A railway track slab crack topological continuity detection method, system and computer

Abstract

The present application relates to the field of railway technology, and provides a railway track slab crack topological continuity detection method, system and computer, the railway track slab crack topological continuity detection method comprises: obtaining an initial night inspection image, noise suppression and dark light enhancement are carried out on the initial night inspection image; Construct horizontal aggregation branch and vertical aggregation branch to obtain a stitching feature map; Analyze the local gradient amplitude and texture direction of the stitching feature map to obtain a spatial feature importance weight matrix to obtain a noise reduction feature map; Obtain a constraint offset field, migrate the positions of a plurality of convolution kernel sampling points, and obtain a crack prediction map based on a plurality of deformed sampling points; Obtain a plurality of crack entities based on the crack prediction map, calculate the crack length, crack average width and crack maximum width to evaluate the track slab damage grade. By using the above method, the topological discontinuity of the track slab crack is repaired, the false detection rate is reduced, and the accuracy of night crack detection is improved.

Description

Technical Field

[0001] This invention relates to the field of railway technology, and in particular to a method, system, and computer for detecting the topological continuity of cracks in railway track slabs. Background Technology

[0002] As a core foundational component supporting high-speed train operation, the surface health of ballastless track slabs for high-speed railways directly affects the smoothness and safety of train operation. Under the combined effects of long-term dynamic load impacts from trains, alternating temperature stress, and rainwater erosion, the track slab surface is highly susceptible to developing micro-cracks. These cracks are initially extremely narrow, typically less than 0.3 mm, but if not detected and repaired with grout in time, they will rapidly expand into penetrating fractures under water pressure and freeze-thaw cycles, leading to steel reinforcement corrosion, track slab separation, and even slab breakage, seriously threatening railway operational safety. Therefore, comprehensive, high-frequency, and meticulous inspection of track slab cracks is a top priority for railway engineering departments in their daily maintenance.

[0003] Due to the operational scheduling rules of high-speed railways, track slab inspection and maintenance can usually only be scheduled during the "maintenance window" from 0:00 to 4:00 AM. Traditional inspection methods mainly rely on manual inspections by hand using handheld high-powered flashlights on foot along the line. This method is not only extremely inefficient, but also prone to missed inspections due to visual fatigue and decreased attention span caused by prolonged nighttime work, especially for micro-cracks less than 0.5 mm wide, where the reliability of manual identification is difficult to guarantee. Therefore, automated inspection technology based on vehicle-mounted imaging is gradually becoming a hot topic.

[0004] However, existing visual inspection technologies have many shortcomings in the high-speed driving inspection environment during nighttime "window periods." First, under the requirements of nighttime high-speed inspection, the imaging system needs to use auxiliary light sources with extremely short exposure times and high-sensitivity acquisition equipment. This approach introduces severe electronic noise. In the imaging results, a large number of high-frequency noise points are mixed with the pores and textures of the concrete surface, resulting in a strong grainy appearance in the image. The features of micro-cracks are easily submerged in the background noise, causing existing visual inspection algorithms to fail. Second, track slab cracks are usually continuous in physical form, but in nighttime imaging, due to the unevenness of the concrete surface and uneven reflection of light, combined with shadow occlusion under strong light and the coverage of track slab stains, continuous cracks appear discontinuous in the image. Existing models struggle to overcome visual discontinuities, easily leading to misjudgments. In addition, the irregular direction of track slab cracks means that feature sampling by convolutional neural networks in existing methods easily introduces a large number of background pixels around the cracks, contaminating the crack features and resulting in large estimation errors. Therefore, existing visual inspection methods have significant shortcomings in detecting track slab cracks during nighttime inspections. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method, system, and computer for detecting the topological continuity of cracks in railway track slabs. This invention specifically addresses the technical problem of insufficient nighttime detection performance for track slab cracks in existing technologies.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: A method for detecting the topological continuity of cracks in railway track slabs includes the following steps: Acquire several initial nighttime inspection images of the railway track slab surface at night, and perform noise suppression and low-light enhancement on the several initial nighttime inspection images to obtain several enhanced nighttime inspection images; Feature tensors are extracted from the enhanced nighttime patrol images, and horizontal and vertical aggregation branches are constructed. The feature tensors are input into the horizontal aggregation branch to obtain a horizontal context description vector. The feature tensors are input into the vertical aggregation branch to obtain a vertical context description vector. Based on the feature tensors, the horizontal context description vector, and the vertical context description vector, a stitched feature map is obtained. The local gradient magnitude and texture direction of the stitched feature map are analyzed to obtain a spatial feature importance weight matrix. The spatial feature importance weight matrix and the stitched feature map are multiplied element-wise to obtain a noise reduction feature map. Several convolutional kernel sampling points are established. Based on the denoised feature map and the several convolutional kernel sampling points, a constraint offset field is obtained. Based on the constraint offset field, the positions of the several convolutional kernel sampling points are shifted to obtain several deformable sampling points. Based on the several deformable sampling points, a crack prediction map is obtained. Based on the crack prediction map, several crack entities are obtained. A skeleton refinement algorithm is used to extract several central skeleton lines from the crack entities. Based on the central skeleton lines, the crack length, average crack width, and maximum crack width are calculated to assess the damage level of the track slab.

[0007] Furthermore, the step of acquiring several initial nighttime inspection images of the railway track slab surface at night, and performing noise suppression and low-light enhancement on the several initial nighttime inspection images to obtain several enhanced nighttime inspection images includes: An image acquisition device is installed at the bottom of the track inspection vehicle. The image acquisition device includes an industrial camera and a linear stroboscopic light source, which acquires several initial night inspection images of the railway track surface during the nighttime maintenance window. The initial nighttime patrol image is subjected to non-local mean filtering to obtain a denoised grayscale patrol image; The denoised grayscale inspection image is divided into several image blocks, and the grayscale histogram of each image block is calculated. Set a clipping limit, determine the gray levels in the gray histogram that are greater than the clipping limit as the clipping gray levels, select the part in the image block that corresponds to the clipping gray level as the clipping sub-block, clip the clipping sub-block from the image block, and evenly distribute it to the remaining gray levels, so as to convert the denoised gray-level inspection image into an equalized inspection image. Enhanced nighttime inspection images are obtained by processing the equalized inspection images using bilinear interpolation.

[0008] Furthermore, the three dimensions of the feature tensor are the number of channels C, height H, and width W, respectively. The steps of constructing horizontal and vertical aggregation branches, inputting the feature tensor into the horizontal aggregation branch to obtain a horizontal context description vector, and inputting the feature tensor into the vertical aggregation branch to obtain a vertical context description vector, and obtaining the stitched feature map based on the feature tensor, the horizontal context description vector, and the vertical context description vector, include: A horizontal aggregation branch is constructed using a long strip pooling kernel of size H*1. The feature tensor is input into the horizontal aggregation branch to obtain a horizontal context description vector, the dimension of which is C*1*W. A vertical aggregation branch is constructed using a long strip pooling kernel of size 1*W. The feature tensor is input into the vertical aggregation branch to obtain a vertical context description vector, the dimension of which is C*H*1. The horizontal context description vector and the vertical context description vector are transformed by 1*1 convolution, and then expanded to H*W size to obtain expanded horizontal feature map and expanded vertical feature map. The extended horizontal feature map, the extended vertical feature map, and the feature tensor are superimposed element by element to obtain the stitched feature map.

[0009] Furthermore, the step of analyzing the local gradient magnitude and texture direction of the stitched feature map to obtain the spatial feature importance weight matrix includes: The suture feature map is sampled using a sliding window to obtain several region features; A parametric prediction local network is composed of several average pooling layers and several convolutional layers. The local gradient magnitude and texture direction of several regional features are analyzed through the parametric prediction local network to obtain several analysis results. Several analysis results are normalized to obtain several spatial feature importance weight values, which are then combined into a spatial feature importance weight matrix.

[0010] Furthermore, the step of establishing several convolutional kernel sampling points and obtaining the constrained offset field based on the denoised feature map and the several convolutional kernel sampling points includes: Define a standard grid coordinate set, and establish several convolution kernel sampling points based on the standard grid coordinate set; Based on the sampling points of the convolution kernel, the denoised feature map is sampled to obtain several local semantic features. Based on the several local semantic features, the initial offset field is predicted. A preset local receptive field is obtained, and the initial offset field is constrained within the preset local receptive field using a hyperbolic tangent function to obtain a constrained offset field.

[0011] Furthermore, the step of shifting the positions of several convolution kernel sampling points based on the constraint offset field to obtain several deformed sampling points, and obtaining a crack prediction map based on several deformed sampling points, includes: The constraint offset field is superimposed on the standard grid coordinate set, causing the positions of several convolution kernel sampling points to shift, resulting in several deformed sampling points. These deformed sampling points exhibit a bending shape that follows the direction of the crack. Based on several deformation sampling points, the noise reduction feature map is sampled to obtain several feature values, and a crack prediction map is obtained based on the several feature values.

[0012] Furthermore, the step of obtaining several crack entities based on the crack prediction map includes: A segmentation threshold is set, and the crack prediction map is binarized based on the segmentation threshold to obtain a crack mask. By using morphological opening and closing operations, the fine adhesions in the crack mask are broken and the internal pores in the crack mask are filled, and isolated noise points are removed to obtain several crack entities.

[0013] Furthermore, the step of calculating the crack length, average crack width, and maximum crack width based on the central skeleton line includes: The crack length is calculated by counting the number of pixels on the central skeleton line and combining the physical resolution. Calculate the Euclidean distance from the pixel on the central skeleton line to the edge of the crack mask along the normal direction to obtain the crack width. Based on several crack widths, obtain the average crack width and the maximum crack width.

[0014] A railway track slab crack topology continuity detection system, employing the railway track slab crack topology continuity detection method as described in the above technical solution, the system comprising: The acquisition module is used to acquire several initial night inspection images of the railway track slab surface at night, and to perform noise suppression and low light enhancement on the several initial night inspection images to obtain several enhanced night inspection images. An aggregation module is used to extract feature tensors from the enhanced nighttime patrol images, construct horizontal aggregation branches and vertical aggregation branches, input the feature tensors into the horizontal aggregation branches to obtain a horizontal context description vector, input the feature tensors into the vertical aggregation branches to obtain a vertical context description vector, and obtain a stitched feature map based on the feature tensors, the horizontal context description vector and the vertical context description vector. The analysis module is used to analyze the local gradient magnitude and texture direction of the stitched feature map to obtain a spatial feature importance weight matrix, and to perform element-wise multiplication of the spatial feature importance weight matrix and the stitched feature map to obtain a noise reduction feature map. The deformation module is used to establish several convolutional kernel sampling points, obtain a constraint offset field based on the denoised feature map and the several convolutional kernel sampling points, migrate the positions of the several convolutional kernel sampling points based on the constraint offset field to obtain several deformation sampling points, and obtain a crack prediction map based on the several deformation sampling points. The evaluation module is used to obtain several crack entities based on the crack prediction map, extract several central skeleton lines from the crack entities using a skeleton refinement algorithm, and calculate the crack length, average crack width, and maximum crack width based on the central skeleton lines to evaluate the damage level of the track slab.

[0015] A computer includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the railway track slab crack topological continuity detection method as described above.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: Through the horizontal and vertical aggregation branches, orthogonal strip-shaped context aggregation is achieved, enabling the perception of long-distance crack directions. This allows for the reconnection of broken crack segments in the feature space, avoiding the limitations of the traditional square receptive field of convolutional kernels, ensuring the topological integrity of the detection results, effectively repairing topological disconnections in track slab cracks, and improving recall. In low signal-to-noise ratio environments at night, adaptive receptive field spatial weighting is introduced. Specifically, by establishing the spatial feature importance weight matrix, the focus is placed on crack edges with drastic gradient changes, analyzing texture direction, and suppressing the response weights of background noise such as ballast and water stains. Combined with image denoising and enhancement, this provides… The system employs nonlocal mean filtering, grayscale equalization, and bilinear interpolation to effectively address the issues of high noise and difficulty in clear identification caused by high-sensitivity nighttime imaging. It also prevents crack details from being obscured by vignetting effects in the image shadows, significantly reducing the false detection rate. For the meandering and varied morphology of track slab cracks, a non-meshable dynamic deformation convolution strategy is used. This causes the convolution sampling points to shift under certain constraints based on image features, allowing the convolution kernel to automatically deform and conform to the curved path of the crack. This avoids the problem of traditional standard meshed convolution kernel sampling involving surrounding background pixels in feature calculations, achieving pixel-level fine segmentation and reducing crack width measurement errors to the sub-pixel level. This provides a reliable basis for crack damage assessment and precise railway maintenance. Attached Figure Description

[0017] Figure 1 This is a flowchart of the railway track slab crack topological continuity detection method in the first embodiment of the present invention; Figure 2 This is a structural block diagram of the railway track slab crack topological continuity detection system in the second embodiment of the present invention; The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0018] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0019] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] Please see Figure 1 The railway track slab crack topological continuity detection method in the first embodiment of the present invention includes the following steps: Step S10: Obtain several initial night inspection images of the railway track slab surface at night, and perform noise suppression and low-light enhancement on the several initial night inspection images to obtain several enhanced night inspection images; Nighttime track slab inspection faces the contradiction of insufficient lighting and high vehicle speeds. Inspection vehicles typically travel at speeds exceeding 80 km / h. To acquire image-free images at high speeds, extremely short exposure times are necessary. However, these short exposure times result in insufficient light intake, which is compensated for by increasing ISO. High ISO, however, introduces a significant amount of noise. This noise, combined with micropores on the concrete surface and ballast dust, creates a strong grainy appearance in the image. Furthermore, the grayscale characteristics of fine cracks in the track slab are weak and easily submerged in background noise. In addition, due to differences in the angle of illumination and the flatness of the concrete surface, the initial image is bright in the center and dark around the edges, and details inside the cracks are easily obscured in shadows. Noise reduction and local contrast enhancement can help overcome these shortcomings.

[0022] Step S10 includes: S110: An image acquisition device is installed at the bottom of the track inspection vehicle. The image acquisition device includes an industrial camera and a linear stroboscopic light source to acquire several initial night inspection images of the railway track surface during the nighttime maintenance window. Preferably, the image acquisition device includes two high-resolution area array industrial cameras with a resolution better than 4096*3000 pixels and a dynamic range greater than 60dB. It also includes a set of high-brightness LED linear stroboscopic flash sources. The cameras in the image acquisition device have an aperture of F4.0 and an ISO setting of 1600~3200 to adapt to low-light environments. The light source wavelength is selected as 600nm~700nm red light, which has better diffuse reflection characteristics for concrete surfaces and provides better contrast rendering for micro-cracks. Hard-triggered synchronous control technology is employed. The system monitors the train's running distance using a rotary encoder at the wheel axle end, sending a trigger pulse every 0.5mm of travel. This pulse simultaneously triggers the camera's shutter and the light source's driver, controlling the exposure time to be 30μs~50μs. The light source's illumination time is strictly synchronized. This microsecond-level synchronization control ensures a high concentration of light source energy, suppresses ambient stray light, and eliminates motion blur. While acquiring images, the system simultaneously records the corresponding line mileage K value, sleeper number, and high-precision GPS geographic coordinates for each frame, writing this spatial attribute information into the image's metadata.

[0023] S120: Perform non-local mean filtering on the initial nighttime patrol image to obtain a denoised grayscale patrol image; Preferably, unlike traditional Gaussian filtering or median filtering which only utilizes pixel neighborhood information, the nonlocal mean filtering algorithm utilizes the structural self-similarity present in natural images. Specifically, for each pixel in the image, the algorithm searches for other pixel blocks with similar structures to its surrounding neighborhood within the entire search window, and calculates weights based on similarity to perform a weighted average. Understandably, it can very well preserve texture features such as cracks with repetitive linear structures, while effectively smoothing randomly distributed Gaussian white noise, thus obtaining a denoised grayscale inspection image with a high signal-to-noise ratio.

[0024] Furthermore, the formula for the denoised pixel value in the denoised grayscale inspection image is:

[0025] in, Indicates the value of the denoised pixels. This represents the pixels in the initial nighttime inspection image. This represents the reference pixel in the initial nighttime inspection image. Represented by pixels The search window is centered on the search. Represents the reference pixel in the initial nighttime inspection image. The original pixel grayscale value;

[0026] in, Represents the reference pixel. Assign pixels The weighting coefficients, Represents the normalization constant. This represents the original pixel grayscale value in the initial nighttime inspection image. Represented by pixels A local neighborhood window centered on the center, Represented by reference pixel A local neighborhood window centered on the center, This represents the square of the Gaussian-weighted Euclidean distance between all corresponding pixel grayscale values ​​within two local neighborhood windows. The standard deviation of the Gaussian kernel is represented. This represents the attenuation parameter that controls the smoothness of the filter.

[0027] Preferably, the normalization constant is used to ensure that the sum of all weights in the local neighborhood is 1. This formula can preserve the texture feature of cracks with a repetitive linear structure very well, while effectively smoothing randomly distributed Gaussian white noise.

[0028] S130: Divide the denoised grayscale inspection image into several image blocks and calculate the grayscale histogram of each image block; Preferably, the denoised grayscale inspection image is divided into several non-overlapping image blocks. Specifically, in this embodiment, the size of the image block is 8*8 pixels.

[0029] S140: Set a clipping limit, determine the gray levels in the gray histogram that are greater than the clipping limit as the clipping gray levels, select the part in the image block that corresponds to the clipping gray level as the clipping sub-block, clip the clipping sub-block from the image block, and evenly distribute it to the remaining gray levels, so as to convert the denoised gray-level inspection image into an equalized inspection image. Preferably, the equalized inspection image includes several equalized image blocks, and the formula for the equalized grayscale histogram of the equalized image blocks is:

[0030] in, This represents a balanced grayscale histogram. Indicates grayscale level. This represents the function that takes the minimum value. This represents a grayscale histogram. Indicates shear limit;

[0031] in, This indicates the total number of pixels exceeding the frame limit. This represents the function that takes the maximum value.

[0032] Preferably, the shearing limit is set to 2.0, and the portion of the histogram exceeding this limit is cut off and evenly distributed to other gray levels to limit the noise amplification caused by excessive enhancement. The noise amplification caused by excessive enhancement is limited by the above formula, and then the cumulative distribution function is changed to complete the gray level equalization.

[0033] S150: The equalized inspection image is processed by bilinear interpolation to obtain an enhanced nighttime inspection image.

[0034] Preferably, bilinear interpolation is used to eliminate the boundary effect between blocks, significantly brighten the shadow areas inside the cracks, and suppress the highlight spots caused by direct strobe light, thereby increasing the contrast between the cracks and the concrete background by more than 30% and obtaining an image with enhanced local contrast.

[0035] Step S20: Extract feature tensors from the enhanced nighttime patrol image, construct horizontal aggregation branches and vertical aggregation branches, input the feature tensors into the horizontal aggregation branches to obtain horizontal context description vectors, input the feature tensors into the vertical aggregation branches to obtain vertical context description vectors, and obtain stitched feature maps based on the feature tensors, the horizontal context description vectors and the vertical context description vectors. Preferably, the track slab cracks are physically continuous, but often appear broken in nighttime images due to light occlusion. Traditional convolutional neural networks typically use global average pooling (GAP) in deep layers to obtain global semantics, but this easily leads to the loss of all spatial location information. The orthogonal strip context aggregation strategy proposed in this embodiment aims to "stitch" these breakpoints in the feature space. The method of this embodiment can construct a topological morphological coupling enhancement network, which includes an encoder, a feature enhancement neck, and a morphological decoder. In the encoder stage, the orthogonal strip context aggregation strategy is implemented to extract the global linear distribution features of the image along the horizontal and vertical directions to establish long-distance pixel dependencies, thereby repairing the topological disconnection of the cracks in the feature extraction stage and obtaining the stitched feature map.

[0036] In step S20, the three dimensions of the feature tensor are the number of channels C, the height H, and the width W, respectively. Step S20 includes: S210: Construct a horizontal aggregation branch using a long strip pooling kernel of size H*1, input the feature tensor into the horizontal aggregation branch, and obtain a horizontal context description vector, wherein the dimension of the horizontal context description vector is C*1*W; Preferably, each point in the horizontal context description vector aggregates all pixel features of the corresponding column, reflecting the probability distribution of whether a crack exists in that column.

[0037] S220: Construct a vertical aggregation branch using a long strip pooling kernel of size 1*W, input the feature tensor into the vertical aggregation branch to obtain a vertical context description vector, the vertical context description vector having a dimension of C*H*1; Preferably, the vertical context description vector aggregates the crack distribution information for each row.

[0038] S230: The horizontal context description vector and the vertical context description vector are transformed by 1*1 convolution, and the horizontal context description vector and the vertical context description vector are expanded to H*W size to obtain an expanded horizontal feature map and an expanded vertical feature map. Preferably, the two vectors are extended using a bilinear interpolation algorithm.

[0039] S240: The extended horizontal feature map, the extended vertical feature map, and the feature tensor are superimposed element by element to obtain the stitched feature map.

[0040] Understandably, if there is a vertical crack in the image, which is broken in the middle due to shadow, the strong response signals at the top and bottom ends of the crack will be compressed and retained in the vertical context description vector through the vertical aggregation branch. When the vector is expanded back to the original image, the originally broken middle area will receive context information from the top and bottom ends. The broken crack area can aggregate global pixel responses in the same row and column, thus achieving the repair of topological connections at the feature level.

[0041] Step S30: Analyze the local gradient magnitude and texture direction of the stitched feature map to obtain the spatial feature importance weight matrix, and multiply the spatial feature importance weight matrix and the stitched feature map element by element to obtain the noise reduction feature map; Preferably, at the feature neck of the network, the feature map integrates deep semantic information and shallow detail information. At this time, the background noise (such as ballast texture and concrete pores) in the night image is the most serious interference. If convolution is performed directly, this noise will be mistaken for crack features. Therefore, in the processing stage of the feature enhancement neck of the topology coupling enhancement network, this step adopts an adaptive receptive field spatial weighting mechanism to dynamically generate the spatial feature importance weight matrix based on local gradient information. It focuses on the crack edges with drastic gradient changes and analyzes the texture direction to enhance the weak crack edges while suppressing background noise. This plays the role of suppressing the response weight of background noise such as ballast and water stains. Combined with the non-local mean filtering, gray-level equalization processing and bilinear interpolation performed in the image preprocessing process, it effectively solves the problem of many noise points and difficulty in clear recognition caused by high-sensitivity imaging at night.

[0042] Step S30 includes: S310: Perform sliding window sampling on the stitching feature map to obtain several region features; Preferably, the window size is 3*3 pixels.

[0043] S320: A parametric prediction local network is formed by combining several average pooling layers and several convolutional layers. The local gradient magnitude and texture direction of several regional features are analyzed through the parametric prediction local network to obtain several analysis results. Preferably, the parameter prediction local network includes an average pooling layer and a 1*1 convolutional layer. The region features are processed by the average pooling layer to aggregate local background information, and then input into the 1*1 convolutional layer for channel integration. Noise is represented by a chaotic gradient, while cracks are usually represented by gradient changes with significant directionality. The response feature matrix of gradient and texture direction can be extracted through the parameter prediction local network.

[0044] S330: Normalize several of the analysis results to obtain several spatial feature importance weight values, and combine the several spatial feature importance weight values ​​into a spatial feature importance weight matrix.

[0045] Preferably, the analysis results are normalized to weight values ​​between 0 and 1 using a Softmax or Sigmoid activation function. Regions where the analysis results show drastic gradient changes with consistent directions are typically considered potential cracks and assigned high weights close to 1. Regions where the analysis results show flat or messy textures are typically considered background and assigned low weights close to 0. Finally, the spatial feature importance weight matrix is ​​multiplied element-wise with the stitching feature map to achieve a feature-weighted response, thereby generating the denoised feature map. In essence, the response intensity of the convolution kernel is dynamically adjusted according to the saliency of the feature content, thereby significantly suppressing the interference of high-frequency background noise while preserving weak crack signals, and significantly improving the detection robustness in low signal-to-noise ratio environments.

[0046] Step S40: Establish several convolutional kernel sampling points, obtain a constraint offset field based on the denoised feature map and the several convolutional kernel sampling points, shift the positions of the several convolutional kernel sampling points based on the constraint offset field to obtain several deformable sampling points, and obtain a crack prediction map based on the several deformable sampling points. Preferably, the shape of the track slab crack is extremely irregular, often serpentine or tree-branch shaped. The standard 3*3 convolution kernel uses a fixed square grid for sampling. When the square convolution kernel covers the curved crack edge, it will inevitably sample the background pixels outside the crack, resulting in blurred and widened crack edges, which seriously affects the measurement accuracy of the width. In the morphology decoder processing stage of the topology morphology coupling enhancement network, a non-mesh dynamic deformation convolution strategy is adopted. The sampling points of the convolution kernel are driven to move position by predicting the coordinate offset, so as to accurately fit the meandering geometry of the crack and generate a high-precision crack prediction map.

[0047] Step S40 includes: S410: Set a standard grid coordinate set, and establish several convolution kernel sampling points based on the standard grid coordinate set; Preferably, a regular grid with standard 3x3 convolution kernels is used.

[0048] S420: Based on the sampling points of the convolution kernel, the denoised feature map is sampled to obtain several local semantic features. Based on the several local semantic features, the initial offset field is predicted. Preferably, the local semantic information of the current feature map is used as input, and an offset field is predicted through a convolutional layer. For each sampling point of the convolutional kernel, the continuous value offset in the X-axis and Y-axis directions is predicted.

[0049] S430: Obtain a preset local receptive field, and use a hyperbolic tangent function to constrain the initial offset field within the preset local receptive field to obtain a constrained offset field.

[0050] Preferably, in order to prevent the sampling points from drifting too far and causing feature divergence, the hyperbolic tangent function (Tanh) is used to constrain the predicted offset within a preset local receptive field. Specifically, the preset local receptive field ranges from [-R, R], where R is the radius of the convolution kernel. The hyperbolic tangent function is used to smoothly map the initially unconstrained coordinate offset and converge it to the interval (-1, 1).

[0051] S440: The constraint offset field is superimposed on the standard grid coordinate set, causing the positions of several convolution kernel sampling points to shift, thereby obtaining several deformed sampling points, which exhibit a bending shape following the crack direction. Preferably, for example, when faced with a curved crack, several of the deformation sampling points will automatically be arranged in a serpentine pattern to conform to the aggregate direction of the crack.

[0052] S450: Based on several deformation sampling points, the noise reduction feature map is sampled to obtain several feature values, so as to obtain a crack prediction map based on several feature values.

[0053] Preferably, the feature values ​​are obtained by bilinear interpolation and convolution operation is performed to finally output a high-precision crack prediction map. Understandably, the non-mesh sampling method breaks the geometric limitations of traditional convolution, allowing the convolution kernel to automatically adapt to the shape of the crack like an "amoeba," avoiding the mixing of background information and achieving pixel-level fine segmentation.

[0054] Step S50: Based on the crack prediction map, obtain several crack entities, use a skeleton refinement algorithm to extract several central skeleton lines from the crack entities, and calculate the crack length, average crack width and maximum crack width based on the central skeleton lines to assess the track slab damage level.

[0055] Preferably, the probability heatmap output by the topology-coupled enhanced network is converted into quantitative indicators required for engineering.

[0056] Step S50: S510: Set a segmentation threshold, and perform binarization processing on the crack prediction map based on the segmentation threshold to obtain a crack mask; Preferably, the optimal segmentation threshold is adaptively determined using the maximum inter-class variance method, and the crack mask is a binary mask.

[0057] S520: By performing morphological opening and closing operations, the fine adhesions in the crack mask are broken and the internal holes in the crack mask are filled, and isolated noise points are removed to obtain several crack entities.

[0058] Preferably, connected component analysis is used to perform erosion followed by dilation and dilation followed by erosion. A preset threshold is set to remove isolated noise points with an area smaller than the preset threshold. Specifically, the preset threshold is 50 pixels. Understandably, this can result in several clean crack entities.

[0059] S530: Count the number of pixels on the central skeleton line and calculate the crack length based on the physical resolution; Preferably, several central skeleton lines are extracted using the Zhang-Suen skeleton thinning algorithm and thinned to a single pixel width.

[0060] S540: Calculate the Euclidean distance of the pixels on the central skeleton line extending along the normal direction to the edge of the crack mask to obtain the crack width. Based on several crack widths, obtain the average crack width and the maximum crack width.

[0061] Preferably, the line extends bidirectionally along the normal direction and intersects with the binary boundary of the crack to generate two intersection points. The Euclidean distance between the two intersection points is calculated and multiplied by the physical resolution calibration coefficient of the image. This coefficient can be obtained in advance by the Zhang Zhengyou calibration method for the camera.

[0062] Preferably, the average classification probability of all pixels within the crack area corresponding to the crack entity can be used as the confidence level of the crack entity. A graded assessment strategy is constructed based on the crack width to comprehensively determine the damage level. The graded assessment strategy is as follows: Level 1 damage: The judgment condition is that the maximum width of the crack is less than 0.3mm, indicating that this type of crack has not yet formed a through-crack trend and the width is far from the maintenance limit. The assessment system outputs the suggestion of "included in the routine log and regular inspection"; Level 2 damage: The judgment condition is that the maximum width of the crack is greater than or equal to 0.3mm and less than or equal to 0.5mm, indicating that although this type of crack has not reached the mandatory maintenance standard stipulated by the rules, it has the risk of expanding into serious defects. The assessment system outputs the suggestion of "increase the observation frequency and pay attention to the expansion trend"; Level 3 damage: The judgment condition is that the maximum width is greater than 0.5mm, or a through-crack is detected along the longitudinal direction of the track. This judgment strictly corresponds to the condition of "should be repaired in time" in the rules. For example, the severe net cracks of the track slab are greater than 0.5mm and the longitudinal through-cracks are greater than 0.5mm. The assessment system outputs the suggestion of "immediate repair or sealing treatment".

[0063] Please see Figure 2 The railway track slab crack topology continuity detection system provided in the second embodiment of the present invention applies the railway track slab crack topology continuity detection method as described in the first embodiment above. The system includes: The acquisition module 10 is used to acquire several initial night inspection images of the railway track slab surface at night, and to perform noise suppression and low light enhancement on the several initial night inspection images to obtain several enhanced night inspection images. The acquisition module 10 includes: The first unit is used to install an image acquisition device at the bottom of the track inspection vehicle. The image acquisition device includes an industrial camera and a linear stroboscopic light source to acquire several initial night inspection images of the railway track surface during the nighttime maintenance window. The second unit is used to perform non-local mean filtering on the initial nighttime patrol image to obtain a denoised grayscale patrol image. The third unit is used to divide the denoised grayscale inspection image into several image blocks and calculate the grayscale histogram of each image block. The fourth unit is used to set the clipping limit, determine the gray levels in the gray level histogram that are greater than the clipping limit as the clipping gray levels, select the part in the image block that corresponds to the clipping gray level as the clipping sub-block, clip the clipping sub-block from the image block, and evenly distribute it to the remaining gray levels, so as to convert the denoised gray level inspection image into an equalized inspection image. The fifth unit is used to process the equalized inspection image through bilinear interpolation to obtain an enhanced nighttime inspection image.

[0064] The aggregation module 20 is used to extract feature tensors from the enhanced nighttime patrol image, construct horizontal aggregation branches and vertical aggregation branches, input the feature tensors into the horizontal aggregation branch to obtain a horizontal context description vector, input the feature tensors into the vertical aggregation branch to obtain a vertical context description vector, and obtain a stitched feature map based on the feature tensors, the horizontal context description vector and the vertical context description vector. In the aggregation module 20, the three dimensions of the feature tensor are the number of channels C, the height H, and the width W, respectively. The aggregation module 20 includes: The sixth unit is used to construct a horizontal aggregation branch using a long strip pooling kernel of size H*1, input the feature tensor into the horizontal aggregation branch, and obtain a horizontal context description vector, wherein the dimension of the horizontal context description vector is C*1*W. The seventh unit is used to construct a vertical aggregation branch through a long strip pooling kernel of size 1*W, input the feature tensor into the vertical aggregation branch, and obtain a vertical context description vector, wherein the dimension of the vertical context description vector is C*H*1. The eighth unit is used to perform channel transformation on the horizontal context description vector and the vertical context description vector through 1*1 convolution, and to expand the horizontal context description vector and the vertical context description vector to the H*W size to obtain an expanded horizontal feature map and an expanded vertical feature map. The ninth unit is used to perform element-wise superposition of the extended horizontal feature map, the extended vertical feature map, and the feature tensor to obtain a stitched feature map.

[0065] Analysis module 30 is used to analyze the local gradient magnitude and texture direction of the stitching feature map to obtain a spatial feature importance weight matrix, and to perform element-wise multiplication of the spatial feature importance weight matrix and the stitching feature map to obtain a noise reduction feature map. The analysis module 30 includes: The tenth unit is used to perform sliding window sampling on the suture feature map to obtain several region features; The eleventh unit is used to form a parameter prediction local network by combining several average pooling layers and several convolutional layers, and to analyze the local gradient magnitude and texture direction of several regional features through the parameter prediction local network to obtain several analysis results. The twelfth unit is used to normalize several of the analysis results to obtain several spatial feature importance weight values, and to combine the several spatial feature importance weight values ​​into a spatial feature importance weight matrix.

[0066] The deformation module 40 is used to establish a number of convolutional kernel sampling points, obtain a constraint offset field based on the denoised feature map and the number of convolutional kernel sampling points, migrate the positions of the number of convolutional kernel sampling points based on the constraint offset field to obtain a number of deformation sampling points, and obtain a crack prediction map based on the number of deformation sampling points. The deformation module 40 includes: The thirteenth unit is used to set a standard grid coordinate set and establish several convolution kernel sampling points based on the standard grid coordinate set; The fourteenth unit is used to sample the denoised feature map based on the sampling points of the convolution kernel to obtain several local semantic features, and to predict the initial offset field based on the several local semantic features. The fifteenth unit is used to obtain a preset local receptive field. The initial offset field is constrained within the preset local receptive field using a hyperbolic tangent function to obtain a constrained offset field.

[0067] The sixteenth unit is used to superimpose the constraint offset field onto the standard grid coordinate set, so that the positions of several convolution kernel sampling points are shifted to obtain several deformed sampling points, and the several deformed sampling points are in a shape that bends along the direction of the crack. The seventeenth unit is used to sample the denoised feature map based on several deformation sampling points to obtain several feature values, so as to obtain a crack prediction map based on several feature values.

[0068] The evaluation module 50 is used to obtain several crack entities based on the crack prediction map, extract several central skeleton lines from the crack entities using a skeleton refinement algorithm, and calculate the crack length, average crack width, and maximum crack width based on the central skeleton lines to evaluate the damage level of the track slab.

[0069] The evaluation module 50: The eighteenth unit is used to set a segmentation threshold and perform binarization processing on the crack prediction map based on the segmentation threshold to obtain a crack mask. The nineteenth unit is used to break the fine adhesions in the crack mask and fill the internal holes in the crack mask through morphological opening and closing operations, respectively, and to remove isolated noise points to obtain several crack entities.

[0070] The twentieth unit is used to count the number of pixels on the central skeleton line and calculate the crack length by combining the physical resolution. The twenty-first unit is used to calculate the Euclidean distance of the pixel points on the central skeleton line extending along the normal direction to the edge of the crack mask, to obtain the crack width, and based on several crack widths, to obtain the average crack width and the maximum crack width.

[0071] The third embodiment of the present invention provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the railway track slab crack topological continuity detection method as described in the first embodiment.

[0072] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0073] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for detecting the topological continuity of cracks in railway track slabs, characterized in that, Includes the following steps: Acquire several initial nighttime inspection images of the railway track slab surface at night, and perform noise suppression and low-light enhancement on the several initial nighttime inspection images to obtain several enhanced nighttime inspection images; Feature tensors are extracted from the enhanced nighttime patrol images, and horizontal and vertical aggregation branches are constructed. The feature tensors are input into the horizontal aggregation branch to obtain a horizontal context description vector. The feature tensors are input into the vertical aggregation branch to obtain a vertical context description vector. Based on the feature tensors, the horizontal context description vector, and the vertical context description vector, a stitched feature map is obtained. The three dimensions of the feature tensor are the number of channels C, height H, and width W, respectively. The steps of constructing horizontal and vertical aggregation branches, inputting the feature tensor into the horizontal aggregation branch to obtain a horizontal context description vector, and inputting the feature tensor into the vertical aggregation branch to obtain a vertical context description vector, and obtaining the stitched feature map based on the feature tensor, the horizontal context description vector, and the vertical context description vector, include: A horizontal aggregation branch is constructed using a long strip pooling kernel of size H*1. The feature tensor is input into the horizontal aggregation branch to obtain a horizontal context description vector, the dimension of which is C*1*W. A vertical aggregation branch is constructed using a long strip pooling kernel of size 1*W. The feature tensor is input into the vertical aggregation branch to obtain a vertical context description vector, the dimension of which is C*H*1. The horizontal context description vector and the vertical context description vector are transformed by 1*1 convolution, and then expanded to H*W size to obtain expanded horizontal feature map and expanded vertical feature map. The extended horizontal feature map, the extended vertical feature map, and the feature tensor are superimposed element by element to obtain the stitched feature map. The local gradient magnitude and texture direction of the stitched feature map are analyzed to obtain a spatial feature importance weight matrix. The spatial feature importance weight matrix and the stitched feature map are multiplied element-wise to obtain a noise reduction feature map. Several convolutional kernel sampling points are established. Based on the denoised feature map and the several convolutional kernel sampling points, a constraint offset field is obtained. Based on the constraint offset field, the positions of the several convolutional kernel sampling points are shifted to obtain several deformable sampling points. Based on the several deformable sampling points, a crack prediction map is obtained. The step of establishing a plurality of convolutional kernel sampling points and obtaining the constrained offset field based on the denoised feature map and the plurality of convolutional kernel sampling points includes: Define a standard grid coordinate set, and establish several convolution kernel sampling points based on the standard grid coordinate set; Based on the sampling points of the convolution kernel, the denoised feature map is sampled to obtain several local semantic features. Based on the several local semantic features, the initial offset field is predicted. A preset local receptive field is obtained, and the initial offset field is constrained within the preset local receptive field using a hyperbolic tangent function to obtain a constrained offset field. Based on the crack prediction map, several crack entities are obtained. A skeleton refinement algorithm is used to extract several central skeleton lines from the crack entities. Based on the central skeleton lines, the crack length, average crack width, and maximum crack width are calculated to assess the damage level of the track slab.

2. The method for detecting the topological continuity of cracks in railway track slabs according to claim 1, characterized in that, The steps of acquiring several initial nighttime inspection images of the railway track slab surface, and performing noise suppression and low-light enhancement on these initial nighttime inspection images to obtain several enhanced nighttime inspection images include: An image acquisition device is installed at the bottom of the track inspection vehicle. The image acquisition device includes an industrial camera and a linear stroboscopic light source, which acquires several initial night inspection images of the railway track surface during the nighttime maintenance window. The initial nighttime patrol image is subjected to non-local mean filtering to obtain a denoised grayscale patrol image; The denoised grayscale inspection image is divided into several image blocks, and the grayscale histogram of each image block is calculated. Set a clipping limit, determine the gray levels in the gray histogram that are greater than the clipping limit as the clipping gray levels, select the part in the image block that corresponds to the clipping gray level as the clipping sub-block, clip the clipping sub-block from the image block, and evenly distribute it to the remaining gray levels, so as to convert the denoised gray-level inspection image into an equalized inspection image. Enhanced nighttime inspection images are obtained by processing the equalized inspection images using bilinear interpolation.

3. The method for detecting the topological continuity of cracks in railway track slabs according to claim 1, characterized in that, The step of analyzing the local gradient magnitude and texture direction of the stitched feature map to obtain the spatial feature importance weight matrix includes: The suture feature map is sampled using a sliding window to obtain several region features; A parametric prediction local network is composed of several average pooling layers and several convolutional layers. The local gradient magnitude and texture direction of several regional features are analyzed through the parametric prediction local network to obtain several analysis results. Several analysis results are normalized to obtain several spatial feature importance weight values, which are then combined into a spatial feature importance weight matrix.

4. The method for detecting the topological continuity of cracks in railway track slabs according to claim 1, characterized in that, The step of shifting the positions of several convolution kernel sampling points based on the constraint offset field to obtain several deformed sampling points, and obtaining a crack prediction map based on several deformed sampling points includes: The constraint offset field is superimposed on the standard grid coordinate set, causing the positions of several convolution kernel sampling points to shift, resulting in several deformed sampling points. These deformed sampling points exhibit a bending shape that follows the direction of the crack. Based on several deformation sampling points, the noise reduction feature map is sampled to obtain several feature values, and a crack prediction map is obtained based on the several feature values.

5. The method for detecting the topological continuity of cracks in railway track slabs according to claim 1, characterized in that, The step of obtaining several crack entities based on the crack prediction map includes: A segmentation threshold is set, and the crack prediction map is binarized based on the segmentation threshold to obtain a crack mask. By using morphological opening and closing operations, the fine adhesions in the crack mask are broken and the internal pores in the crack mask are filled, and isolated noise points are removed to obtain several crack entities.

6. The method for detecting the topological continuity of cracks in railway track slabs according to claim 5, characterized in that, The steps for calculating the crack length, average crack width, and maximum crack width based on the central skeleton line include: The crack length is calculated by counting the number of pixels on the central skeleton line and combining the physical resolution. Calculate the Euclidean distance from the pixel on the central skeleton line to the edge of the crack mask along the normal direction to obtain the crack width. Based on several crack widths, obtain the average crack width and the maximum crack width.

7. A railway track slab crack topological continuity detection system, employing the railway track slab crack topological continuity detection method as described in any one of claims 1 to 6, characterized in that, The system includes: The acquisition module is used to acquire several initial night inspection images of the railway track slab surface at night, and to perform noise suppression and low light enhancement on the several initial night inspection images to obtain several enhanced night inspection images. An aggregation module is used to extract feature tensors from the enhanced nighttime patrol images, construct horizontal aggregation branches and vertical aggregation branches, input the feature tensors into the horizontal aggregation branches to obtain a horizontal context description vector, input the feature tensors into the vertical aggregation branches to obtain a vertical context description vector, and obtain a stitched feature map based on the feature tensors, the horizontal context description vector and the vertical context description vector. In the aggregation module, the three dimensions of the feature tensor are the number of channels C, the height H, and the width W, respectively. The aggregation module includes: The sixth unit is used to construct a horizontal aggregation branch using a long strip pooling kernel of size H*1, input the feature tensor into the horizontal aggregation branch, and obtain a horizontal context description vector, wherein the dimension of the horizontal context description vector is C*1*W. The seventh unit is used to construct a vertical aggregation branch through a long strip pooling kernel of size 1*W, input the feature tensor into the vertical aggregation branch, and obtain a vertical context description vector, wherein the dimension of the vertical context description vector is C*H*1. The eighth unit is used to perform channel transformation on the horizontal context description vector and the vertical context description vector through 1*1 convolution, and to expand the horizontal context description vector and the vertical context description vector to the H*W size to obtain an expanded horizontal feature map and an expanded vertical feature map. The ninth unit is used to perform element-wise superposition of the extended horizontal feature map, the extended vertical feature map, and the feature tensor to obtain a stitched feature map; The analysis module is used to analyze the local gradient magnitude and texture direction of the stitched feature map to obtain a spatial feature importance weight matrix, and to perform element-wise multiplication of the spatial feature importance weight matrix and the stitched feature map to obtain a noise reduction feature map. The deformation module is used to establish several convolutional kernel sampling points, obtain a constraint offset field based on the denoised feature map and the several convolutional kernel sampling points, migrate the positions of the several convolutional kernel sampling points based on the constraint offset field to obtain several deformation sampling points, and obtain a crack prediction map based on the several deformation sampling points. The deformation module includes: The thirteenth unit is used to set a standard grid coordinate set and establish several convolution kernel sampling points based on the standard grid coordinate set; The fourteenth unit is used to sample the denoised feature map based on the sampling points of the convolution kernel to obtain several local semantic features, and to predict the initial offset field based on the several local semantic features. The fifteenth unit is used to obtain a preset local receptive field. The initial offset field is constrained within the preset local receptive field using a hyperbolic tangent function to obtain a constrained offset field. The evaluation module is used to obtain several crack entities based on the crack prediction map, extract several central skeleton lines from the crack entities using a skeleton refinement algorithm, and calculate the crack length, average crack width, and maximum crack width based on the central skeleton lines to evaluate the damage level of the track slab.

8. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the railway track slab crack topological continuity detection method as described in any one of claims 1 to 6.

Images