Rail surface defect detection system based on computer vision

Abstract

The application discloses a track surface defect detection system based on computer vision and relates to the technical field of defect detection. Continuous images of a track surface after rain are collected, and adjacent frame brightness fluctuation change tracks are recorded synchronously. In a time advancing process, dark line appearance frequency and moving direction information are extracted to generate dynamic dark line initial identification. For the region corresponding to the dynamic dark line initial identification, the brightness change rhythm in the continuous images is compared frame by frame, and the region with periodic fluctuation characteristics is marked as a flow interference region. The application analyzes the continuous images of the track surface through the time dimension, extracts dynamic dark line characteristics and forms a hierarchical recognition path, realizes effective separation of water film disturbance and crack characteristics, improves defect detection stability, simultaneously strengthens stable texture through reverse convergence and edge information superposition, restores real crack characteristics, improves recognition accuracy and reduces the risk of missed detection.

Description

Technical Field

[0001] This invention relates to the field of defect detection technology, and more specifically to a computer vision-based track surface defect detection system. Background Technology

[0002] Track surface defect detection refers to the process of identifying and judging various abnormal states (such as cracks, peeling, wear, crushing, corrosion, etc.) that occur on the running surface of railway rails during long-term loading, friction and environmental effects. Its core is to continuously monitor the structural integrity and service status of the track surface, and to achieve early detection and location of defects as they gradually evolve from small initiations to macroscopic damage, thereby preventing further expansion of defects and the resulting track failure or traffic safety risks.

[0003] Computer vision-based track surface defect detection introduces image perception and pattern recognition technologies into the aforementioned detection process. By continuously imaging the track surface, image data containing texture features, grayscale distribution, and morphological changes are acquired. Based on this, preprocessing, feature extraction, and defect identification are performed on the images. For example, edge information, local contrast changes, or deep learning models are used to identify abnormal areas, transforming defect identification from manual experience-based judgment to a data-driven automated analysis process. During operation, high-speed, non-contact detection can be achieved, and the trend of defect changes can be continuously characterized over time, thereby improving detection efficiency and recognition accuracy, while enhancing the consistency and stability of detection results.

[0004] In scenarios where residual water film covers the rail surface after rain, the water film will undergo continuous deformation and micro-scale flow along the rail surface under the action of train vibration, thus gradually forming directional dynamic dark lines in a continuously acquired image sequence. These dark lines show a high similarity to early cracks in terms of grayscale distribution and morphological characteristics.

[0005] During this process, dynamic dark lines continuously shift in position and evolve in shape over time, thus constantly interfering with the defect identification process. This causes unstable fluctuations in the detection results across multiple keyframes, leading the system to misjudge water film disturbances as structural cracks. Simultaneously, the contrast of actual cracks is weakened under the water film coverage, and their boundary information is gradually obscured in the image. This causes a shift in the defect feature extraction process, ultimately resulting in the neglect and delayed treatment of true defects, which further increases the risk of crack propagation over time.

[0006] The process by which the residual water film covering the track surface after rain is formed is as follows: After rainfall ends, the surface of the rails is not completely dry, and a thin, continuously distributed layer of water remains. This water layer is typically formed by rainwater trapped within the microscopic unevenness of the rail surface. In this state, the water film is not static but undergoes slight flow or shape changes under the influence of vibrations caused by passing trains, airflow disturbances, and temperature variations on the rail surface, thus forming a transparent coating with dynamic characteristics. This water film alters the original light reflection characteristics of the rail surface, causing brightness fluctuations, enhanced local reflections, or blurred textures during image acquisition, thereby interfering with vision-based defect identification.

[0007] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] The purpose of this invention is to provide a computer vision-based track surface defect detection system. By analyzing continuous track surface images in the time dimension, dynamic dark texture features are extracted and a hierarchical recognition path is formed, which effectively separates water film disturbance and crack features, thereby improving the stability of defect detection. At the same time, by enhancing stable texture through reverse convergence and edge information superposition, the true crack features are restored, improving recognition accuracy and reducing the risk of missed detection, thus solving the problems in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a computer vision-based track surface defect detection system, comprising a dynamic dark pattern marking generation module, a flow interference area identification module, a coverage influence area extraction module, a dark pattern trajectory convergence processing module, and a stable texture enhancement and recovery module; The dynamic dark freckle marker generation module acquires continuous images of the track surface after rain and simultaneously records the brightness fluctuation trajectory of adjacent frames. During the time progression, it extracts the frequency of dark freckle occurrence and movement direction information to generate dynamic dark freckle initial markers that evolve over time. The flow interference area identification module compares the brightness change rhythm in the continuous image frame by frame for the area corresponding to the initial mark of the dynamic dark pattern, and marks the area with periodic fluctuation characteristics as the flow interference area, so as to separate the active area affected by the water film in the time dimension. The coverage influence area extraction module, based on the flow interference area, progressively compares the edge clarity of the same location in the previous and next images, extracts the continuously weakened edge segments, and determines the area where the edge disappears intermittently as the coverage influence area, which is used to depict the process of real cracks being covered by water film. The dark texture trajectory convergence processing module performs reverse convergence processing on the trajectory of dark texture changes on the track surface around the coverage area, gradually compressing the texture that has drifted over time back to its initial position, thereby weakening the continuous impact of flow interference on crack identification. The stable texture enhancement and recovery module performs continuous enhancement extraction on the preserved area after inverse convergence processing, and superimposes edge stability information along the time direction to highlight textures that have existed for a long time and are in a stable position, so as to restore the real crack features and suppress dynamic dark texture interference.

[0010] Preferably, the generation of the initial identifier for the dynamic dark pattern includes: Collect continuous images of the track surface after rain and construct spatial coordinate correspondence. Extract the brightness difference at the same position in adjacent frames and stitch them together in time order to form a brightness fluctuation trajectory, while marking the direction of brightness change. Segmented analysis is performed on the trajectory of brightness fluctuations, extracting the alternating segments of brightness increase and decrease and recording the time range, counting the number of occurrences of the changing segments to obtain the frequency of dark patterns, and connecting the center position of brightness change to form a spatial movement path and marking the direction of movement. By combining the frequency of dark pattern occurrence, spatial movement path, and brightness fluctuation trajectory for unified association expression, the brightness change duration in continuous time nodes is recorded and the continuously changing area is extracted to form dynamic change description data with time sequence. The system summarizes the brightness fluctuation trajectory, dark pattern occurrence frequency, and movement direction information corresponding to the continuously changing areas, and outputs them with unified markings to generate the initial identifier of the dynamic dark pattern that evolves over time.

[0011] Preferably, the marking of the flow interference zone includes: Extract the brightness values ​​corresponding to the initial markers of dynamic dark patterns in spatial locations and arrange them in chronological order to form a brightness change sequence. Perform frame-by-frame comparison to record the brightness increase and decrease processes, and string them together to form a continuous brightness change rhythm record. The brightness change rhythm is analyzed to construct brightness fluctuation units and record the time range and time interval. The number of occurrences of brightness fluctuation units is counted and merged to form a repeating brightness fluctuation structure. Regions with periodic fluctuation characteristics are marked. The brightness change trajectory of the corresponding spatial region is integrated by combining the repetitive brightness fluctuation structure and connecting them in time sequence to form a continuous change trajectory and complete the marking of the flow interference area.

[0012] Preferably, the process of analyzing the rhythm of brightness changes and constructing brightness fluctuation units includes: pairing the brightness increase process with the immediately following brightness decrease process to form brightness fluctuation units, recording the corresponding time range and time interval, arranging the brightness fluctuation units in chronological order and counting the number of occurrences, and merging them to form a repeating brightness fluctuation structure.

[0013] Preferably, the determination of the coverage area includes: The spatial range corresponding to the flow interference area is defined, and the image content at the same location in a continuous image sequence is extracted. The gray-scale transition of the neighboring region is recorded and the trajectory of the change in edge clarity is formed. The process involves comparing adjacent time points to identify the trajectory of changes in edge sharpness, extracting the intervals where the edge changes from continuous to discontinuous, and organizing these intervals to form segments where the edge continuously weakens. By combining the time nodes of grayscale boundary line interruption in the continuously weakened edge segment and recording the interruption process, a set of time nodes of intermittent edge disappearance is formed. The sets of continuously weakening edge segments and intermittently disappearing edge time nodes are integrated and sequentially linked, and the corresponding spatial locations are marked as coverage influence areas.

[0014] Preferably, the extraction of the continuously weakened edge segment includes: performing a comparison of the change trajectory of edge clarity at adjacent time nodes, identifying the process of grayscale boundary line changing from a continuous state to an intermittent state and recording the corresponding time interval to form the continuously weakened edge segment.

[0015] Preferably, performing reverse convergence processing on the trajectory of dark pattern changes on the track surface includes: Select the spatial range corresponding to the coverage area and mark the dark pattern position of the starting frame of the time series. Extract the dark pattern region in the continuous image sequence frame by frame and record the center position and boundary range. Connect them to form the dark pattern morphology change trajectory. Based on the morphological change trajectory of dark patterns, frame-by-frame backtracking is performed. The dark pattern regions at each time node are aligned with the previous time node position according to the time sequence and continuously moved towards the initial position. The spatial coordinate changes are recorded to form the shrinkage process. The dark texture areas that have been aligned in position are gathered and spatially superimposed to form a concentrated texture expression centered on the initial position, thereby achieving reverse convergence of the dark texture morphology change trajectory.

[0016] Preferably, the frame-by-frame backtracking process based on the dark pattern morphology change trajectory includes: extracting the spatial coordinates of each time node along the dark pattern morphology change trajectory, aligning the current dark pattern area with the corresponding position of the previous time node in chronological order, while maintaining the continuity of grayscale distribution within the dark pattern area, forming a spatial change process that gradually shrinks to the initial position.

[0017] Preferably, the texture enhancement process includes: Extract the continuous image sequence of the retained region after reverse convergence processing and scan the pixel position frame by frame to record the gray-level transition path and extension range of the neighborhood, forming a set of edge information across time series. The edge information set is subjected to adjacent time node comparison processing, the repeated edge transition paths are marked and superimposed in time order to form an edge path cumulative expression; Continuity analysis is conducted based on the cumulative expression of edge paths, preserving edge paths that maintain connectivity while weakening scattered distribution paths to construct a stable texture structure; By integrating and uniformly arranging the spatial distribution of stable texture structures, textures that are long-lasting and stable in position are highlighted, thus restoring the true crack features and suppressing dynamic dark texture interference.

[0018] Preferably, during the process of performing adjacent time node comparison processing on the edge information set, the edge transition paths that appear repeatedly in the same spatial location are matched point by point and the consistency of the path extension direction is recorded. The paths are then superimposed in chronological order so that the edge transition paths form a continuous cumulative expression in the same spatial location.

[0019] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention performs layered analysis on continuous images of the track surface over time, systematically extracting the frequency of occurrence, direction of movement, and periodic brightness changes of dynamic dark patterns. Furthermore, it establishes a hierarchical identification path between the flow interference zone and the coverage influence zone, clearly separating the dynamic disturbances caused by the water film during the time evolution process. This effectively avoids misjudging water film disturbances as structural cracks, maintains stable output of identification results during continuous detection, reduces identification offset caused by image fluctuations, and ensures consistency of detection results across different time points.

[0020] This invention performs reverse convergence processing on the morphological change trajectory of dark lines within the coverage area, and then superimposes edge stability information in the time direction on the preserved area, thereby continuously enhancing the long-term existing and positionally stable textures. This achieves the gradual recovery of the true crack features and can still highlight the spatial continuity of structural defects even when dynamic interference persists. It also allows the crack boundaries that are covered by water film to reappear, which helps to improve the accuracy of defect identification and reduce the probability of missed detection. Attached Figure Description

[0021] 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.

[0022] Figure 1 This is a schematic diagram of the computer vision-based track surface defect detection system of the present invention. Detailed Implementation

[0023] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0024] This invention provides, for example Figure 1 The computer vision-based track surface defect detection system shown includes a dynamic dark pattern marking generation module, a flow interference area identification module, a coverage influence area extraction module, a dark pattern trajectory convergence processing module, and a stable texture enhancement and recovery module. The dynamic dark freckle marker generation module acquires continuous images of the track surface after rain and simultaneously records the brightness fluctuation trajectory of adjacent frames. During the time progression, it extracts the frequency of dark freckle occurrence and movement direction information to generate dynamic dark freckle initial markers that evolve over time. The initial markers for generating dynamic dark patterns that evolve over time are as follows: First, continuous image acquisition is performed on the track surface after rain. During the acquisition process, multiple frames of track surface images are acquired at fixed time intervals, and each frame is numbered and arranged according to the order of acquisition time to maintain a stable time interval relationship between adjacent images.

[0025] In each frame of the image, a spatial coordinate correspondence is established at the pixel level to ensure that the images at different time points maintain a consistent spatial mapping. Then, for two adjacent frames, brightness values ​​are extracted point by point at the same spatial location, and the brightness difference between the current frame and the previous frame is recorded sequentially in chronological order, forming a brightness change record with time as the horizontal progression dimension and spatial location as the vertical index. In multiple consecutive frames, the brightness difference at the same spatial location is continuously stitched together so that each pixel location corresponds to a complete brightness fluctuation trajectory. This trajectory reflects the brightness change process at that location throughout the entire time period, and the direction of change in the trajectory is marked to distinguish between brightness increase segments and brightness decrease segments, thus forming a set of brightness fluctuation trajectory with directional information.

[0026] In the set of existing brightness fluctuation trajectory, each trajectory is segmented and analyzed in chronological order. The segments in the trajectory where brightness increases and decreases alternately are extracted, and the start and end times of each interval are recorded. At the same time, the number of times such intervals occur throughout the entire time range is counted to obtain the frequency of dark patterns at the corresponding spatial locations.

[0027] After obtaining the frequency information, the location of the brightness change center in the continuous images is tracked. By comparing the locations of the areas with the most concentrated brightness changes in adjacent time nodes, the spatial coordinates of these areas in consecutive frames are connected to form a continuous spatial movement path. During path generation, the direction of each displacement segment is labeled in chronological order to clarify its movement direction in the image coordinates. The direction information from multiple time nodes is then concatenated to form a complete sequence of movement direction information, ensuring that each area experiencing a brightness change has a corresponding frequency of occurrence and a description of its movement direction.

[0028] After obtaining the information on the frequency of dark pattern occurrence and the direction of movement, the brightness fluctuation trajectory, occurrence frequency, and direction of movement information at the same location are uniformly correlated based on spatial location, so that this information can be expressed in the same coordinate system.

[0029] Subsequently, for each region with complete information, the persistence of its brightness changes is recorded frame-by-frame in a continuous time series. Regions showing brightness changes at multiple consecutive time points are marked as continuously changing regions, while regions showing changes only at individual time points are further differentiated. Within continuously changing regions, the change process is further refined according to time sequence, retaining the direction and magnitude of brightness changes at each time point. This information is then arranged in chronological order to form a time-series descriptive data that fully reflects the dynamic change process of the region, ensuring continuity and traceability of the region in the time dimension.

[0030] After completing the above information integration, all continuously changing regions are uniformly summarized and processed. The brightness fluctuation trajectory, dark pattern frequency, and movement direction information of each region in the entire time series are centrally organized and numbered according to time sequence, so that each region corresponds to a complete set of time evolution records.

[0031] During this process, the spatial distribution relationship between different regions is maintained so that the relative positional relationship of each region in the image does not change. Then, these regions with complete temporal evolution information are marked and output as a whole, thereby generating a dynamic dark freckle initial label that evolves over time. This dynamic dark freckle initial label not only contains spatial location information, but also temporal change trajectory, occurrence frequency and movement direction information, providing continuous temporal reference and spatial positioning basis for subsequent identification of flowing interference areas.

[0032] The flow interference area identification module compares the brightness change rhythm in the continuous image frame by frame for the area corresponding to the initial mark of the dynamic dark pattern, and marks the area with periodic fluctuation characteristics as the flow interference area, so as to separate the active area affected by the water film in the time dimension. Regions exhibiting periodic fluctuations are designated as flow disturbance zones. The specific steps are as follows: Starting from the spatial location covered by the initial marker of the dynamic dark pattern, this location is located frame by frame in a continuous image sequence. The brightness value of the corresponding location is extracted in each frame. The brightness values ​​of multiple consecutive time points are arranged in chronological order to form a brightness change sequence for that spatial location. Within this brightness change sequence, frame-by-frame comparison is performed on adjacent time points. The brightness value of the current time point is compared with the brightness values ​​of the previous and next time points. The starting time point, duration, and end time point of the brightness change from low to high are recorded, as well as the starting time point, duration, and end time point of the brightness change from high to low.

[0033] After completing the above recording, all brightness increase and decrease processes at the same spatial location over the entire time range are connected in chronological order, so that the area corresponding to each initial marker of the dynamic dark pattern forms a continuous and complete record of brightness change rhythm. This record not only includes the time and location of each brightness change, but also the duration and direction of the change, thus allowing the brightness change rhythm to be continuously expressed in the time dimension.

[0034] In the established records of brightness variation rhythms, the brightness change process at the same spatial location is segmented. Successive brightness increases are paired with immediately following brightness decreases, forming a complete brightness fluctuation unit for each pair of increases and decreases. The start time, end time, and duration of each brightness fluctuation unit are recorded. After obtaining multiple brightness fluctuation units, these units are arranged in chronological order, and the time intervals between adjacent units are measured to form a complete time interval record.

[0035] Based on this, the occurrence frequency of all brightness fluctuation units within the entire time range is counted, and the duration of each brightness fluctuation unit and the time interval between adjacent units are compared item by item. Brightness fluctuation units with durations within the same time range and adjacent time intervals with consistent distribution are merged, thereby extracting recurring brightness fluctuation structures in the time series. These recurring structures are then marked in spatial location, so that the corresponding regions have the characteristics of continuous and repetitive brightness change rhythm.

[0036] After extracting the repetitive brightness fluctuation structure, the spatial region with the rhythmic characteristics of repetitive brightness changes is matched with the initial marker of the dynamic dark pattern. The brightness change process of this region in the entire continuous image sequence is integrated as a whole, and the brightness fluctuation units are connected in chronological order so that the region forms a continuous change trajectory covering the entire time range.

[0037] In this continuous change trajectory, the brightness state at each time point is recorded one by one, and the brightness change state at the current time point is connected with the brightness change state at the previous and next time points to keep the entire change process continuous and consistent in the time dimension. Subsequently, based on this continuous change trajectory, the corresponding spatial region is uniformly marked and identified as the flow interference zone. This region is presented as an independent identifier in the continuous image and is distinguished from other regions that do not form a repeating brightness change rhythm. Thus, the active area affected by the water film is separated in the time dimension, providing a clear temporal evolution reference for the subsequent extraction of the coverage influence area.

[0038] The coverage influence area extraction module, based on the flow interference area, progressively compares the edge clarity of the same location in the previous and next images, extracts the continuously weakened edge segments, and determines the area where the edge disappears intermittently as the coverage influence area, which is used to depict the process of real cracks being covered by water film. The areas where the edges disappear intermittently are identified as the coverage influence areas. The specific steps are as follows: Within the spatial range covered by the flow interference zone, each frame of the continuous image sequence is processed frame-by-frame for localization. Using fixed spatial coordinates in the image as a reference, the image content at the same location at all time points is extracted accordingly. In each frame, a fixed neighborhood region centered on that location is selected, and the transition of pixel grayscale from bright to dark within that neighborhood region is meticulously recorded. By scanning row by row and column by column, it is determined whether a clear grayscale boundary line exists at that location in the current frame, and the continuity, transition width, and directional consistency of the grayscale boundary line are recorded item by item.

[0039] After completing the single-frame recording, the edge state of that position in all time nodes is arranged in chronological order, so that each spatial position forms a complete record of edge sharpness change. Each time node corresponds to a set of specific edge description information, including the start and end positions of grayscale changes, transition range, and edge continuity, thereby constructing an edge change trajectory covering the entire time range.

[0040] In the established edge change trajectory, the edge clarity at the same spatial location between adjacent time nodes is compared item by item. The edge description information of the current time node is matched one by one with the edge description information of the previous time node. The process of grayscale boundary line changing from concentrated to dispersed and from continuous to discontinuous is recorded, and the start and end positions of the change process on the time axis are clearly marked. In multiple consecutive time nodes, the change segments with gradually weakening edge clarity are extracted. Each time interval that transitions from a continuous and clear edge state to a discontinuous edge state is recorded as an independent segment, and the duration of the segment in the time series is statistically analyzed one by one.

[0041] Subsequently, all segments that meet the characteristics of continuous change are arranged in chronological order, so that the same spatial location forms multiple segments with continuously weakening edges over the entire time range. The start time, end time, and duration of each segment are fully preserved, thus obtaining a set of continuous segments that can reflect the gradual weakening process of the edges.

[0042] After obtaining the continuously weakened edge segment, the edge state of the same spatial location in the entire time series is further refined. The time node where the gray-level boundary line is completely interrupted is found in the edge change trajectory. That is, there is no continuous gray-level transition path in the neighborhood area at the time node. At the same time, the edge state changes before and after the time node are recorded. The process of the edge changing from a discontinuous state to a completely interrupted state and from a completely interrupted state back to a discontinuous state is marked one by one.

[0043] In a continuous time period, the time points at which multiple grayscale boundary lines completely break are summarized and arranged in chronological order, so that each spatial location forms a set of time points where the edge disappears intermittently. At the same time, the start time, end time, and duration of each edge break are recorded, thus forming a complete record of the edge intermittent disappearance change.

[0044] After extracting the sets of continuously weakening edge segments and intermittently disappearing edge time nodes, the two are spatially integrated to uniformly express the edge change process in the time dimension at the same spatial location. The continuously weakening edge segment is regarded as the gradual stage of edge change, and the intermittently disappearing edge time node is regarded as the interruption stage of edge change. The two stages are connected in chronological order so that each spatial location forms a complete time evolution chain containing gradual change and interruption change.

[0045] Subsequently, all spatial locations that simultaneously contain both continuously weakened edge segments and intermittently disappearing edge features are uniformly marked and identified as coverage influence areas. These coverage influence areas not only include specific spatial ranges but also complete temporal change records, thus enabling continuous depiction of the entire process of real cracks gradually changing from being clearly visible to being obscured under the action of water film coverage. This provides a clear temporal evolution basis for subsequent texture restoration processing.

[0046] The dark texture trajectory convergence processing module performs reverse convergence processing on the trajectory of dark texture changes on the track surface around the coverage area, gradually compressing the texture that has drifted over time back to its initial position, thereby weakening the continuous impact of flow interference on crack identification. Around the area of ​​influence, reverse convergence processing is performed on the trajectory of dark stripe morphology changes on the track surface. The specific steps are as follows: Within the spatial range defined by the coverage area, each frame of the continuous image sequence is processed for frame-by-frame localization. The starting frame of the time series is used as the reference frame, and the specific spatial location of the dark pattern is marked in the reference frame, which is then used as the initial position reference point. Subsequently, in the images of each subsequent time node, the same spatial coordinate range is used as the search area, and the texture region corresponding to the dark pattern in the reference frame is searched frame by frame. By scanning pixel by pixel, the dark pattern region in the current frame that is brighter than the surrounding area and is continuously distributed is located, and the center position and boundary range of the region are recorded.

[0047] After completing single-frame localization, the position of the dark freckle region in the current frame is compared one by one with the position of the corresponding dark freckle region in the previous time node. The displacement changes in the horizontal and vertical directions are recorded, and the displacement changes between frames are concatenated in chronological order, so that the spatial position changes of the same dark freckle throughout the entire time series form a continuous record. In this process, the above localization and recording operations are performed on the dark freckle region in each frame of the image, so that each dark freckle in the coverage area forms a complete dark freckle morphology change trajectory. This trajectory includes the specific displacement path and change sequence from the initial position to each subsequent time node, thus providing an accurate trajectory description for subsequent back-convergence processing.

[0048] After obtaining the complete trajectory of dark pattern changes, the initial position in the reference frame is used as a unified reference point. Frame-by-frame backtracking is then performed on the dark pattern regions in subsequent time nodes. At each time node, based on the current position of the dark pattern region in the trajectory, this position is aligned with the corresponding position of the previous time node along the trajectory path, causing the dark pattern region in the current frame to shift towards its position in the previous frame. After aligning the current frame with the previous frame, the aligned dark pattern region is gradually moved closer to its position in an earlier time node in the same way, causing the spatial position of the dark pattern in consecutive time nodes to shrink frame by frame towards its initial position.

[0049] In this process, the position adjustment operation is performed on the dark texture area in each frame of the image, and its new spatial coordinates are recorded after each position adjustment. This makes the spatial distribution of dark texture gradually change from scattered to concentrated throughout the time series. At the same time, the gray-scale distribution inside the dark texture area is maintained during the position adjustment process, so that the dark texture will not be broken or missing during the movement, thus forming a continuous change process of gradually returning to the initial position from back to front.

[0050] After completing frame-by-frame backtracking and position alignment of the dark texture regions in all time nodes, the dark texture regions of each frame that have returned to the vicinity of the initial position are uniformly integrated. The image content corresponding to the same dark texture in different time nodes is spatially superimposed, so that the dark texture in multiple time nodes forms an overlapping expression in the same spatial position.

[0051] During the overlay process, the dark texture regions that have been traced back in each frame of the image are added frame by frame in chronological order, so that the texture information that appears repeatedly at multiple time points gradually accumulates, while the textures that only appear briefly and drift in the time series are compressed to a smaller range during the tracing process, thus forming a concentrated texture distribution centered on the initial position in the overlay result.

[0052] This centralized expression effectively compresses the drift of dark patterns caused by flow interference within the coverage area, and the trajectory of dark pattern morphology changes converges spatially, thereby weakening the continuous impact of flow interference on the crack identification process and providing a unified spatial benchmark for the subsequent enhanced extraction of stable textures.

[0053] The stable texture enhancement and recovery module targets the preserved area after reverse convergence processing, performs continuous enhancement extraction, and superimposes edge stability information along the time direction to highlight textures that have existed for a long time and are stable in position, so as to restore the real crack features and suppress dynamic dark texture interference. To highlight long-standing and location-stable textures, restore realistic crack characteristics, and suppress dynamic dark texture interference, the specific steps are as follows: After the reverse convergence process is completed, the reserved regions corresponding to each time node in the continuous image sequence are located frame by frame. Under a unified spatial coordinate system, all pixel positions in the reserved region of each frame image are scanned point by point. During the scanning process, a fixed range of neighborhood regions is selected with each pixel position as the center, and the gray-level distribution in the neighborhood is read row by row and column by column. The continuous pixel sequence transitioning from the bright area to the dark area is recorded. During the recording process, the start position, end position and continuity of the transition path of the gray-level change are marked item by item. At the same time, the extension length of the transition path in the horizontal and vertical directions is recorded.

[0054] After edge extraction of a single frame image is completed, the edge records of the same spatial location at different time nodes are arranged in chronological order, so that each pixel position corresponds to a set of edge information spanning the entire time range. This set contains the edge existence status, transition path position and path extension at each time node, thus forming the basic data of edge information covering the continuous time series.

[0055] In the established set of edge information, edge records of the same spatial location in adjacent time nodes are compared item by item. The edge transition path of the current time node is matched with the edge transition path of the previous time node, and the overlap in spatial location and the consistency of extension direction are recorded. In multiple consecutive time nodes, the repeated occurrence of edge transition paths in the same spatial location is accumulated and recorded. The locations where edge transition paths exist in multiple consecutive time nodes are marked, and the time node numbers of their occurrence are registered one by one.

[0056] Subsequently, these edge transition paths that repeatedly appear in multiple time nodes are superimposed in chronological order. During the superposition process, the edge paths in each time node are superimposed point by point to the same coordinate position according to their spatial location. This allows the edge paths that repeatedly appear in multiple time nodes to accumulate gradually, while the edge paths that only appear in a few time nodes remain weakly expressed due to the limited number of superpositions. As a result, an edge distribution structure dominated by multiple repeated paths is formed in spatial distribution.

[0057] After completing the edge path overlay, the edge continuity in the overlay result is analyzed point by point. The edge path connection between adjacent pixel positions is checked one by one in the spatial range. Edge paths that maintain continuous connection in multiple time points are retained, and the extension direction of these continuous paths is uniformly recorded so that the edge paths continuously distributed in the same direction form a complete linear structure.

[0058] In this process, edge paths that have shifted in position in the time series and are compressed during the reverse convergence process are distinguished and processed. Since the corresponding positions of such paths are different at different time nodes, they are distributed in a scattered state in the superposition result. These scattered edge paths are weakened point by point so that they cannot form a continuous connection structure in space, thereby enabling the long-term existing and stable edge paths to form a continuous and concentrated texture expression in space.

[0059] After completing the edge continuity processing, the edge overlay results of all pixel positions within the retained area are uniformly integrated. The cumulative edge path situation of each pixel position throughout the entire time series is summarized and rearranged according to spatial coordinates. This allows edge paths that repeatedly appear at multiple time points and maintain stable spatial positions to form a concentrated expression in the same spatial location. During the integration process, the length, direction, and frequency of occurrence of each continuous edge path in the time series are fully preserved, so that these paths present a continuous and extended texture structure in the final image. Edge paths that only appear briefly in the time series and whose positions change are gradually weakened during the integration process, thus forming a spatial distribution dominated by stable textures in the final result.

[0060] Through the above processing, the long-term existing and positionally stable textures in the retained area after reverse convergence processing are highlighted, thereby restoring the true crack characteristics and suppressing the interference of dynamic dark lines caused by water film flow, thus improving the stability and consistency of track surface defect identification.

[0061] By performing layered analysis on continuous images of the track surface over time, the frequency of occurrence, direction of movement, and periodic brightness changes of dynamic dark patterns are systematically extracted. Furthermore, a hierarchical identification path is formed between the flow interference zone and the coverage influence zone, clearly separating the dynamic disturbances caused by the water film during temporal evolution. This effectively avoids misjudging water film disturbances as structural cracks, maintaining stable output of identification results during continuous detection, reducing identification offset caused by image fluctuations, and ensuring consistency of detection results across different time points. By performing reverse convergence processing on the morphological change trajectory of dark patterns within the coverage influence zone, and then superimposing temporal edge stability information on the preserved areas, long-term and positionally stable textures are continuously enhanced, thus achieving gradual recovery of true crack features. Even with persistent dynamic interference, the spatial continuity of structural defects is highlighted, and crack boundaries obscured by the water film are re-emerged, contributing to improved defect identification accuracy and reduced false negatives.

[0062] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A computer vision-based track surface defect detection system, characterized in that, It includes a dynamic dark texture marker generation module, a flow interference area identification module, an overlay influence area extraction module, a dark texture trajectory convergence processing module, and a stable texture enhancement and restoration module; The dynamic dark ripple marker generation module acquires continuous images of the track surface after rain and simultaneously records the trajectory of brightness fluctuations in adjacent frames. During the time progression, it extracts the frequency of dark ripple occurrence and the direction of movement information to generate the initial marker of the dynamic dark ripple. The motion interference area identification module compares the brightness change rhythm in the continuous image frame by frame for the area corresponding to the initial mark of the dynamic dark pattern, and marks the area with periodic fluctuation characteristics as the motion interference area. The coverage influence area extraction module performs a progressive comparison of the edge clarity at the same location in the previous and next images based on the flow interference area, extracts the segments where the edges are continuously weakened, and identifies the areas where the edges disappear intermittently as the coverage influence area; The dark texture trajectory convergence processing module performs reverse convergence processing on the trajectory of dark texture changes on the track surface around the coverage area, gradually compressing the texture that has drifted over time back to its initial position. The stable texture enhancement and restoration module targets the preserved area after reverse convergence processing, performs continuous enhancement extraction, and superimposes edge stability information along the time direction to highlight long-standing and positionally stable textures, restore real crack features and suppress dynamic dark texture interference.

2. The computer vision-based track surface defect detection system according to claim 1, characterized in that, The generation of the initial marker for the dynamic dark pattern includes: Collect continuous images of the track surface after rain and construct spatial coordinate correspondence. Extract the brightness difference at the same position in adjacent frames and stitch them together in time order to form a brightness fluctuation trajectory. Segmented analysis is performed on the trajectory of brightness fluctuations, extracting the alternating segments of brightness increase and decrease and recording the time range, counting the number of occurrences of the changing segments to obtain the frequency of dark patterns, and connecting the center position of brightness change to form a spatial movement path and marking the direction of movement. By combining the frequency of dark pattern occurrence, spatial movement path and brightness fluctuation trajectory for unified association expression, the brightness change duration in continuous time nodes is recorded and the continuously changing area is extracted. The system summarizes the brightness fluctuation trajectory, dark pattern occurrence frequency, and movement direction information corresponding to the continuously changing areas, and outputs them with unified markings to generate the initial identifier of the dynamic dark pattern that evolves over time.

3. The computer vision-based track surface defect detection system according to claim 1, characterized in that, The markings for the flow interference zone include: Extract the brightness values ​​corresponding to the initial markers of dynamic dark patterns in spatial locations and arrange them in chronological order to form a brightness change sequence. Perform frame-by-frame comparison to record the brightness increase and decrease processes, and string them together to form a continuous brightness change rhythm record. The brightness change rhythm is analyzed to construct brightness fluctuation units and record the time range and time interval. The number of times the brightness fluctuation units appear is counted and merged to form a repeating brightness fluctuation structure. The brightness change trajectory of the corresponding spatial region is integrated by combining the repetitive brightness fluctuation structure and connecting them in time sequence to form a continuous change trajectory and complete the marking of the flow interference area.

4. The computer vision-based track surface defect detection system according to claim 3, characterized in that, The process of analyzing the rhythm of brightness changes and constructing brightness fluctuation units includes: pairing the brightness increase process with the immediately following brightness decrease process to form brightness fluctuation units, recording the corresponding time range and time interval, arranging the brightness fluctuation units in chronological order and counting the number of occurrences, and merging them to form a repeating brightness fluctuation structure.

5. The computer vision-based track surface defect detection system according to claim 3, characterized in that, The determination of the coverage area includes: The spatial range corresponding to the flow interference area is defined, and the image content at the same location in a continuous image sequence is extracted. The gray-scale transition of the neighboring region is recorded and the trajectory of the change in edge clarity is formed. The process involves comparing adjacent time points to identify the trajectory of changes in edge sharpness, extracting the intervals where the edge changes from continuous to discontinuous, and organizing these intervals to form segments where the edge continuously weakens. Search for the grayscale boundary line interruption time nodes around the continuously weakening edge segment and record the interruption process, and summarize them to form a set of edge intermittent disappearance time nodes; The sets of continuously weakening edge segments and intermittently disappearing edge time nodes are integrated and sequentially linked, and the corresponding spatial locations are marked as coverage influence areas.

6. The computer vision-based track surface defect detection system according to claim 5, characterized in that, Extraction of persistently weakened edge segments includes: comparing the trajectory of edge sharpness changes between adjacent time nodes, identifying the process of grayscale boundary lines changing from continuous to discontinuous states and recording the corresponding time intervals to form persistently weakened edge segments.

7. The computer vision-based track surface defect detection system according to claim 5, characterized in that, Performing reverse convergence processing on the trajectory of dark pattern changes on the track surface includes: Select the spatial range corresponding to the coverage area and mark the dark pattern position of the starting frame of the time series. Extract the dark pattern region in the continuous image sequence frame by frame and record the center position and boundary range. Connect them to form the dark pattern morphology change trajectory. Based on the trajectory of dark pattern changes, frame-by-frame backtracking is performed, and the dark pattern areas of each time node are aligned with the previous time node position according to the time sequence and continuously moved towards the initial position. The dark texture areas that have been aligned in position are gathered and spatially superimposed to form a concentrated texture expression centered on the initial position, thereby achieving reverse convergence of the dark texture morphology change trajectory.

8. The computer vision-based track surface defect detection system according to claim 7, characterized in that, The frame-by-frame backtracking process based on the dark pattern morphology change trajectory includes: extracting the spatial coordinates of each time node along the dark pattern morphology change trajectory, aligning the current dark pattern area with the corresponding position of the previous time node in chronological order, while maintaining the continuity of grayscale distribution within the dark pattern area, forming a spatial change process that gradually shrinks to the initial position.

9. The computer vision-based track surface defect detection system according to claim 7, characterized in that, Texture highlighting processing includes: Extract the continuous image sequence of the retained region after reverse convergence processing and scan the pixel position frame by frame to record the gray-level transition path and extension range of the neighborhood, forming a set of edge information across time series. The edge information set is subjected to adjacent time node comparison processing, the repeated edge transition paths are marked and superimposed in time order to form an edge path cumulative expression; Continuity analysis is conducted based on the cumulative expression of edge paths, preserving edge paths that maintain connectivity while weakening scattered distribution paths to construct a stable texture structure; By integrating and uniformly arranging the spatial distribution of stable texture structures, textures that are long-lasting and stable in position are highlighted, thus restoring the true crack features and suppressing dynamic dark texture interference.

10. The computer vision-based track surface defect detection system according to claim 9, characterized in that, During the process of performing adjacent time node comparison processing on the edge information set, the edge transition paths that appear repeatedly in the same spatial location are matched point by point and the consistency of the path extension direction is recorded. The paths are then superimposed in chronological order so that the edge transition paths form a continuous cumulative expression in the same spatial location.

Images