A pavement image preprocessing method for crack fine structure preservation

Abstract

The present application relates to the technical field of image processing, and discloses a kind of for crack slender structure fidelity pre-processing method of pavement image.The method reads pavement original image, acquisition log, camera calibration data, detection task data and pre-processing configuration data, forms image access record and pixel scale record;According to the effective field of view of camera, distortion boundary, lane range and pavement texture, demarcate pavement effective area, form acquisition quality record and local background unit;According to base gray, adjacent pixel connection relationship, direction extension relationship and background texture record, form the candidate slender structure with center dark band, side edge band, end point and branch point;Again form gray valley extension, direction continuity, width retention, endpoint adjacency, branch reservation and pseudo slender interference state, execute narrow band fidelity enhancement, original reservation to be checked, interference suppression and prohibition processing, and carry out end point, width, branch and interference isolation check.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, specifically to a method for preprocessing road surface images to preserve the fidelity of slender crack structures. Background Technology

[0002] Road surface crack images are typically acquired using vehicle-mounted cameras, handheld inspection terminals, or unmanned inspection equipment. These images are used for subsequent crack segmentation, length measurement, width estimation, and damage severity assessment. In actual images, early-stage cracks, shallow cracks, and fine cracks often appear as narrow, elongated structures with low grayscale contrast, local discontinuity, and slow directional changes. The same image may also contain interference such as aggregate textures, construction joints, wheel tracks, shadow boundaries, oil and water stains, localized reflections, and slight motion blur. These interferences are easily similar to the elongated structure of cracks in terms of grayscale, edge characteristics, and directional morphology.

[0003] Existing methods for processing road crack images typically perform steps such as grayscale conversion, denoising, homogenization, contrast enhancement, edge detection, thresholding, morphological processing, skeletonization, and breakpoint connection before crack identification. The processed image is then input into a crack segmentation model, a traditional image segmentation algorithm, or a geometric measurement module. These methods can reduce some random noise and highlight crack edges, but their preprocessing often aims to improve overall image quality and usually does not separately constrain the endpoint positions, local widths, directional continuity, and weak grayscale extension relationships of the crack's slender structure during the preprocessing stage.

[0004] Therefore, in road surface images with weak contrast, coarse texture, local shadows, and surface contamination, general denoising and smoothing may weaken the weak grayscale segments of real fine cracks into the background. Thresholding and morphological processing may cause crack chain breaks, endpoint loss, width contraction, edge expansion, and branch deletion. At the same time, pseudo-elongated structures formed by joints, shadow boundaries, and oil and water stains may be simultaneously enhanced and included in subsequent segmentation. Subsequent skeleton repair or model recognition can only continue processing on the already distorted preprocessed results, making it difficult to restore the true continuity and geometric scale of cracks, thus affecting the calculation results of crack length, width, branch morphology, and damage extent. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for preprocessing pavement images to preserve the fidelity of slender crack structures, thus solving the aforementioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for preprocessing pavement images to preserve the fidelity of slender crack structures, executed by a pavement image processing terminal, comprising: S1. Read the original road surface image, acquisition log, camera calibration data, detection task data and preprocessing configuration data to form an image access record and pixel scale record; S2. Based on the camera's effective field of view, distortion boundary, lane range, and road surface texture, delineate the effective area of ​​the road surface and form acquisition quality records and local background units; S3. Based on the base gray level, adjacent pixel connection relationship, directional extension relationship and background texture record in the local background unit, a candidate slender structure with central dark band, side edge band, endpoint and branch point is formed. S4. Based on the central dark band, side edge band, endpoint adjacent units, branch connection positions and background texture records of the candidate slender structure, the corresponding states of gray valley extension, direction continuity, width preservation, endpoint adjacency, branch preservation and pseudo-slender interference are formed. S5. Based on the corresponding states, determine the preprocessing execution path according to the priority of prohibiting processing, preserving as is pending verification, interference suppression, and narrowband fidelity enhancement. Perform corresponding processing on the region where the candidate slender structure is located, and perform endpoint, width, branch, and interference isolation verification on the processing results. If the verification fails, restore the original grayscale of the corresponding position and write it into the rollback record.

[0007] Furthermore, in S1: The image access record number is formed by the detection task number, road segment number, lane number, acquisition date, camera number, and image sequence number in a fixed order. The image access record registers the original image storage location, acquisition log source, camera calibration data number, and preprocessing configuration data number.

[0008] Furthermore, in S1: The pixel scale record is formed based on at least one of the following: camera number, lens number, image resolution, acquisition platform mark, and camera mounting height mark and shooting distance mark corresponding to the acquisition platform; The registration scale is valid when all of the above content falls within the effective range of the camera calibration data; if any inconsistency exists, the registration scale needs to be verified.

[0009] Furthermore, in S2: The effective area of ​​the road surface is the pixel area that is located within the effective field of view of the camera, does not fall outside the distortion boundary, is within the lane range registered for the detection task, and where the difference between the grayscale of the road surface base and the local texture falls within the grayscale processing boundary registered in the preprocessing configuration data.

[0010] Furthermore, in S2: The acquisition quality record registers the exposure quality status, the supplementary lighting consistency status, the motion clarity status, and the reflection occlusion status. The above statuses are formed based on the original image grayscale, camera exposure time, supplementary lighting working mark, acquisition vehicle speed mark, road surface particle edge morphology, highlight area, and acquisition remarks, respectively.

[0011] Furthermore, in S3: Background texture records include the grayscale range of the road surface base, the impact of particle texture, the impact of directional texture, the impact of shadow, the impact of joints, the impact of road markings, the impact of oil and water stains, and the background processability status.

[0012] Furthermore, in S3: The candidate elongated structure is formed by a low grayscale narrow band that extends continuously along the connection relationship of adjacent pixels. The central region of the low grayscale narrow band is lower than the grayscale of the road surface base on both sides, while the road surface texture can still be identified on both sides. The width falls within the width boundary of the candidate elongated structure, and the extension position is not completely covered by the background unusable state.

[0013] Furthermore, in S4: The extension state of gray valleys is formed based on the gray level of the central dark band, the gray levels of the adjacent areas on both sides, the continuity of the low gray level narrow band, and the discernible boundary of the gray valley. The directional continuity state is formed based on the low grayscale narrow band continuation direction, natural bending boundaries, and directional texture influences. The width retention status is formed based on the side edge band spacing, the width boundary of the candidate slender structure, and the influence of the seam. The adjacency state of the endpoints is formed based on the continuation of the low gray-scale narrow band, the side edge band, and the weak extension of the low gray-scale band on the outer side of the endpoints. The branch retention status is determined based on the branch point, branch width, branch direction, and branch connection position; Low grayscale weak extension is a low grayscale continuous pixel segment located in the adjacent unit outside the endpoint, below the grayscale range of the corresponding road surface base, whose width does not meet the stable registration conditions of the candidate slender structure, and whose direction is consistent with the extension direction of the endpoint. The above gray valley extension status is registered as one of the following: extension established, extension pending verification, and extension not established; directional continuity status is registered as one of the following: directional continuity, directional pending verification, and directional discontinuity; width preservation status is registered as one of the following: width preservation, width pending verification, and width not conforming; endpoint adjacency status is registered as one of the following: adjacency established, adjacency pending verification, and adjacency not established; branch retention status is registered as one of the following: branch retention, branch pending verification, and branch not retained.

[0014] Furthermore, in S4: The pseudo-elongated interference state is recorded based on the background texture record, acquisition quality record, gray valley extension state, directional continuity state, width preservation state, endpoint adjacency state, and branch preservation state, and is registered as one of the following: grain dark texture interference, shadow boundary interference, seam interference, mark edge interference, oil and water stain edge interference, reflective edge interference, motion blur edge interference, and non-interference. When the source of interference in the same candidate slender structure only covers part of the location, the interference segment is registered according to the covered location, and the corresponding fidelity verification record is retained for the uncovered location.

[0015] Furthermore, in S5: The road surface image processing terminal forms a preprocessing execution path in the order of prohibition processing, original preservation pending verification, interference suppression, and narrowband fidelity enhancement; When at least one of the following conditions exists: exposure quality status is exposure unavailable, background processable status is background unavailable, reflection occlusion status is reflection occlusion pending verification and endpoint adjacency status is not registered as adjacency established, pixel scale record is scale pending verification and actual size data needs to be output, registration prohibits processing. If at least one of the following exists: extension to be verified, width to be verified, adjacent to be verified, and branch to be verified, and the pseudo-slender interference state is not registered as any of the following: grain dark line interference, shadow boundary interference, seam interference, mark edge interference, oil and water stain edge interference, reflective edge interference, and motion blur edge interference, then the original state is retained for verification. If the pseudo-elongated interference state is registered as any one of the following: grain dark line interference, shadow boundary interference, seam interference, mark edge interference, oil and water stain edge interference, reflective edge interference, and motion blur edge interference, and the gray valley extension state is not registered as extension established, then interference suppression is registered. When the gray valley extension state is established, the width maintenance state is maintained, the pseudo-slender interference state is non-interference, and it does not fall into the prohibited processing area or the original preservation waiting area, register narrowband fidelity enhancement; The output includes a high-fidelity preprocessed image, a crack candidate high-fidelity layer, a thin structure layer to be verified, and a layer of prohibited enhancement areas. When the endpoint consistency verification, width consistency verification, branch consistency verification, and interference isolation verification fail, the original grayscale of the corresponding position is restored and written to the corresponding rollback record.

[0016] Compared with existing technologies, the present invention provides a method for preprocessing pavement images to preserve the fidelity of slender crack structures, which has the following advantages: 1. By reading the original road surface image, acquisition log, camera calibration data, detection task data, and preprocessing configuration data, image access records, pixel scale records, effective road surface area, acquisition quality records, and local background units are formed. Furthermore, candidate elongated structures are formed based on the base gray level, adjacent pixel connection relationship, directional extension relationship, and background texture record. Then, gray level valley extension, directional continuity, width preservation, endpoint adjacency, branch preservation, and pseudo-elongated interference are verified. This makes the road surface image preprocessing no longer rely solely on general denoising, enhancement, and thresholding. This processing can distinguish between real fine cracks and pseudo-elongated structures such as aggregate dark lines, marking edges, joint boundaries, shadow boundaries, and oil and water stain edges during the preprocessing stage. It reduces the occurrence of chain breaks, endpoint loss, width expansion, width contraction, and branch deletion of real cracks during the preprocessing process, providing a more structurally stable image input for subsequent crack segmentation and geometric reading.

[0017] 2. After verification, candidate slender structures are processed into narrowband fidelity enhancement, original preservation for verification, interference suppression, and prohibition paths respectively. Before output, endpoint, width, branch, and interference isolation verification are performed again. Thus, the crack candidate fidelity layer, the slender structure layer to be verified, and the prohibited enhancement area layer can respectively accept real crack candidates, slender dark bands with insufficient evidence, and interference areas such as joints, marking edges, shadow boundaries, oil and water stain edges, and motion blur edges. This reduces the situation where pseudo slender structures enter the crack enhancement range and reduces the missed detection, false detection, and geometric scale deviation of real fine cracks caused by false deletion, false connection, and over-enhancement during preprocessing. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the process for a road surface image preprocessing method for preserving the fidelity of slender crack structures according to the present invention. Detailed Implementation

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

[0020] Example: Figure 1 A pavement image preprocessing method for preserving the slender structure of cracks is presented, executed by a pavement image processing terminal, including: S1. Read the original road surface image, acquisition log, camera calibration data, detection task data, and preprocessing configuration data to form an image access record and pixel scale record, specifically: The road surface image processing terminal reads raw road surface images, acquisition logs, camera calibration data, inspection task data, and preprocessing configuration data. The raw road surface images are sourced from at least one of the following: an area array industrial camera, a line array camera, or a handheld acquisition camera on the inspection vehicle. The acquisition logs are sourced from at least one of the following: the camera acquisition control unit, the data acquisition host on the inspection vehicle, and the local acquisition records on the handheld acquisition terminal. The acquisition logs include the camera number, lens number, image resolution, exposure time, gain value, supplementary light working mark, acquisition time, image sequence number, acquisition platform mark, and shooting distance mark.

[0021] The data acquisition platform marker is used to distinguish between fixed data acquisition by the inspection vehicle, fixed data acquisition by the unmanned inspection equipment, and close-range data acquisition by the handheld terminal; the shooting distance marker is derived from at least one of the following: camera installation and acceptance record, fixed bracket height record, handheld terminal shooting auxiliary ruler record, and on-site data acquisition notes.

[0022] The camera calibration data comes from the shooting and verification records of calibration boards, lane markings and crack samples of known width before data acquisition. It includes the camera's effective field of view, distortion boundary, horizontal pixel scale, vertical pixel scale, camera installation height, applicable shooting distance range, lens focal length and calibration effective range.

[0023] The scope of calibration is jointly defined by the camera number, lens number, image resolution, camera mounting height, applicable shooting distance, and data acquisition platform markings.

[0024] When the inspection vehicle collects data at a fixed location or the unmanned inspection equipment collects data at a fixed location, the camera installation height is determined by the equipment installation acceptance record and the pre-collection verification record. When the handheld terminal collects data at close range, the fixed camera installation height is not used as the sole basis. Instead, the effective boundary of the pixel scale record is determined by the applicable range of shooting distance, auxiliary scale records, and on-site collection notes.

[0025] The data for the inspection tasks comes from the road inspection task book and the on-site data acquisition plan, and includes the inspection task number, road segment number, lane number, acquisition direction, acquisition speed boundary, minimum visible crack width requirement, and re-acquisition conditions. The preprocessing configuration data comes from test images of the same inspection task, manually verified crack samples, joint samples, road marking samples, shadow samples, oil and water stain samples, reflective samples, and coarse aggregate texture samples, and includes local background unit boundaries, candidate slender structure width boundaries, endpoint retention boundaries, side edge protection boundaries, and discontinuity allowable boundaries.

[0026] The aforementioned boundaries are formed by samples taken under the same conditions of camera, lens, image resolution, acquisition platform, road surface type, and acquisition distance, and are not used as fixed parameters across devices, road sections, and road surface types.

[0027] The road image processing terminal generates an image access record number in a fixed order based on the detection task number, road section number, lane number, acquisition date, camera number, and image sequence number. The image access record also records the original image storage location, acquisition log source, camera calibration data number, and preprocessing configuration data number.

[0028] The acquisition date is derived from the acquisition time; if the acquisition time is missing but the image file write time, adjacent image acquisition time and acquisition task batch record can correspond to the same acquisition batch, the acquisition date is written to the batch pending verification mark.

[0029] Image sequence numbers are derived from the camera's continuous acquisition logs. When an image sequence number is missing, the road image processing terminal generates a temporary sequence number based on the image file write time, adjacent image acquisition time, road segment number, lane number, and sampling location, and writes a sequence check mark in the image access record. Images awaiting sequence check can undergo pixel-level preprocessing of a single image and do not participate in endpoint adjacency verification or cross-image verification that depend on the sequence of adjacent images.

[0030] Image access records are used to trace back to the original acquired image for subsequent high-fidelity preprocessed images, crack candidate high-fidelity layers, thin structure layers to be verified, and prohibited enhancement area layers. If at least one of the following exists—missing acquisition time, missing camera number, or missing image resolution—the image access record is registered as an image to be verified; images to be verified only undergo pixel-level preprocessing and do not generate actual width, actual length, or lesion area data.

[0031] If both the camera number and image resolution can be completed using at least one of the image file header information, acquisition log backup records, and device batch records, and the completion result is consistent with adjacent images within the same acquisition batch, the access pending status can be changed to access available, and the completion source record will be retained.

[0032] The road image processing terminal reads camera calibration data and forms a pixel-scale record based on at least one of the following: camera number, lens number, image resolution, acquisition platform mark, and camera installation height mark and shooting distance mark corresponding to the acquisition platform in the current image.

[0033] If the camera number, lens number, image resolution, and acquisition platform mark of the current image all fall within the calibration effective range of the camera calibration data, and the camera installation height mark corresponding to the fixed acquisition by the inspection vehicle and the fixed acquisition by the unmanned inspection equipment falls within the calibration effective range, and the shooting distance mark corresponding to the close-range acquisition by the handheld terminal falls within the calibration effective range, the pixel scale record is registered as available; if any of the above is inconsistent with the calibration effective range, the pixel scale record is registered as pending verification.

[0034] The scaled image allows for subsequent pixel-level preprocessing results and actual size data; the scaled image to be verified retains the pixel-level image processing results, but does not output actual width, actual length, and disease area data.

[0035] If any of the following changes occur: camera, lens, installation height, applicable shooting distance, image resolution, or acquisition platform, the current preprocessing configuration data will no longer be used for newly added images after the change; image access records, pixel scale records, and preprocessing output results generated before the change will still be traced according to the camera calibration data number and preprocessing configuration data number bound at the time of generation.

[0036] Newly added images after the changes need to be re-preprocessed based on new calibration data, test images, and manually reviewed samples to form new preprocessing configuration data. Before the new preprocessing configuration data is formed, newly added images are only allowed to enter access verification, scale verification, and pixel-level preprocessing, and are not allowed to enter actual size output and batch crack geometry reading.

[0037] S2. Based on the camera's effective field of view, distortion boundary, lane range, and road surface texture, delineate the effective area of ​​the road surface and form acquisition quality records and local background units, specifically: The road image processing terminal delineates the effective area of ​​the road surface based on the effective field of view and distortion boundary of the camera recorded at the pixel scale, combined with the lane range registered in the detection task data and the visible road surface texture in the original image.

[0038] The camera's effective field of view and distortion boundary are derived from the pixel-scale record formed by S1. The lane range is derived from the lane number, acquisition direction, and lane detection boundary in the detection task data. The road surface texture is derived from the pixel information corresponding to the road surface base grayscale, aggregate texture, surface roughening texture, old crack traces, and road surface pollution traces in the original image.

[0039] Pixels located within the camera's effective field of view, not falling outside the distortion boundary, within the lane registered for the detection task, and whose surface grayscale and local texture differences fall within the grayscale processing boundary registered in the preprocessing configuration data, are written into the effective surface area. Detection vehicle body edges, camera bracket shadows, image black borders, objects outside the lane, obstructions from the acquisition equipment, and roadside objects unrelated to the detection lane are not written into the effective surface area.

[0040] Lane markings, manhole covers, water stains, shadows, and localized stains located within the detection lane are not directly removed during the delineation of the effective road surface area. If the above areas can still display at least one of the following: road surface base grayscale, aggregate texture, and surface roughening texture, they are retained within the effective road surface area and recorded in the local background impact record.

[0041] The local background impact record registers the corresponding types of impacts from road markings, manhole covers, water stains, shadows, and dirt, while also recording the boundaries of the affected area, the source of the impact, and the connection between the affected area and the surrounding pavement texture. This process is designed to prevent cracks from being truncated during the effective pavement area formation stage when they cross road markings, manhole cover edges, water stain edges, shadow boundaries, and dirt boundaries.

[0042] If the above area completely covers the road surface texture and the difference between the grayscale of the road surface base and the local texture cannot be distinguished, then the area is written into an unprocessable impact record and is not used as the basis for terminating the endpoint of the subsequent slender structure.

[0043] The road surface image processing terminal generates a quality record for the effective area of ​​the road surface. This record registers the exposure quality status, illumination consistency status, motion clarity status, and reflection / occlusion status. The exposure quality status is determined by the original image grayscale, camera exposure time, gain value, and the grayscale processing boundaries registered in the test image.

[0044] The grayscale processable boundary is registered by test images and manually verified samples under the same detection task, and is used to distinguish processable road surface texture, local texture to be verified, and unprocessable areas.

[0045] When the grayscale of the road surface substrate, aggregate texture, and thin dark bands can be distinguished within the effective area of ​​the road surface, the exposure quality status is registered as "exposure available." When local areas are too dark or too bright, but the grayscale of the road surface substrate and some textures can still be distinguished, the exposure quality status is registered as "exposure pending verification." When excessively dark or bright areas meet the re-acquisition conditions registered in the detection task data, and the differences in the grayscale of the road surface substrate and local textures cannot be distinguished, the exposure quality status is registered as "exposure unavailable." Areas with unavailable exposure will not be processed for enhancement, suppression, or connection; they will only serve as a re-acquisition prompt and a reason for prohibition.

[0046] The consistency of the supplementary lighting is determined by the supplementary lighting work markers, the grayscale values ​​of the image center region, the grayscale values ​​of the image edge region, and the grayscale changes of adjacent images. The image center region and the image edge region are formed according to the position range registered in the preprocessing configuration data within the camera's effective field of view. The position range is registered from the test images and the camera supplementary lighting acceptance records.

[0047] When the supplementary light is working normally and the grayscale changes in the image center and edge regions fall within the supplementary light processing boundaries registered in the preprocessing configuration data, the supplementary light consistency status is registered as supplementary light consistency. When the supplementary light is missing a working mark, but the grayscale changes in adjacent images within the same acquisition batch still fall within the supplementary light processing boundaries, the supplementary light consistency status is registered as supplementary light pending verification. When the image has a cross-regional light-dark boundary, the light-dark boundary passes through the thin, dark band to be processed, and the grayscale change records of adjacent images cannot support supplementary light stability, the supplementary light consistency status is registered as supplementary light inconsistency.

[0048] If the image access record in S1 is in sequential pending verification and a reliable adjacent image cannot be called, then the grayscale change record of the adjacent image will not be used in the consistent lighting state. Instead, it will be formed by the grayscale processing boundary of the current image's center area, edge area, and test image, and the adjacent image will be marked as unavailable synchronously.

[0049] The motion clarity is determined by the vehicle speed marker, exposure time, road surface particle edge morphology, and texture location of overlapping areas in adjacent images. The vehicle speed marker is derived from at least one of the following: the detection vehicle's acquisition log, wheel speed records, and acquisition platform records. The exposure time is derived from the acquisition log, and the road surface particle edge morphology is derived from the edge extension direction of aggregate particles, road marking edges, manhole cover edges, and thin dark stripes on the road surface in the current image.

[0050] When the vehicle speed and exposure time fall within the acquisition boundary of the detection task data registration, and the edges of road surface particles do not form a dragging pattern along the same direction, the motion clear state is registered as motion clear; when the vehicle speed mark is missing, but the boundaries of road surface particles, marking edges, and thin dark bands to be processed can still be distinguished, the motion clear state is registered as motion pending verification; when the edges of road surface particles, marking edges, and thin dark bands to be processed all form a dragging pattern along the acquisition direction, the motion clear state is registered as motion blur.

[0051] If adjacent images are unavailable, the motion clarity status is formed solely based on the edge morphology of the current image, exposure time, and acquisition vehicle speed, and is written into the adjacent image unavailable flag.

[0052] The reflective occlusion status is formed by the bright areas, reflective boundaries, acquisition notes, and the visibility of textures in adjacent images. The bright areas and reflective boundaries are derived from the areas in the original image whose brightness reaches the bright boundaries registered in the preprocessing configuration data. The acquisition notes are derived from the records of on-site acquisition personnel regarding humidity, water accumulation, reflection, and contamination.

[0053] When the reflective spot does not cover the main extension area of ​​the slender dark band to be treated, the reflective occlusion state is registered as "reflection does not affect". When the reflective spot covers a part of the road surface, but the central area and the texture on both sides of the slender dark band to be treated are still identifiable, it is registered as "partial reflective nucleus". When the reflective spot covers the slender dark band to be treated, and the central dark band is covered, it is registered as "reflective occlusion nucleus". The reflective occlusion nucleus area is not used as the termination position of the slender structure, and no enhancement or suppression treatment is performed.

[0054] Subsequently, the road image processing terminal divides the effective area of ​​the road surface into local background units based on image resolution, pixel scale recording, road surface type data, aggregate texture scale, and minimum visible crack width requirements. The boundaries of the local background units are registered by preprocessing configuration data, which comes from test images and manually verified samples under the same detection task.

[0055] The size of local background cells is used to limit the formation range of subsequent base grayscale, background texture, and thin dark bands, and is not used as a fixed size across cameras, road surface types, and acquisition heights. Regions with insufficient full size located at the edge of the effective road surface area, near the lane boundary, and inside the distortion boundary can still form edge local background cells; edge local background cells only participate in subsequent processing within the effective road surface area they actually cover, and are written into the edge cell marker.

[0056] In the urban asphalt pavement example, the image resolution is 4096 pixels by 3000 pixels. Manually reviewed samples show that fine cracks in the image are mostly one to nine pixels wide. The preprocessing configuration data registers local background units as 64 pixels by 64 pixels. This value is only effective within the same camera, the same installation height, the same pavement type, and the same detection task. After the local background units are formed, the pavement image processing terminal transmits the acquisition quality record, local background influence record, and edge unit marker to S3 for the formation of background texture records and candidate elongated structures.

[0057] S3. Based on the base grayscale, adjacent pixel connectivity, directional extension, and background texture records within the local background unit, a candidate elongated structure with a central dark band, side edge bands, endpoints, and branch points is formed, specifically: The road surface image processing terminal forms a background texture record within each local background unit and forms candidate slender structures based on the background texture record. The background texture record registers the grayscale range of the road surface substrate, the influence of grain texture, the influence of directional texture, the influence of shadow, the influence of joints, the influence of road markings, the influence of oil and water stains, and the background processable status.

[0058] The grayscale range of the road surface base is formed by road surface pixels within the local background unit that are not affected by markings, joints, water stains, shadows, dirt, or unprocessable effects. When there is a thin, elongated dark band to be processed within the local background unit, the road surface pixels that are still recognizable on both sides of the thin, elongated dark band to be processed are used as the source of the base grayscale. The thin, elongated dark band itself does not participate in the formation of the road surface base grayscale range, thus avoiding the grayscale of fine cracks being incorporated into the background aperture.

[0059] The particle texture effect record is formed by the aggregate dark patterns in the form of dots, blocks, or short lines within local background units. The location, span, and whether the dark pattern enters the adjacent local background unit are recorded. The dark pattern span boundary is derived from coarse aggregate texture samples, test images, and manually verified samples from the same detection task, and is not used as a fixed value across road sections.

[0060] The directional texture influence record is formed by the continuous entry of the same directional texture into adjacent local background units, and the texture direction, continuous unit range and texture source are registered; when the source of the same directional texture is unknown and its direction difference falls into the direction to be kernel boundary registered in the preprocessing configuration data, the directional texture to be kernel mark is written.

[0061] The shadow effect record is formed by the shadow effect transmitted by S2, the light and dark transition boundary of adjacent local background units, and the shadow coverage area; the joint effect record is formed by the location data of contraction joints and construction joints in the inspection task data, as well as the wide and straight dark bands in the image; the lane marking effect record is formed by the lane marking location, the light and dark changes of the marking edge, and the lane range in the inspection task data; the oil and water stain effect record is formed by the dark areas, the boundary curvature, the reflective boundary, and the on-site collection notes.

[0062] When the difference between the grayscale of the road surface substrate and the local texture can be distinguished within a local background unit, the background is registered as usable. When a local background unit is affected by at least one of the following: shadows, water stains, road markings, joints, or dirt, but still retains the source of the road surface substrate grayscale, the background is registered as pending verification.

[0063] When a local background unit is covered by reflection, occlusion, overexposure, or underexposure, and cannot form a grayscale range of the road surface substrate, the background processing state is registered as unusable. Unusable background units are not used as the termination basis for candidate slender structures; when the slender dark band to be processed extends to the boundary of the unusable background unit, the road image processing terminal registers the extension to be kerneled mark and transfers the boundary to the endpoint adjacency state formation process of S4.

[0064] Edge local background units only form background texture records within the effective area of ​​the actually covered road surface. Edge local background units that are not large enough to be considered as background unusable are not considered as reasons for background unusability.

[0065] The candidate elongated structure is formed by the original image grayscale, local background units, road surface base grayscale range, adjacent pixel connectivity, directional extension relationships, and background texture records. Adjacent pixel connectivity is formed by the connections between horizontally adjacent, vertically adjacent, and diagonally adjacent pixels. Directional extension relationships are formed by the extension directions of the elongated dark band to be processed between adjacent pixel positions, adjacent short segment positions, and adjacent local background units.

[0066] Within the same local background unit, a low grayscale narrow band that extends continuously along the connection relationship of adjacent pixels, whose central region is lower than the grayscale of the road surface base on both sides, and whose road surface texture can still be identified on both sides, and whose width falls within the candidate slender structure width boundary registered in the preprocessing configuration data, and whose extension position is not completely covered by the background unavailable state, is registered as a candidate slender structure by the road surface image processing terminal.

[0067] The width boundary of candidate slender structures is jointly formed by pixel-scale records, manually verified fine crack samples, and the minimum visible crack width requirement registered in the detection task data. The minimum extension boundary of candidate slender structures is jointly formed by test images, manually verified crack samples, and coarse aggregate texture samples, and is used to distinguish real slender cracks from individual aggregate dark lines.

[0068] In the urban asphalt pavement example, the image resolution is 4096 pixels by 3000 pixels. Under the same camera, installation height, and pavement type, manually reviewed samples show that fine cracks in the image are mostly one to nine pixels wide. The preprocessing configuration data registers the width boundaries of candidate slender structures as one to nine pixels. Manually reviewed samples also show that individual aggregate dark lines typically do not continuously enter adjacent local background units. The preprocessing configuration data registers the minimum extension boundary as a unidirectional extension trace entering adjacent local background units. These boundaries only apply within the corresponding detection task.

[0069] When a low-grayscale narrow band is located only at the edge of a single aggregate, does not enter the adjacent local background unit, and does not reach the minimum extension boundary, the writing of the particle texture affects the recording and is not directly written to the candidate slender structure.

[0070] When the width of the low-grayscale narrow band reaches the boundary corresponding to the joint influence record, its direction is consistent with the construction joint and contraction joint data, and the boundary is in a continuous and straight shape, it is written into the joint influence record.

[0071] The low grayscale narrow band is located at the boundary between light and dark, with one side being a shadow-covered area and the other side being a normal road surface area. When the central area lacks a narrow grayscale valley relative to the grayscale of the road surface base on both sides, the shadow effect record is written.

[0072] When a low-grayscale narrow band is adjacent to the edge of the lane marking and extends along the edge of the marking, it is written into the lane marking impact record.

[0073] When a low-grayscale narrow band is located at the boundary of a large dark area of ​​oil and water stains, and there is a large grayscale transition inside and outside the boundary, it is written into the oil and water stain impact record.

[0074] Each candidate slender structure simultaneously forms a central dark band, side edge bands, endpoints, and branch points. The central dark band is formed by continuous low-grayscale positions within the candidate slender structure and is used to indicate the main extension position of the candidate slender structure; the central dark band must not be formed separately by the datum edge, seam boundary, or shadow boundary.

[0075] The lateral edge band is formed by the position where the gray level changes from dark to bright on both sides of the central dark band, while still within the width boundary of the candidate slender structure. It is used to represent the local width boundary of the candidate slender structure. The endpoint is formed by the position where the candidate slender structure temporarily stops extending at the boundary of the current local background unit, the edge local background unit, the background to be kernel unit, or the background unavailable unit. The endpoint is only used as the subsequent adjacent verification object and is not used as the actual termination position of the crack in S3.

[0076] A branch point is formed by connecting two or more slender branches that satisfy the width boundary of the candidate slender structure from a candidate slender structure. Only the connection position, branch direction and local background unit where the branch is located are registered at the branch point. Branch deletion and branch connection are not performed in S3.

[0077] The road surface image processing terminal transmits the candidate slender structure number, central dark band, side edge band, endpoint, branch point, corresponding local background unit number and background texture record to S4 to form a fidelity verification record.

[0078] S4. Based on the central dark band, lateral edge band, endpoint adjacent units, branch connection positions, and background texture records of the candidate elongated structure, corresponding states of gray-level valley extension, direction continuity, width preservation, endpoint adjacency, branch preservation, and pseudo-elongated interference are formed, specifically: The road image processing terminal generates a fidelity verification record based on the candidate elongated structure numbers, central dark bands, lateral edge bands, endpoints, branch points, corresponding local background unit numbers, and background texture records generated by S3. The fidelity verification record registers the grayscale valley extension state, directional continuity state, width preservation state, endpoint adjacency state, branch preservation state, and pseudo-elongated interference state. All of these states are derived from the original image grayscale, local background units, acquisition quality records, background texture records, and preprocessing configuration data, and are not based on subjective human judgment as a direct basis for registration.

[0079] The grayscale valley extension state is formed by the grayscale of the central dark band, the grayscale of the neighboring areas on both sides of the central dark band, the continuity of low grayscale narrow bands within adjacent local background units, the distinguishable boundary of the grayscale valley, and the acquisition quality record. The distinguishable boundary of the grayscale valley is jointly registered by the test images of the same detection task, manually verified fine crack samples, and coarse aggregate texture samples, and is used to define the distinguishable grayscale difference that the central dark band should have relative to its neighboring areas on both sides.

[0080] When the central dark band extends continuously along the direction of the candidate slender structure, and the gray level of the central dark band reaches the distinguishable boundary of the gray level valley, and there is still a narrow band of low gray level in the same direction after entering the adjacent local background unit, the gray level valley extension state is registered as the extension is established.

[0081] When the central dark band is interrupted in at least one of the following regions: shadow effect, reflection occlusion, water stain effect, motion blur, but the interruption front point and the interruption back point both point to the same adjacent cell, and there is at least one of the following in the adjacent cell: same-direction low gray level narrow band, side edge band continuation, low gray level weak extension, the gray level valley extension state is registered as extension pending kernel.

[0082] When the central dark band is formed by any one of the following: grain edge, gradation edge, seam boundary, or shadow boundary line, and no narrow side bands are formed on both sides, the gray valley extension state is registered as extension not valid.

[0083] The directional continuity is formed by the direction of the central dark band, the direction of adjacent short segments, the direction of endpoint extension, the direction of adjacent local background units, and the natural bending boundary. The natural bending boundary is jointly registered by manually verified crack samples, joint samples, and test images of the same road segment, and is used to distinguish between the natural bending of cracks, the straight extension of joints, and the extension of shadow boundaries.

[0084] When the directional change of a candidate slender structure between adjacent short segments falls within the natural bending boundary, and the directional change does not align with the joint influence record, marking influence record, or shadow influence record, the directional continuity state is registered as directional continuity.

[0085] When a candidate slender structure is locally interrupted in the shadow, water stain, reflective, or blurred areas, but the extension directions of the endpoints at both ends of the interruption point to the same adjacent unit, the directional continuity state is registered as directional nullable.

[0086] When the direction of the candidate slender structure is consistent with the edge of the marking line, the boundary of the seam, and the boundary of the shadow, and the gray valley extension state is not valid, the direction continuity state is registered as direction discontinuity.

[0087] The width preservation status is formed by the pixel width between side edge bands, the width variation at adjacent positions, the width boundary of candidate slender structures, the seam influence record, and the acquisition quality record. The width boundary of candidate slender structures is jointly registered by pixel-scale records, manually verified fine crack samples, and the minimum visible crack width requirement.

[0088] When the width between the side edge bands falls within the width boundary of the candidate slender structure in adjacent positions, and the side edge bands are not incorporated into the dark texture of adjacent particles, the edge of the marking line, or the boundary of the seam, the width retention state is registered as width retention.

[0089] When the candidate slender structure widens locally in the nucleation area affected by water stains, shadows, or reflective occlusion, but the central dark band still reaches the distinguishable boundary of the gray valley, the width is maintained and registered as the width nucleation state.

[0090] When the width of a candidate slender structure reaches the width boundary corresponding to the joint influence record, and the boundary is continuous and straight with the same direction as the joint influence record, the width is recorded as not conforming.

[0091] The endpoint adjacency state is formed by the candidate slender structure endpoint, the adjacent unit outside the endpoint, the endpoint extension direction, the overlapping area of ​​adjacent images, the image access record, and the acquisition quality record. The adjacent unit outside the endpoint is the next local background unit that the endpoint enters along the extension direction; when the endpoint is located in the edge local background unit, the adjacent unit is the adjacent area that still belongs to the effective area of ​​the road surface in the extension direction of the endpoint.

[0092] If at least one of the following exists within the adjacent unit outside the endpoint: a narrow band of low grayscale in the same direction, a continuation of the side edge band, or a weak extension of low grayscale, and none of the above conditions indicate a clear source of interference, the endpoint adjacency status is registered as established. A weak extension of low grayscale is a continuous low grayscale pixel segment located within the adjacent unit outside the endpoint, below the grayscale range of the corresponding road surface substrate, with a width that does not meet the stable registration conditions for candidate slender structures, and a direction consistent with the extension direction of the endpoint.

[0093] When the adjacent unit outside the endpoint is in one of the following conditions: background unavailable, reflective occlusion waiting for kernel, motion blur, strong shadow transition, the endpoint adjacency state is registered as adjacent for kernel.

[0094] When an image access record is pending sequentially and a reliable adjacent image cannot be accessed, the endpoint adjacency state does not use the overlapping area of ​​adjacent images. It is formed only based on the current image's adjacency unit, and an unavailable adjacent image is written synchronously. When the adjacent unit outside the endpoint is available in the background and there are no unidirectional low-grayscale narrow bands, side edge band continuations, or low-grayscale weak extensions, the endpoint adjacency state is registered as adjacency not established.

[0095] The branch retention status is formed by the branch point, the branch center dark band, the branch side edge band, the branch width, the branch direction, and the connection position between the branch and the main candidate elongated structure. When a branch has at least two of the following: the center dark band extends, the side edge band is identifiable, and the width falls within the width boundary of the candidate elongated structure, and the branch connection position is continuous with the center dark band of the main candidate elongated structure, the branch retention status is registered as branch retention.

[0096] If the branch length does not reach the branch stability boundary registered in the preprocessed configuration data, but is located near the endpoint of the candidate slender structure, and there is low grayscale weak extension in the adjacent background unit, the branch retention state is registered as a branch waiting for kernel.

[0097] When a branch is located at any of the following locations: the edge of a single particle, a burr on a marking line, or the edge of an oil or sewage stain, and it lacks a central dark band connection with the main candidate slender structure, the branch retention status is registered as "branch not retained".

[0098] The branch stability boundary is jointly registered by manually verified fine crack branch samples and coarse aggregate texture samples, and is not used as a fixed parameter across road sections.

[0099] The pseudo-elongated interference state is formed by the background texture record, acquisition quality record, gray valley extension state, directional continuity state, width preservation state, endpoint adjacency state, and branch preservation state.

[0100] When a candidate slender structure is located only at the edge of a single aggregate and the adjacency is not established, it is registered as particle dark texture interference; when it is located at the boundary between light and dark areas and the central dark band does not reach the distinguishable boundary of the gray valley, it is registered as shadow boundary interference; when the width and direction are consistent with the joint influence record, it is registered as joint interference; when it is adjacent to the edge of the lane marking and extends along the edge of the marking, it is registered as lane marking edge interference; when it is located at the edge of a patch of dark area of ​​oil and sewage stains and the gray level is in a patchy transition, it is registered as oil and sewage stain edge interference; when it is located at the boundary of bright reflective light and the reflective occlusion state is local reflective nucleus or reflective occlusion nucleus, it is registered as reflective edge interference; when the direction is consistent with the dragging direction of the entire image and the motion is clear but motion blurry, it is registered as motion blur edge interference.

[0101] If a candidate slender structure simultaneously satisfies at least three of the following conditions: extension, direction continuity, width preservation, and adjacency, and does not meet any of the interference formation conditions, the pseudo-slender interference state is registered as non-interference.

[0102] When the same candidate slender structure simultaneously meets the non-interference condition and any explicit interference formation condition, the road image processing terminal does not register it as non-interference, but registers it as the corresponding interference or pending verification state; when the explicit interference source cannot cover the entire candidate slender structure, only the interference coverage location is written into the interference section, and the remaining locations of the candidate slender structure continue to retain the corresponding fidelity verification record.

[0103] When only some locations in the same candidate slender structure meet the above-mentioned interference formation conditions, the road image processing terminal registers the interference section according to the location that meets the interference formation conditions. The location that does not meet the interference formation conditions continues to retain the fidelity verification record corresponding to the gray valley extension state, direction continuity state, width preservation state, endpoint adjacency state and branch preservation state.

[0104] After the verification record is generated, the road image processing terminal transmits the candidate slender structure number, the value of each state, the interference section, the reason for verification, and the unusable mark of adjacent images to S5 to form the preprocessing execution path.

[0105] S5. Based on the corresponding states, determine the preprocessing execution path according to the priority of prohibition processing, preservation of the original state pending verification, interference suppression, and narrowband fidelity enhancement. Perform corresponding processing on the region where the candidate slender structure is located, and perform endpoint, width, branch, and interference isolation verification on the processing results. If the verification fails, restore the original grayscale of the corresponding position and write it to the rollback record, specifically: The road surface image processing terminal forms a preprocessing execution path based on the fidelity verification record generated by S4, and performs corresponding image processing on the regions where candidate slender structures are located. The preprocessing execution path includes narrowband fidelity enhancement, preserving the original image for verification, interference suppression, and prohibition processing.

[0106] When forming a path, first read the candidate slender structure number, gray valley extension status, directional continuity status, width preservation status, endpoint adjacency status, branch preservation status, pseudo-slender interference status, interference segment, reason for verification, and acquisition quality record, and then restrict it in the following order: prohibition processing, original preservation for verification, interference suppression, and narrowband fidelity enhancement.

[0107] This sequence is used to prevent the direct enhancement of elongated dark bands in areas where exposure is unavailable, background is unavailable, reflections obscure the kernel, endpoints are adjacent to the kernel, and interference sources are not isolated.

[0108] When at least one of the following exists: exposure quality status is exposure unavailable, background processable status is background unavailable, reflection occlusion status is reflection occlusion pending verification and endpoint adjacency status is not registered as adjacency established, scale record is scale pending verification and actual size data needs to be output, the preprocessing execution path is registered as prohibited processing.

[0109] The original image grayscale is not allowed to be read, the reason for the prohibition is not allowed, the corresponding local background unit number and the candidate elongated structure number are not allowed, the original grayscale of the corresponding area is maintained during execution, and enhancement, suppression, connection and deletion are not performed.

[0110] When the prohibited processing area is located outside the endpoint of the candidate slender structure, this area is not considered as the actual termination location of the crack; only the reason for prohibition and a prompt for re-acquisition are recorded.

[0111] If at least one of the following exists: gray valley extension state is extension pending kernel, width preservation state is width pending kernel, endpoint adjacency state is adjacency pending kernel, and branch retention state is branch pending kernel, and no explicit interference is registered in the pseudo-slender interference state, the preprocessing execution path is registered as unchanged pending kernel.

[0112] The original image grayscale, reason for kernel entry, central dark band, side edge band, endpoints, branch points, and local background influence records are preserved as is. During execution, the original grayscale of the narrow band region where the candidate slender structure is located is maintained, the endpoints are kept in their original positions, the branches are kept in their original shapes, and automatic connection is not performed at the interrupted positions.

[0113] The road surface image processing terminal writes the candidate slender structure number, reason for verification, endpoint location, branch point location, adjacent background unit number, and acquisition quality impact type to the layer of slender structures to be verified. If the reason for verification is due to at least one of the following: shadow effect, water stain effect, reflection occlusion, motion blur, sequential verification, or unavailable adjacent images, a re-acquisition prompt and a review prompt are simultaneously written. The layer to be verified is retained as is and is not output as a confirmed crack; it is only used as limiting data for subsequent crack segmentation and manual review.

[0114] If the pseudo-slender interference state is registered as any one of the following: grain dark line interference, shadow boundary interference, seam interference, mark edge interference, oil and water stain edge interference, reflective edge interference, and motion blur edge interference, and the gray valley extension state is not an extension, the preprocessing execution path is registered as interference suppression.

[0115] Interference suppression reads the pseudo-slender interference status, background texture record, pavement base grayscale range, interference section, and the position of adjacent candidate slender structures. During execution, the corresponding interference section is prohibited from entering the crack enhancement range. Particle dark texture interference is limited to the edge range of a single aggregate and does not cross into adjacent candidate slender structures; shadow boundary interference is limited to the light-dark boundary range and does not change the original light-dark relationship on both sides of the shadow; joint interference is written to the joint influence record, retaining the joint position but not writing it to the crack candidate fidelity layer; marking edge interference is written to the marking influence record, retaining the marking edge position but not writing it to the crack candidate fidelity layer; oil and water stain edge interference is written to the oil and water stain influence record, retaining the pollution boundary but not writing it to the crack candidate fidelity layer; reflective edge interference and motion blur edge interference are written to the acquisition quality influence record but not written to the crack candidate fidelity layer.

[0116] When the gray valley extension state is "extended", the width preservation state is "width preserved", the pseudo-slender interference state is "non-interference", and the candidate slender structure does not fall into the prohibited processing region or the original preserved waiting region, the preprocessing execution path is registered as narrowband fidelity enhancement.

[0117] Narrowband fidelity enhancement reads the original image's grayscale, central dark band, side edge band, road base grayscale range, side edge protection boundary, endpoint retention boundary, and width preservation status. During execution, the position of the central dark band remains unchanged, the enhancement range does not exceed the side edge protection boundary, the endpoint position does not exceed the endpoint retention boundary, the side edge band does not merge with adjacent aggregate dark lines, marking edges, and joint boundaries, and the enhanced width does not exceed the side edge band boundary.

[0118] The lateral edge protection boundary is formed by camera calibration data, manual verification of crack samples, and acquisition of image sharpness data; the endpoint retention boundary is formed by manual verification of crack endpoint samples, acquisition of image sharpness data, and test images. In the urban asphalt pavement example, the lateral edge protection boundary is registered as one to two pixels outside the lateral edge strip, and the endpoint retention boundary is registered as one to three pixels outside the endpoint. This value is only valid within the same camera, the same installation height, the same pavement type, and the same inspection task.

[0119] When the same candidate slender structure has both non-interfering and interfering sections, the road surface image processing terminal performs processing on a section-by-section basis.

[0120] When non-interference sections meet the narrowband fidelity enhancement condition, narrowband fidelity enhancement is only performed on non-interference sections; interference suppression is performed on interference sections; conflict pending records are written at the boundary between the two types of sections, and the original grayscale is maintained. When a candidate slender structure passes through a joint, marking edge, shadow boundary, or oil / stain edge, if the central dark band satisfies the extension condition before and after the crossing, and the endpoint adjacency state is adjacency established, the crossing position is not directly deleted, but written into the conflict pending record, and the original grayscale is maintained. This processing is used to avoid the accidental deletion of real fine cracks when they pass through interference areas, and also to avoid interference boundaries being enhanced into cracks.

[0121] After preprocessing, the road image processing terminal outputs a high-fidelity preprocessed image, a crack candidate high-fidelity layer, a thin structure layer to be kerneled, and a prohibited enhancement region layer. The high-fidelity preprocessed image saves the pixel results after narrowband high-fidelity enhancement, preservation of the kernel, interference suppression, and prohibition processing.

[0122] The crack candidate fidelity layer registers the candidate slender structure number, central dark band, side edge band, endpoints, branch points, and corresponding local background unit number for inclusion in the narrowband fidelity enhancement path. The slender structure layer to be kerneled registers the candidate slender structure number, kernel reason, endpoint position, branch point position, and adjacent background unit number for inclusion in the original kernel path and conflict kernel records.

[0123] The prohibited enhancement region layer registers the region boundaries, interference type, reason for prohibition, and corresponding local background unit number for areas entering the interference suppression and prohibition processing path. The crack candidate fidelity layer is used for subsequent crack image segmentation and geometric reading; the elongated structure to be verified layer is used for subsequent verification and re-acquisition prompts; the prohibited enhancement region layer is used to restrict subsequent processing from using pseudo-elongated structures as crack input.

[0124] Before outputting the image, the road surface image processing terminal performs endpoint consistency verification, width consistency verification, branch consistency verification, and interference isolation verification. The endpoint consistency verification reads the original endpoints, preprocessed endpoints, endpoint retention boundaries, and endpoint adjacency status. When a preprocessed endpoint crosses the endpoint retention boundary, the original grayscale near the endpoint is restored, and the endpoint rollback record is written.

[0125] Width consistency verification reads the original side edge band, the pre-processed side edge band, the side edge protection boundary, and the width retention status; if at least one of the following occurs after pre-processing: width expansion, side edge contraction, or interference intrusion, the original grayscale of the corresponding position is restored, and the width rollback record is written.

[0126] Branch consistency verification reads the branch point, branch retention status, and preprocessed branch shape; when a retained branch position is deleted, the original grayscale near the branch is restored and written to the branch rollback record. Interference isolation verification reads the crack candidate fidelity layer, the slender structure layer to be verified, and the prohibited enhancement area layer; when joints, marking edges, shadow boundaries, oil and water stain edges, reflective edges, and motion blur edges are written to the crack candidate fidelity layer, the corresponding areas are rewritten as conflict to be verified, the original grayscale is restored, and the interference rollback record is written.

[0127] Each rollback record includes the original image number, candidate slender structure number, rollback location, rollback reason, and processing paths before and after the rollback, which are used for subsequent review of preprocessing configuration data.

[0128] The boundaries of distinguishable gray-level valleys, natural bend boundaries, light-processable boundaries, vehicle speed boundaries, branch stability boundaries, width boundaries corresponding to joint influence records, re-acquisition conditions, and discontinuous allowable boundaries are all jointly registered using camera calibration data, test images, manually verified samples, road inspection task sheets, and on-site acquisition plans for the same inspection task. These boundaries are not considered universal fixed parameters across cameras, road sections, road surface types, or acquisition platforms. The aforementioned boundaries are formed separately from real crack samples and corresponding interference samples in the manually verified samples. During registration, the sample number, image number, camera number, road surface type, pixel scale record, and manually verified conclusion are simultaneously saved. Isolated gray-level differences in a single image are not used as fixed thresholds.

[0129] In an urban asphalt pavement inspection task, the pavement image processing terminal reads the original pavement image captured by the industrial camera on the inspection vehicle. The image resolution is 4096 pixels by 3000 pixels, corresponding to camera number CAM-A01 and lens number LEN-12, and the acquisition platform is marked as fixed acquisition by the inspection vehicle.

[0130] The camera calibration data records the camera number, lens number, image resolution, and installation height, covering the acquisition conditions of the image. The pixel scale record is registered as available. The detection task data registers the current detection object as the first lane on the right, the acquisition direction as the mileage increasing direction, and the minimum visible crack width requirement as one to nine pixels in the image.

[0131] The preprocessing configuration data is formed from test images and manually reviewed samples of the same detection task. Local background units are registered as 64 pixels by 64 pixels, candidate slender structure width boundaries are registered as one to nine pixels, side edge protection boundaries are registered as one to two pixels outside the side edge band, and endpoint preservation boundaries are registered as one to three pixels outside the endpoint.

[0132] The example parameters above are only used to illustrate the state formation and path execution process in this embodiment. In actual application, they are jointly registered by the corresponding detection task data, camera calibration data, test images and manual verification samples, and are not used as fixed thresholds for cross-device, cross-road segment and cross-road surface types.

[0133] After delineating the effective area of ​​the road surface in the image, the road surface image processing terminal divides the effective area into multiple local background units. Within each local background unit located in the center of the image, there exists a narrow low-grayscale band. The width of this low-grayscale band is between three and five pixels. The central area is lower than the grayscale of the road surface base on both sides, while the aggregate texture can still be discerned on both sides. Furthermore, this low-grayscale band continues to extend in the same direction after entering adjacent local background units.

[0134] The road surface image processing terminal registers the low grayscale narrow band as a candidate elongated structure, and within this candidate elongated structure, registers a central dark band, side edge bands, two endpoints, and one branch point. The fidelity verification record shows that the central dark band of the candidate elongated structure extends in the same direction within adjacent local background units, the width of the side edge band remains within the width boundary of the candidate elongated structure, a low grayscale narrow band in the same direction exists outside the endpoints, and the branch position still has a discernible central dark band and side edge bands. Based on this, the grayscale valley extension state is registered as extension established, the directional continuity state is registered as directional continuity, the width preservation state is registered as width preservation, the endpoint adjacency state is registered as adjacency established, the branch retention state is registered as branch retention, and the pseudo-elongated interference state is registered as non-interference.

[0135] The road image processing terminal registers the preprocessing execution path of the candidate slender structure as narrow band fidelity enhancement and performs enhancement within the side edge protection boundary. The position of the central dark band remains unchanged, the endpoint does not cross the endpoint preservation boundary, and the enhanced side edge band does not merge into the dark texture of the adjacent aggregate.

[0136] Near the edge of the lane marking in the same image, there is also a straight dark band with a width of six to eight pixels. This dark band is adjacent to the edge of the bright marking and runs in the same direction as the edge of the marking. The central area does not form a narrow gray valley relative to the gray level of the road surface base on both sides.

[0137] The road image processing terminal registers this dark band as a road marking influence, the corresponding pseudo-elongated interference state as a road marking edge interference, and the preprocessing execution path as interference suppression. This dark band is retained in the prohibited enhancement area layer and is not written to the crack candidate fidelity layer.

[0138] The water stain on the right side of the image forms a curved dark band. This dark band is located at the boundary of a large dark area. The area inside and outside the boundary shows a large gray-scale transition. The endpoint adjacency state is not valid. The corresponding pseudo-slender interference state is registered as oil and water stain edge interference. The preprocessing execution path is registered as interference suppression.

[0139] After preprocessing, the road image processing terminal outputs a high-fidelity preprocessed image, a crack candidate high-fidelity layer, a layer of slender structures to be verified, and a layer of prohibited enhancement areas. The crack candidate high-fidelity layer registers the numbers of candidate slender structures in the center of the image, the central dark band, the side edge band, the endpoints, the branch points, and the corresponding local background unit numbers; the prohibited enhancement area layer registers the interference areas of road marking edges and oil and water stain edges; the layer of slender structures to be verified does not include the edges of road markings and oil and water stains as confirmed cracks.

[0140] Before output, the road image processing terminal performs a consistency check on the candidate slender structure in the middle of the image. The check results show that the endpoints of the candidate slender structure are still within the endpoint preservation boundary, the enhanced width does not exceed the side edge boundary, and the branch points and corresponding branch shapes are preserved; at the same time, the interference areas of the marking edge and the interference areas of the oil and sewage stain edge are not written into the crack candidate fidelity layer.

[0141] The above content is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for preprocessing pavement images to preserve the fidelity of slender crack structures, characterized in that, Performed by the road surface image processing terminal, including: S1. Read the original road surface image, acquisition log, camera calibration data, detection task data and preprocessing configuration data to form an image access record and pixel scale record; S2. Based on the camera's effective field of view, distortion boundary, lane range, and road surface texture, delineate the effective area of ​​the road surface and form acquisition quality records and local background units; S3. Based on the base gray level, adjacent pixel connection relationship, directional extension relationship and background texture record in the local background unit, a candidate slender structure with central dark band, side edge band, endpoint and branch point is formed. S4. Based on the central dark band, lateral edge band, endpoint adjacent units, branch connection positions, and background texture records of the candidate slender structure, corresponding states of gray-level valley extension, directional continuity, width preservation, endpoint adjacency, branch retention, and pseudo-slender interference are formed. The gray-level valley extension state is formed based on the gray level of the central dark band, the gray levels of the two neighboring areas, the continuation of the low gray-level narrow band, and the discernible boundary of the gray-level valley. The directional continuity state is formed based on the continuation direction of the low gray-level narrow band, the natural bending boundary, and the influence of directional texture records. The width preservation state is formed based on the spacing of the lateral edge bands, the width boundary of the candidate slender structure, and the influence of seams records. The endpoint adjacency state is formed based on the same-direction low gray-level narrow band outside the endpoint, the continuation of the lateral edge band, and the low gray-level weak extension. The low grayscale weak extension is defined as a continuous low grayscale pixel segment located in the adjacent unit outside the endpoint, below the grayscale range of the corresponding road surface base, with a width that does not meet the stable registration conditions for candidate slender structures, and a direction consistent with the extension direction of the endpoint; the branch retention state is formed based on the branch point, branch width, branch direction, and branch connection position; the pseudo-slender interference state is registered based on the background texture record, acquisition quality record, grayscale valley extension state, direction continuity state, width retention state, endpoint adjacency state, and branch retention state as one of the following: grain dark texture interference, shadow boundary interference, joint interference, marking edge interference, oil and water stain edge interference, reflective edge interference, motion blur edge interference, and non-interference. S5. Based on the corresponding states, determine the preprocessing execution path according to the priority of prohibiting processing, preserving as is pending verification, interference suppression, and narrowband fidelity enhancement. Perform corresponding processing on the region where the candidate slender structure is located, and perform endpoint, width, branch, and interference isolation verification on the processing results. If the verification fails, restore the original grayscale of the corresponding position and write it into the rollback record.

2. The method for preprocessing pavement images for preserving the fidelity of slender crack structures according to claim 1, characterized in that, In S1: The image access record number is formed by the detection task number, road segment number, lane number, acquisition date, camera number, and image sequence number in a fixed order. The image access record registers the original image storage location, acquisition log source, camera calibration data number, and preprocessing configuration data number.

3. The method for preprocessing pavement images for preserving the fidelity of slender crack structures according to claim 1, characterized in that, In S1: The pixel scale record is formed based on at least one of the following: camera number, lens number, image resolution, acquisition platform mark, and camera mounting height mark and shooting distance mark corresponding to the acquisition platform; The registration scale is valid when all of the above content falls within the effective range of the camera calibration data; if any inconsistency exists, the registration scale needs to be verified.

4. The method for preprocessing pavement images for preserving the fidelity of slender crack structures according to claim 1, characterized in that, In S2: The effective area of ​​the road surface is the pixel area that is located within the effective field of view of the camera, does not fall outside the distortion boundary, is within the lane range registered for the detection task, and where the difference between the grayscale of the road surface base and the local texture falls within the grayscale processing boundary registered in the preprocessing configuration data.

5. The method for preprocessing pavement images for preserving the fidelity of slender crack structures according to claim 1, characterized in that, In S2: The acquisition quality record registers the exposure quality status, the supplementary lighting consistency status, the motion clarity status, and the reflection occlusion status. The above statuses are formed based on the original image grayscale, camera exposure time, supplementary lighting working mark, acquisition vehicle speed mark, road surface particle edge morphology, highlight area, and acquisition remarks, respectively.

6. The method for preprocessing pavement images for preserving the fidelity of slender crack structures according to claim 1, characterized in that, In S3: Background texture records include the grayscale range of the road surface base, the impact of particle texture, the impact of directional texture, the impact of shadow, the impact of joints, the impact of road markings, the impact of oil and water stains, and the background processability status.

7. The method for preprocessing pavement images for preserving the fidelity of slender crack structures according to claim 1, characterized in that, In S3: The candidate elongated structure is formed by a low grayscale narrow band that extends continuously along the connection relationship of adjacent pixels. The central region of the low grayscale narrow band is lower than the grayscale of the road surface base on both sides, while the road surface texture can still be identified on both sides. The width falls within the width boundary of the candidate elongated structure, and the extension position is not completely covered by the background unusable state.

8. The method for preprocessing pavement images for preserving the fidelity of slender crack structures according to claim 1, characterized in that, In S4: The above gray valley extension status is registered as one of the following: extension established, extension pending verification, and extension not established; directional continuity status is registered as one of the following: directional continuity, direction pending verification, and directional discontinuity; width preservation status is registered as one of the following: width preservation, width pending verification, and width not conforming; endpoint adjacency status is registered as one of the following: adjacency established, adjacency pending verification, and adjacency not established. The branch retention status is registered as one of the following: branch retained, branch pending approval, or branch not retained.

9. A method for preprocessing pavement images for preserving the fidelity of slender crack structures according to claim 1, characterized in that, In S4: When the source of interference in the same candidate slender structure only covers part of the location, the interference segment is registered according to the covered location, and the corresponding fidelity verification record is retained for the uncovered location.

10. A method for preprocessing pavement images for preserving the fidelity of slender crack structures according to claim 1, characterized in that, In S5: The road surface image processing terminal forms a preprocessing execution path in the order of prohibition processing, original preservation pending verification, interference suppression, and narrowband fidelity enhancement; When at least one of the following conditions exists: exposure quality status is exposure unavailable, background processable status is background unavailable, reflection occlusion status is reflection occlusion pending verification and endpoint adjacency status is not registered as adjacency established, pixel scale record is scale pending verification and actual size data needs to be output, registration prohibits processing. If at least one of the following exists: extension to be verified, width to be verified, adjacent to be verified, and branch to be verified, and the pseudo-slender interference state is not registered as any of the following: grain dark line interference, shadow boundary interference, seam interference, mark edge interference, oil and water stain edge interference, reflective edge interference, and motion blur edge interference, then the original state is retained for verification. If the pseudo-slender interference state is registered as any one of the following: grain dark line interference, shadow boundary interference, seam interference, mark edge interference, oil and water stain edge interference, reflective edge interference, and motion blur edge interference, and the gray valley extension state is not registered as extension established, then interference suppression is registered. When the gray valley extension state is extended, the width maintenance state is width maintenance, the pseudo-slender interference state is non-interference, and it does not fall into the prohibited processing area or the original preservation waiting area, register narrowband fidelity enhancement; The output includes a high-fidelity preprocessed image, a crack candidate high-fidelity layer, a thin structure layer to be verified, and a layer of prohibited enhancement areas. When the endpoint consistency verification, width consistency verification, branch consistency verification, and interference isolation verification fail, the original grayscale of the corresponding position is restored and written to the corresponding rollback record.

Images