive similarity is calculated. The median value between the minimum comprehensive similarity value for the compacted state and the maximum comprehensive similarity value for the voided state is taken as the preset similarity threshold. For each circumferential angle index, the overall similarity of the feature vectors of each sub-image under that circumferential angle index is determined frame by frame along the axial direction to see if it is lower than a preset similarity threshold. Frames with an overall similarity lower than the preset similarity threshold are marked as empty frames, and frames with an overall similarity not lower than the preset similarity threshold are marked as dense frames. Intervals with consecutive empty frames are searched along the axial direction. The axial position corresponding to the starting empty frame in the consecutive interval is taken as the starting position of the slurry empty region at that circumferential angle index in the axial direction, and the axial position corresponding to the ending empty frame in the consecutive interval is taken as the ending position of the slurry empty region at that circumferential angle index in the axial direction. When there are no consecutive empty frames under a certain circumferential angle index, it is determined that there is no slurry empty region at that circumferential angle index. Intervals with fewer than a preset frame count threshold are discarded. The preset frame count threshold is set according to the ratio of the axial length of the segment to be reviewed to the axial frame spacing of the panoramic image to avoid misjudging single-frame noise as an empty region.

[0068] All candidate slurry de-cavitation boundaries are connected sequentially according to their circumferential angles to form slurry de-cavitation boundary lines. The method for determining the global zero-point reference for the circumferential angle is as follows: Before pushing the panoramic camera device into the prestressed duct, a fixed circumferential reference mark is set along the duct's axial direction on the inner wall of the prestressed duct. This circumferential reference mark is a reflective marking strip extending continuously along the duct's axial direction, forming a high-contrast strip region with a continuous axial direction and a fixed circumferential position in the uniformly illuminated panoramic image sequence. After acquiring the initial panoramic image sequence, the pixel region of the circumferential reference mark is searched along the circumferential direction for each frame of the panoramic image. The circumferential angle position of this pixel region is taken as the absolute zero point of the circumferential direction for that frame. The circumferential angle coordinates of each circumferential angle sub-image within that frame are recalculated starting from this zero point, ensuring that the circumferential angle coordinates of all axial frames are aligned with the same physical generatrix, thereby eliminating the circumferential angle drift caused by the panoramic camera device's own rotation. Based on this, when the sub-images of the same circumferential angle index are arranged in the order of the axial direction of the channel to form a multi-angle local image, each sub-image corresponds to the same circumferential region near the same generatrix on the inner wall of the channel.

[0069] After obtaining the axial start and end range of the voided area at each circumferential angle index, the circumferential angle index of the voided area is marked as the candidate boundary of the voided slurry, and the candidate boundaries of the voided slurry are arranged in order of the size of their circumferential angle index. In the polar coordinate plane of the prestressed duct cross section, using the circumferential angle corresponding to the circumferential angle index as the polar angle and the inner wall radius of the duct section to be checked as the polar diameter, each candidate boundary of the voided slurry is mapped to a boundary point on the polar coordinate plane, and adjacent boundary points are connected in order of polar angle to form a candidate boundary curve extending along the circumferential direction; for cases where the circumferential angle index is discontinuous and there is a break, a linear interpolation method is used to supplement transition boundary points between two adjacent candidate boundaries of the voided slurry, so that the candidate boundary curve is continuous in the circumferential direction. Subsequently, the axial start and end positions corresponding to each candidate boundary for slurry void removal are mapped onto the same polar coordinate plane, forming start and end boundary points. The start boundary points of each candidate boundary for slurry void removal are connected in polar angle order to form a start-end closed curve, and the end boundary points of each candidate boundary for slurry void removal are connected in polar angle order to form a end-end closed curve. The candidate boundary curves, start-end closed curves, and end-end closed curves together form a closed slurry void removal boundary line. When the number of candidate boundaries for slurry void removal is less than three, it is determined that there are insufficient boundaries to form an effective closed region. The slurry void removal boundary line is set to an empty set, and the area of ​​the void region is determined to be zero, and the grouting density test value of the duct is one.

[0070] The void boundary line of the slurry is projected onto the cross-section of the section to be reviewed, resulting in the projected profile of the voided region. After the void boundary line of the slurry, defined by the circumferential angle and axial length, is closed along the circumferential direction, its radial position is mapped onto the standard circular cross-sectional geometric plane of the section to be reviewed. The radial mapping uses the inner wall radius value of the duct corresponding to the circumferential angle position, forming the projected profile of the voided region on a two-dimensional plane.

[0071] Calculate the first area value enclosed by the projected outline of the empty region. The area of ​​the planar closed region enclosed by the projected outline of the empty region is calculated using Green's formula or its discrete form. The first area value is obtained by summing the coordinates of the polygon vertices of the projected outline of the empty region using the polygon area formula.

[0072] The inner diameter of the duct corresponding to the section to be verified in the initial panoramic image sequence is obtained, and the second area value of the duct cross-section of the section to be verified is calculated based on the inner diameter value. The inner diameter value of the duct is obtained through the calibration parameters of the panoramic camera device, including the radial distance from the panoramic camera device to the inner wall of the prestressed duct. The second area value of the duct cross-section of the section to be verified is calculated using the formula for the area of ​​a circle, which is equal to pi multiplied by the square of half the inner diameter value of the duct.

[0073] Dividing the first area value by the second area value yields the percentage of voided area in the section to be reviewed. The percentage of voided area reflects the proportion of the voided slurry region in the cross-section of the duct.

[0074] Subtract the percentage of voided area from the numerical value to obtain the grouting density test value. The grouting density test value is expressed as a decimal or percentage, with a higher value indicating denser grouting in the section to be checked.

[0075] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for detecting the grouting density of prestressed ducts based on visual inspection, characterized in that, The method includes: Obtain an initial panoramic image sequence of the inner wall of the prestressed duct; Perform illumination non-uniformity correction on the initial panoramic image sequence to obtain a uniformly illuminated panoramic image sequence; Multiple local texture feature points of the inner wall of the channel are extracted based on the uniformly illuminated panoramic image sequence; Based on the distribution density of multiple local texture feature points in the uniformly illuminated panoramic image sequence, multiple slurry-filled regions and multiple void regions on the inner wall of the channel are identified. The width of the slurry interface transition zone on the inner wall of the duct is calculated based on the boundary morphology between the multiple slurry-filled regions and the multiple void regions. When the width value of the transition zone at the slurry interface exceeds a preset width threshold, the corresponding pore section is marked as a section to be verified, and multi-angle local images of the section to be verified in the initial panoramic image sequence are obtained. Gradient direction histogram feature extraction is performed on the multi-angle local images to obtain multiple gradient direction feature vectors; Based on the similarity matching results among multiple gradient direction feature vectors, the slurry delamination boundary line of the section to be verified is determined; The grout compaction test value is obtained by comparing the area enclosed by the grout void boundary line with the cross-sectional area of ​​the duct in the section to be checked.

2. The method for detecting the grouting density of prestressed ducts based on visual inspection according to claim 1, characterized in that, The step of obtaining the initial panoramic image sequence of the inner wall of the prestressed duct includes: The panoramic camera device is pushed into the prestressed duct at a constant speed along the axial direction of the duct. During the push process, the panoramic camera device is controlled to rotate continuously around the axis at a preset angular velocity to capture multiple circumferential images; The images of each circumferential ring are stitched together in the order of acquisition time to form a single-ring panoramic circumferential image. The multiple panoramic ring images are arranged sequentially according to the pushing direction, and the overlapping areas of adjacent two panoramic ring images are registered and fused at the pixel level to obtain the initial panoramic image sequence.

3. The method for detecting the grouting density of prestressed ducts based on visual inspection according to claim 1, characterized in that, The step of performing illumination unevenness correction on the initial panoramic image sequence to obtain a uniformly illuminated panoramic image sequence includes: Each frame of the initial panoramic image sequence is divided into multiple image sub-blocks of equal size; Calculate the average gray value within each of the image sub-blocks, and construct the illumination background surface based on the average gray value of all the image sub-blocks; The illuminated background surface is smoothed by interpolation to obtain the estimated background illumination value for each pixel position. Divide the original grayscale value of each pixel in each frame of the panoramic image by the background illumination estimate corresponding to that pixel location to obtain the corrected grayscale value of each pixel; All the corrected grayscale values ​​are rearranged according to their original pixel positions to generate a corrected image for each frame of the panoramic image. All the corrected images are then arranged in the original sequence order to obtain the uniformly illuminated panoramic image sequence.

4. The method for detecting the grouting density of prestressed ducts based on visual inspection according to claim 1, characterized in that, The step of extracting multiple local texture feature points of the inner wall of the channel based on the uniformly illuminated panoramic image sequence includes: Perform multi-scale Gaussian filtering on each frame of the uniformly illuminated panoramic image sequence to construct the image Gaussian pyramid of that frame. In each layer of the Gaussian pyramid of the image, the Hessian matrix response value of each pixel is calculated; Pixels whose Hessian matrix response value in the same layer is greater than the preset response threshold of that layer are marked as candidate feature points of that layer. Spatial location matching is performed on the candidate feature points between different layers, and multiple candidate feature points with spatial location deviations less than a preset deviation threshold are grouped into the same group. The candidate feature point with the largest Hessian matrix response value in the same group is taken as the local texture feature point of the group, and multiple local texture feature points are obtained.

5. The method for detecting the grouting density of prestressed ducts based on visual inspection according to claim 1, characterized in that, The step of identifying multiple slurry-filled regions and multiple void regions on the inner wall of the pore based on the distribution density of multiple local texture feature points in the uniformly illuminated panoramic image sequence includes: Each frame of the uniformly illuminated panoramic image in the uniformly illuminated panoramic image sequence is divided into multiple statistical grids of equal area; The number of local texture feature points contained in each statistical grid is counted to obtain the feature point count of each statistical grid; The statistical grids with a feature point count greater than a preset high-density threshold are marked as slurry-filled grids, and the statistical grids with a feature point count less than a preset low-density threshold are marked as void grids. All the slurry-filled meshes are merged into four-connected regions to obtain multiple slurry-filled regions; All the aforementioned void meshes are merged into four connected domains to obtain multiple void regions.

6. The method for detecting the grouting density of prestressed ducts based on visual inspection according to claim 1, characterized in that, The step of calculating the width of the slurry interface transition zone on the inner wall of the duct based on the boundary morphology between the plurality of slurry-filled regions and the plurality of void regions includes: Extract the outer boundary contour of the slurry body in each slurry-filled region, and extract the outer boundary contour of the voids in each void region; The shortest distance between the outer boundary contour of the slurry body and the outer boundary contour of the void is taken as the initial boundary spacing between the slurry filling region and the void region. Based on the initial boundary spacing, all void regions adjacent to each of the slurry-filled regions are selected; Select the void region with the smallest initial boundary spacing from all void regions adjacent to each of the slurry filling regions as the paired void region of that slurry filling region; The initial boundary distance between the slurry-filled region and the paired void region is used as the width value of the slurry interface transition zone at that location.

7. The method for detecting the grouting density of prestressed ducts based on visual inspection according to claim 1, characterized in that, The step of marking the corresponding pore section as a section to be verified when the width value of the slurry interface transition zone exceeds a preset width threshold, and obtaining multi-angle local images of the section to be verified in the initial panoramic image sequence, includes: Obtain the image position coordinates of the slurry-filled area in the uniformly illuminated panoramic image sequence, corresponding to the width value of the slurry interface transition zone. The axial position of the channel corresponding to the width value of the transition zone of the slurry interface is determined according to the image position coordinates, and the channel section between the previous preset distance and the next preset distance of the axial position of the channel is marked as the section to be verified. Extract multiple panoramic images covering the section to be reviewed from the initial panoramic image sequence; Each frame of the panoramic image is cut according to the circumferential angle to obtain multiple circumferential angle sub-images; The circumferential angle sub-images with the same circumferential angle are arranged in order of the channel axis to form the multi-angle local image corresponding to each circumferential angle.

8. The method for detecting the grouting density of prestressed ducts based on visual inspection according to claim 1, characterized in that, The step of extracting gradient direction histogram features from the multi-angle local images to obtain multiple gradient direction feature vectors includes: For each frame of the multi-angle local image at each angle, calculate the gradient magnitude and gradient direction of each pixel. Each frame of the sub-image is divided into multiple cells of equal size; Calculate the gradient direction distribution of all pixels in each cell and generate a gradient direction histogram for that cell; The gradient direction histograms of all cells in the same frame of the sub-image are concatenated and stitched together according to the cell position order to generate the sub-image feature vector of the sub-image of that frame. The feature vector of each sub-image under the same circumferential angle index is used as the gradient direction feature vector of that frame. Multiple gradient direction feature vectors are obtained by arranging them in axial order. The gradient direction feature vectors of each frame under the same circumferential angle index have the same dimension, which is the total number of cells multiplied by 9. For the same multi-angle local image, the feature vector of each sub-image in the multi-angle local image arranged in axial order is taken as the gradient direction feature vector of that frame, and arranged in axial order to form the gradient direction feature vector sequence corresponding to the circumferential angle index. Perform this operation on each of the circumferential angle indices to obtain a sequence of gradient direction feature vectors corresponding to each of the multiple circumferential angle indices.

9. The method for detecting the grouting density of prestressed ducts based on visual inspection according to claim 1, characterized in that, The step of determining the slurry void boundary line of the section to be verified based on the similarity matching results among multiple gradient direction feature vectors includes: Select a first feature vector located at the beginning of the segment to be verified and a second feature vector located at the end of the segment to be verified from a plurality of gradient direction feature vectors; Calculate the first cosine similarity between the first feature vector and each of the remaining gradient direction feature vectors, and calculate the second cosine similarity between the second feature vector and each of the remaining gradient direction feature vectors; The first cosine similarity and the second cosine similarity corresponding to each gradient direction feature vector are weighted and summed to obtain the comprehensive similarity of the gradient direction feature vector; The circumferential angle positions corresponding to the gradient direction feature vectors whose comprehensive similarity is lower than a preset similarity threshold are marked as candidate boundaries for slurry delamination. All the candidate slurry devastation boundaries are connected in circumferential angle order to form the slurry devastation boundary line.

10. The method for detecting the grouting density of prestressed ducts based on visual inspection according to claim 1, characterized in that, The step of obtaining the grouting density test value based on the ratio of the area enclosed by the grout void boundary line to the cross-sectional area of ​​the duct in the section to be checked includes: The void boundary line of the slurry is projected onto the cross-section of the section to be verified to obtain the projected outline of the void region. Calculate the first area value enclosed by the projected contour of the voided region; Obtain the inner diameter value of the channel corresponding to the segment to be verified in the initial panoramic image sequence, and calculate the second area value of the channel cross-section of the segment to be verified based on the inner diameter value of the channel. Divide the first area value by the second area value to obtain the percentage of the empty area of ​​the section to be reviewed; Subtracting the percentage of voided area from the numerical value yields the measured value of grout compaction in the duct.

Images