Image recognition-based vehicle damage automatic evaluation system

Abstract

This invention discloses an automatic vehicle damage assessment system based on image recognition. The system includes: a coordinate grid module: acquiring a set of vehicle images to be assessed, performing vehicle localization, body component segmentation, viewpoint annotation, and quality screening on each image, and constructing a component parameter coordinate grid for the corresponding body components based on component seams, rib lines, headlight boundary tangent points, wheel arch boundary tangent points, and handlebar positioning points; and a damage characterization module: extracting candidate damage zones within the component parameter coordinate grid, refining and completing each candidate damage zone to obtain a damage skeleton line, a skeleton node sequence, and a cross-sectional width sequence generated along the normal of the damage skeleton line. This invention improves the accuracy, stability, and consistency of automatic vehicle damage assessment through damage cross-image normalization and the collaborative processing of separating old and new damage.

Description

Technical Field

[0001] This application relates to the field of vehicle damage assessment technology, and in particular to an automatic vehicle damage assessment system based on image recognition. Background Technology

[0002] With the application of image recognition and deep learning technologies in the field of auto insurance claims, existing technologies are now able to identify vehicle damage and assist in damage assessment based on vehicle images. Image-based vehicle damage assessment solutions, multi-view vehicle damage recognition solutions, and non-simultaneous accident damage recognition solutions have emerged.

[0003] However, existing methods still mainly focus on damage detection, component identification, and multi-view information fusion, and do not adequately consider some problems that are easily overlooked in actual automated damage assessment: Firstly, the same physical damage often presents different outlines and areas in distant views, close-up views, and different shooting angles. If only image-level deduplication or result-level fusion is performed, the same damage may be counted repeatedly, resulting in inflated repair items and damage assessment amounts.

[0004] Secondly, historical paint touch-ups, sheet metal repairs, old putty layers, or abnormal textures after disassembly and reassembly may spatially overlap with new scratches and dents caused by the current accident. While existing solutions can identify damage from different accidents, they lack an effective mechanism for quantifying only the actual new damage in the current accident when historical repair marks are mixed with newly added damage. Therefore, it is necessary to propose an automatic vehicle damage assessment system based on image recognition to improve the accuracy and consistency of automatic damage assessment results. Summary of the Invention

[0005] To address the aforementioned problems, embodiments of the present invention provide an image recognition-based automatic vehicle damage assessment system, the system comprising: Coordinate Grid Module: Acquires the image set of the vehicle to be evaluated, performs vehicle localization, body component segmentation, viewpoint annotation and quality screening on each image, and constructs the component parameter coordinate grid of the corresponding body component based on the component seam line, rib line, headlight boundary tangent point, wheel arch boundary tangent point and handle positioning point; Damage characterization module: Extracts candidate damage bands within the component parameter coordinate grid, performs refinement and fracture completion on each candidate damage band, and obtains damage skeleton line, skeleton node sequence, and cross-sectional width sequence generated along the normal of the damage skeleton line; Cross-graph unification module: Based on the projection position of the damage skeleton line in the component parameter coordinate grid, the order preservation relationship of the skeleton node sequence, the continuity relationship of the cross-sectional width sequence, and the closure relationship of adjacent boundaries, cross-graph skeleton splicing is performed on candidate damage bands belonging to the same body component and mapped to the same component parameter coordinate grid to obtain damage skeleton objects. New damage identification module: Extracts local low-frequency undulation layer and high-frequency coating residual layer around the damaged skeleton object, generates repair trace constraint contour based on the grinding texture and disassembly offset zone in the local low-frequency undulation layer and the coating transition zone in the high-frequency coating residual layer, and determines the new damage domain based on the overlap area between the repair trace constraint contour and the damaged skeleton object. Damage assessment output module: Generates a damage assessment unit based on the damage skeleton object, the newly added damage domain, the corresponding body parts, and the repair operation mapping relationship, and outputs the automatic vehicle damage assessment result based on the damage assessment unit.

[0006] Furthermore, the method for constructing the component parameter coordinate mesh includes: generating a first parameter line along the seam line of the body component, generating a second parameter line along the rib line of the body component, using the tangent point of the lamp body boundary, the tangent point of the wheel arch boundary, and the handle positioning point as anchor points, performing component interpolation on the first parameter line and the second parameter line to form a parameterized mesh unit that corresponds one-to-one with the corresponding body component.

[0007] Furthermore, when extracting candidate damage bands, the damage response region is first determined within the component parameter coordinate grid, and then the damage response region is segmented into strip constraints to obtain the candidate damage bands. When refining the candidate damage bands, the central connected path is extracted as the damage skeleton line, and the damage skeleton line, the skeleton node sequence, and the cross-sectional width sequence are written into the candidate damage band description structure.

[0008] Furthermore, when performing cross-graph skeleton stitching, the candidate damage zone description structure is used as the matching unit. A stitching relationship matrix is ​​constructed based on the overlapping relationship of the projection interval, the order preservation relationship of the skeleton node sequence, the continuity relationship of the cross-sectional width sequence, and the consistency relationship of the boundary closure direction. Multiple candidate damage zone description structures are merged into the damage skeleton object according to the stitching relationship matrix.

[0009] Furthermore, when there are intersecting damage skeleton lines in the candidate damage band description structure, the bifurcation nodes corresponding to the intersection positions are extracted. Based on the valley values ​​of the cross-sectional width on both sides of the bifurcation nodes, the curvature of the skeleton turning point, and the local enclosing boundary, the intersecting damage skeleton lines are split and recombined, and the split and recombined results are rewritten into the candidate damage band description structure and the splicing relationship matrix.

[0010] Furthermore, when extracting the local low-frequency undulation layer and the high-frequency coating residual layer, the coverage area of ​​the damaged skeleton object in the component parameter coordinate grid is used as a local window. The image within the local window is decomposed into two layers, and the decomposed local low-frequency undulation layer and the high-frequency coating residual layer are mapped back to the component parameter coordinate grid.

[0011] Furthermore, when generating the repair trace constraint profile, the grinding texture and disassembly offset zone are extracted from the local low-frequency undulation layer, the coating transition zone is extracted from the high-frequency coating residual layer, and the repair trace constraint profile is generated based on the closed bounding boundary of the grinding texture, the disassembly offset zone and the coating transition zone.

[0012] Furthermore, when determining the newly added damage domain, the overlapping area is first established based on the repair trace constraint contour and the damage skeleton object. Then, the fresh edge band is constructed along both sides of the damage skeleton object. The overlapping area is then segmented based on the substrate exposure response, paint layer fracture response, and specular shadow reversal relationship. The area that is connected to the fresh edge band and located within the coverage area of ​​the damage skeleton object is determined as the newly added damage domain.

[0013] Furthermore, when generating the damage assessment unit, the vehicle body component identifier, the damage skeleton object identifier, the boundary information of the newly added damage domain, and the maintenance operation label are written into the same damage assessment unit. The N damage assessment units are merged according to the coverage relationship in the component parameter coordinate grid, the overlap relationship of the damage skeleton object, and the boundary relationship of the newly added damage domain, where N is an integer greater than 1.

[0014] The image recognition-based automatic vehicle damage assessment system also includes a pseudo-edge verification module, which is used to extract the highlight inversion boundary and the attached texture boundary in the overlapping area, remove boundary segments that are discontinuous with the damaged skeleton object, and correct newly added damage domains.

[0015] The technical effects and advantages of the image recognition-based automatic vehicle damage assessment system provided by this invention are as follows: This invention improves the accuracy, stability, and consistency of automatic vehicle damage assessment by combining damage cross-image normalization with the separation of old and new damage. It utilizes a component parameter coordinate grid to uniformly map candidate damage zones across different images and merges the same physical damage using damage skeleton objects, reducing the problem of repeated identification and counting of the same damage under multiple viewpoints and scene conditions. By jointly extracting the constraint contour of repair traces through a local low-frequency fluctuation layer and a high-frequency coating residual layer, and identifying newly added damage domains within the overlap area with the damage skeleton object, it can distinguish between historical repair traces and newly added damage. Through the unified generation and merging of damage assessment units, the assessment results have better structured representation and output consistency. Attached Figure Description

[0016] Figure 1 This is a connection diagram of the image recognition-based automatic vehicle damage assessment system in Example 1; Figure 2 This is a flowchart of the cross-graph normalization and new injury separation collaborative processing method in Example 1; Figure 3 This is a connection diagram of the image recognition-based automatic vehicle damage assessment system in Example 2. Detailed Implementation

[0017] 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.

[0018] Example 1: Please see Figure 1 As shown, embodiments of the present invention provide an automatic vehicle damage assessment system based on image recognition, the system comprising: Coordinate Grid Module: Acquires the image set of the vehicle to be evaluated, performs vehicle localization, body component segmentation, viewpoint annotation and quality screening on each image, and constructs the component parameter coordinate grid of the corresponding body component based on the component seam line, rib line, headlight boundary tangent point, wheel arch boundary tangent point and handle positioning point; Damage characterization module: Extracts candidate damage bands within the component parameter coordinate grid, performs refinement and fracture completion on each candidate damage band, and obtains damage skeleton line, skeleton node sequence, and cross-sectional width sequence generated along the normal of the damage skeleton line; Cross-graph unification module: Based on the projection position of the damage skeleton line in the component parameter coordinate grid, the order preservation relationship of the skeleton node sequence, the continuity relationship of the cross-sectional width sequence, and the closure relationship of adjacent boundaries, cross-graph skeleton splicing is performed on candidate damage bands belonging to the same body component and mapped to the same component parameter coordinate grid to obtain damage skeleton objects. New damage identification module: Extracts local low-frequency undulation layer and high-frequency coating residual layer around the damaged skeleton object, generates repair trace constraint contour based on the grinding texture and disassembly offset zone in the local low-frequency undulation layer and the coating transition zone in the high-frequency coating residual layer, and determines the new damage domain based on the overlap area between the repair trace constraint contour and the damaged skeleton object. Damage assessment output module: Generates a damage assessment unit based on the damage skeleton object, the newly added damage domain, the corresponding body parts, and the repair operation mapping relationship, and outputs the automatic vehicle damage assessment result based on the damage assessment unit.

[0019] In this embodiment, the component parameter coordinate grid is constructed by a coordinate grid module. This module takes the previously obtained set of images of the vehicle to be evaluated as input and combines the vehicle positioning results, body component segmentation results, and viewpoint annotation results to establish a unified positional description basis for each body component. The set of images of the vehicle to be evaluated refers to the set of images corresponding to the same vehicle to be evaluated. The body component segmentation results refer to the component areas such as doors, fenders, bumpers, hoods, and trunk lids separated from the images. The viewpoint annotation results refer to the markings of the image shooting orientation, which are used to define the correspondence of the same body component in different images.

[0020] In the specific construction process, stable structural features are first extracted in each body component area. The body component seam line refers to the dividing line between adjacent body components, the rib line refers to the shape fold line on the body surface, the lamp body boundary tangent point refers to the tangent position between the outer contour of the lamp body and the boundary of the target body component, the wheel arch boundary tangent point refers to the tangent position between the wheel arch contour and the boundary of the target body component, and the handle positioning point refers to the center position of the door handle area. The above structural features are used to define the direction and anchoring position of the internal coordinates of the body component.

[0021] After extracting the above structural features, a first parameter line is generated along the seam line of the body component, and a second parameter line is generated along the rib line of the body component. Using the tangent point of the lamp body boundary, the tangent point of the wheel arch boundary, and the handle positioning point as anchor points, component interpolation is performed on the first parameter line and the second parameter line to form a parametric mesh unit. The first parameter line is used to characterize the expansion trend of the component boundary, the second parameter line is used to characterize the extension trend of the internal shape of the component, and the anchor points are used to constrain the intersection position of the parameter lines and limit mesh drift. The area enclosed by adjacent mesh lines constitutes a parametric mesh unit, which is used to express the relative positional relationship and adjacency relationship of each region within the same body component.

[0022] The component parameter coordinate grid description structure output by the coordinate grid module includes: vehicle body component identifier, first parameter line set, second parameter line set, anchor point set, and adjacency relationship of parameterized grid cells; when the subsequent damage characterization module extracts candidate damage bands within the component parameter coordinate grid, it uses this description structure as a unified reference, so that the positions of candidate damage bands in different images can be mapped to the same parameter space, thereby providing a basis for subsequent cross-image skeleton stitching.

[0023] For example: Using the side corner area of ​​the bumper as the target vehicle body component, the front image is first extracted from the vehicle positioning results. Then, the bumper area is obtained based on the vehicle body component segmentation results. Since there is a clear seam between the bumper and the fender, this seam can be used as the boundary guide for the first parameter line. The styling ribs on the bumper surface serve as the internal guide for the second parameter line. The tangent point of the headlight boundary is located at the adjacent position of the headlight and the bumper, and the tangent point of the wheel arch boundary is located at the transition position between the wheel arch and the bumper. After the coordinate grid module performs component interpolation based on these features, a set of stable parametric grid cells can be formed in the side corner area of ​​the bumper. Even if the appearance of the bumper changes in another oblique side view image, as long as the above structural features are still identified, the component parameter coordinate grid corresponding to the previous image can still be constructed. In this way, scratches or cracks that subsequently appear in the side corner area of ​​the bumper can be projected onto the same parameter space for unified expression and subsequent processing.

[0024] In this embodiment, the extraction of candidate damage bands is performed by the damage characterization module, and the aforementioned component parameter coordinate grid is used as a unified reference. The damage response area refers to the area within the component parameter coordinate grid that exhibits edge abruptness, texture destruction, abnormal brightness, or interruption of morphological continuity relative to the surrounding vehicle surface. This area is not directly used as a candidate damage band, but is used as the input range for subsequent strip constraint segmentation to avoid directly including reflections, shadows, and local stains into the damage characterization object.

[0025] In practice, the vehicle body component image is first mapped to the corresponding component parameter coordinate grid. Surface anomalies are detected segment by segment within the grid cells to obtain the damage response region. Subsequently, banded constraint segmentation is performed on the damage response region. Banded constraint segmentation refers to the continuous filtering and boundary convergence of the damage response region according to the morphological characteristics of damage that typically extends continuously in a certain direction and has a relatively limited lateral width. This transforms the segmentation result from a blocky abnormal region into a banded region with an extending direction, and the resulting result is the candidate damage band. Through this processing, discrete noise regions on the same vehicle body component can be excluded, while retaining the main damage channel corresponding to the edges of scratches, cracks, and dents.

[0026] After obtaining the candidate damage zones, they are refined, and the central connected path is extracted as the damage skeleton line. The damage skeleton line refers to the connected path located at the center of the candidate damage zone and reflecting the main extension trend of the damage. The skeleton node sequence refers to the set of nodes recorded sequentially along the extension direction of the damage skeleton line, used to characterize the direction change, turning points and branching relationships of the skeleton line. In order to enable subsequent cross-graph normalization to utilize both positional and width features, after extracting the damage skeleton line, corresponding sampling is performed on both sides of the candidate damage zone along the normal of the damage skeleton line to form a cross-sectional width sequence. The cross-sectional width sequence refers to the bandwidth change sequence recorded segment by segment along the extension direction of the damage skeleton line, used to characterize the width fluctuation of the damage zone at different locations.

[0027] After completing the above processing, the damage skeleton line, skeleton node sequence, and cross-sectional width sequence are written into the candidate damage band description structure. The candidate damage band description structure is used to uniformly save the morphological characterization results of the same candidate damage band, so that the subsequent cross-graph normalization module can directly call the description structure to compare and stitch candidate damage bands belonging to the same body part.

[0028] For example: In the outer panel area of ​​a car door, a scratch may appear as an irregularly shaped strip-shaped abnormal area in the original image; after the damage response region is extracted, the abnormal area is confined within the corresponding component parameter coordinate grid; after strip constraint segmentation, candidate damage bands extending along the length direction of the car door can be obtained; after further refinement, the damage skeleton line located at the center of the candidate damage band can be extracted, and the corresponding skeleton node sequence and cross-sectional width sequence can be formed, which are then written into the candidate damage band description structure for subsequent steps.

[0029] In this embodiment, cross-image skeleton stitching is performed by the cross-image unification module. This module takes the aforementioned candidate damage band description structure as input and no longer directly uses the pixel regions in the original image as comparison objects. The candidate damage band description structure refers to the structured description data established for a single candidate damage band, which includes at least the projection interval of the candidate damage band in the component parameter coordinate grid, the corresponding damage skeleton line, the skeleton node sequence, the cross-sectional width sequence, and the boundary orientation information. By using the candidate damage band description structure as a matching unit, the damage characterization results of the same vehicle body component in different images can be converted into a unified data object, providing a foundation for subsequent cross-image stitching.

[0030] In practice, pairwise comparisons are first performed on multiple candidate damage band description structures mapped to the same component parameter coordinate grid. The overlapping relationship of the projection intervals refers to whether the coverage intervals of each candidate damage band in the component parameter coordinate grid are in the same or adjacent position ranges. The order preservation relationship of the skeleton node sequence refers to whether the skeleton nodes in the description structures of two candidate damage bands maintain the same sequential arrangement trend along the damage extension direction. The continuity relationship of the cross-sectional width sequence refers to whether the bandwidth changes of the two candidate damage bands at corresponding positions have continuous connection characteristics. The consistency of the boundary closure direction refers to whether the local enclosure direction enclosed by the two sides of the candidate damage band is consistent. The above relationships are not used in isolation, but together serve as the basis for judging whether two candidate damage bands originate from the same physical damage.

[0031] After comparing the candidate damage band description structures, a splicing relationship matrix is ​​constructed. This matrix records whether any two candidate damage band description structures meet the splicing conditions and the degree to which they do so. Subsequently, based on the splicing relationship matrix, multiple candidate damage band description structures that are interconnected, continuous, and positionally corresponding are merged to form a single damage skeleton object. The damage skeleton object is a unified representation object obtained by merging candidate damage bands for the same physical damage in multiple images. It is used to eliminate the problem of the same damage being repeatedly identified and counted in different images and to provide a unique carrier object for the subsequent generation of new damage domain identification and damage assessment units.

[0032] For example: A scratch on the same car door appears in both the frontal side view and the oblique side view. After preprocessing, the two images form corresponding candidate damage band description structures. If the projection intervals of the two images in the component parameter coordinate grid are connected, the extension order of the skeleton node sequence is consistent, the change trend of the cross-sectional width sequence is continuous, and the boundary closure direction is consistent, then the cross-image unification module will write the two images into the same splicing relationship matrix connected region and merge them into the same damage skeleton object. Subsequent steps will then be carried out around this damage skeleton object, instead of processing the candidate damage bands in the two images separately.

[0033] In this embodiment, the splitting and recombination process is performed by the cross-graph unification module before or during the cross-graph skeleton splicing. Cross-damage skeleton lines refer to two or more damage skeleton lines that intersect, overlap, or share a local path within the same candidate damage zone description structure. Bifurcation nodes refer to the skeleton connection nodes corresponding to the intersection positions, used to characterize the damage path branching or merging at that point.

[0034] In practice, the skeleton connectivity is first detected in the candidate damage zone description structure. When multiple skeleton lines intersect in the same area within a single candidate damage zone, the bifurcation node corresponding to the intersection position is extracted. Subsequently, with the bifurcation node as the center, the local cross-sectional width change, skeleton orientation change, and boundary enclosure state are extracted along each branch direction. The cross-sectional width valley value refers to the local contraction position in the cross-sectional width sequence on both sides of the bifurcation node, which is used to identify the transition boundary between different damage paths. The skeleton turning curvature refers to the degree of turning change of the skeleton line near the bifurcation node, which is used to distinguish between continuous extension paths and temporary intersection paths. The local enclosure boundary refers to the local closed boundary formed by the candidate damage zone in the intersection area, which is used to limit the actual coverage area corresponding to each branch.

[0035] After obtaining the above information, the cross-damage skeleton line is split and recombined. Specifically, the priority splitting position is determined according to the valley value of the cross-section width, the main extension direction of each branch is determined according to the skeleton turning curvature, and then the skeleton segments originally shared in the cross area are split into corresponding independent skeleton segments in combination with the local enclosing boundary. Skeleton segments with continuous extension relationship are then recombined. After splitting and recombining, the overlapping damage paths in the original candidate damage zone description structure are decomposed into characterization results with clearer boundaries and more stable directions.

[0036] After the splitting and recombination is completed, the updated damage skeleton lines, skeleton node sequences and cross-sectional width sequences are rewritten into the candidate damage band description structure, and the splicing relationship matrix is ​​updated simultaneously. This ensures that subsequent cross-graph skeleton splicing is based on the corrected matching units, thus avoiding the misidentification of two adjacent damages as a single damage skeleton object due to local intersection.

[0037] For example: In the corner area of ​​a bumper, a transverse scratch and a diagonal crack may intersect in the image; if they are directly involved in cross-image skeleton stitching, they are easily misclassified; by extracting the bifurcation node at the intersection and combining the valley value of the cross-sectional width on both sides of the bifurcation node, the curvature of the skeleton turning point, and the local enclosing boundary to perform splitting and recombination, the transverse scratch and the diagonal crack can be written into the updated candidate damage zone description structure, thereby improving the accuracy of subsequent damage skeleton object normalization.

[0038] In this embodiment, the extraction of the local low-frequency fluctuation layer and the high-frequency coating residual layer is performed by the new damage identification module. Its input is the damage skeleton object obtained in the previous step and the corresponding component parameter coordinate grid. The coverage area of ​​the damage skeleton object in the component parameter coordinate grid refers to the grid area occupied by the skeleton line, boundary projection and associated region corresponding to the damage skeleton object in the component parameter coordinate grid. Using this coverage area as a local window, the subsequent analysis can be limited to the local area of ​​the component directly related to the target damage, avoiding interference from irrelevant areas to the identification of repair traces.

[0039] In practice, the image region within a local window is first extracted based on the position range of the damaged skeleton object in the component parameter coordinate grid. Then, a two-layer decomposition is performed on the image within this local window. The local low-frequency undulation layer is an image layer that reflects the slowly changing shape of the vehicle body surface and is used to characterize continuous surface morphology such as sheet metal undulations, grinding transitions, and local unevenness. The high-frequency coating residual layer is a layer of detail changes that is retained after removing the slowly changing background and is used to characterize coating anomalies such as paint texture interruptions, paint repair edges, repair lines, and fine cracks. Through the two-layer decomposition, the surface change information and coating change information in the same local window can be separated, so that the extraction of subsequent grinding textures, disassembly offsets, and coating transitions can be established on the corresponding data layers.

[0040] After completing the two-layer decomposition, the obtained local low-frequency undulation layer and high-frequency coating residual layer are mapped back to the component parametric coordinate grid. Mapping back means that the two-layer decomposition results are re-mapped to the parameterized grid cells covered by the local window, so that each grid cell not only retains the original positional relationship, but also carries the corresponding low-frequency undulation information and high-frequency residual information. In this way, when generating the repair trace constraint contour in the future, it is possible to directly determine which areas show abnormal continuous surface undulation and which areas show abnormal coating texture transition within the same component parametric coordinate grid, thereby maintaining a consistent coordinate basis with the aforementioned damage skeleton object and the subsequent identification of newly added damage domains.

[0041] For example: When there is a damaged skeleton object at the bottom of the car door, after performing a two-layer decomposition using the coverage area of ​​the damaged skeleton object in the component parameter coordinate grid as a local window, the local low-frequency undulation layer can highlight the slow concave transition of the door panel surface, and the high-frequency coating residual layer can highlight the texture changes formed by the paint repair boundary; after mapping back to the component parameter coordinate grid, the two can be used as the basic data for the subsequent generation of repair trace constraint contours.

[0042] In this embodiment, the generation of the repair trace constraint contour is performed by the new damage identification module, and the aforementioned local low-frequency undulation layer and high-frequency coating residual layer mapped back to the component parameter coordinate grid are used as inputs; grinding texture refers to an abnormal area that is continuously distributed in a band in the local low-frequency undulation layer, reflecting that the surface undulation tends to be smoother after grinding but still has a directional transition; disassembly and assembly offset zone refers to a band-shaped abnormal area in the local low-frequency undulation layer caused by boundary misalignment, gap offset or local contour discontinuity formed after component disassembly and assembly; coating transition zone refers to a band-shaped area in the high-frequency coating residual layer that shows abrupt changes in paint texture, transition of paint repair edges or discontinuity in surface detail response; the above three types of band-shaped areas respectively characterize the residual features of different levels in the historical repair process, among which grinding texture and disassembly and assembly offset zone focus on surface morphology changes, and coating transition zone focuses on paint texture changes.

[0043] In practice, firstly, continuous undulation bands are searched along the area surrounding the damaged skeleton object within the local low-frequency undulation layer. Regions that satisfy the requirements of extension continuity and local directional stability are extracted as grinding texture bands. Simultaneously, local misalignment regions near component boundaries, seams, or connection points are extracted as disassembly / assembly offset bands. Then, texture transition regions distributed around the damaged skeleton object are extracted within the high-frequency coating residual layer to obtain coating transition bands. After the above extraction is completed, the grinding texture bands, disassembly / assembly offset bands, and coating transition bands are projected onto the same component parameter coordinate grid. Based on the spatial connection and enclosing relationship of the three, a closed enclosing boundary is determined. The closed enclosing boundary refers to the closed contour formed by connecting the above-mentioned strip-shaped regions end to end or enclosing each other. It is used to define the outer boundary of historical repair traces within the current local area of ​​the component. Subsequently, this closed enclosing boundary is determined as the repair trace constraint contour for subsequent identification of new damage domains in the overlapping area with the damaged skeleton object.

[0044] For example: In a suspected area of ​​repeated repair on the outer panel of a car door, the local low-frequency undulation layer can reflect the grinding texture distributed along the surface of the door panel and the disassembly offset zone near the door seam; the high-frequency coating residual layer can further reflect the coating transition zone formed by the paint touch-up edge; after mapping the three to the same component parameter coordinate grid, a closed bounding boundary can be formed around the area, thereby obtaining the corresponding repair trace constraint contour.

[0045] like Figure 2 As shown, in this embodiment, the determination of the newly added damage domain is performed by the new damage identification module, and the aforementioned repair trace constraint contour and damage skeleton object are used as inputs. The overlapping area refers to the local area formed by the intersection of the area defined by the repair trace constraint contour and the coverage area of ​​the damage skeleton object in the component parameter coordinate grid. The purpose of setting this overlapping area is to limit the judgment range to the common area where "historical repair traces may exist and the current damage skeleton object has actually passed through", and avoid including irrelevant areas outside the repair trace constraint contour in the new damage identification.

[0046] In practice, firstly, an overlapping area is established between the repair trace constraint contour and the damaged skeleton object. Then, fresh edge bands are constructed along both sides of the damaged skeleton object. Fresh edge bands refer to strip-shaped areas distributed along the extension direction of the damaged skeleton object that can reflect the newly formed boundary characteristics of the current damage. When constructing the fresh edge band, the damaged skeleton object is taken as the center, and boundary areas with continuous edge response, adjacent positions and consistent directions are searched on both sides of it to maintain a correspondence between the fresh edge band and the damaged skeleton object.

[0047] After a fresh edge band is formed, the overlapping area is segmented. The exposed substrate response refers to the local response formed when the substrate is exposed after the coating is damaged; the paint layer fracture response refers to the edge response formed after the continuity of the coating is broken; the highlight-shadow reversal relationship refers to the light and dark reversal correspondence caused by the change in surface reflection state due to newly formed scratches, cracks or dents. Based on the above responses and relationships, the new damage identification module delineates the boundaries of each local area within the overlapping area and identifies the area connected to the fresh edge band and located within the coverage of the damage skeleton object as the newly added damage domain. Through this process, the part of the area covered by historical repair traces that truly belongs to the newly added damage can be separated from the overlapping area.

[0048] Furthermore, to ensure that the segmentation of newly added damage domains within the overlapping region has a feasible quantitative basis, the new damage identification module can identify any candidate sub-region within the overlapping region. Calculate the new judgment value The methods include: ; In the formula, Indicates candidate subregion The continuity coefficient with the damaged skeleton object is used to characterize whether the candidate sub-region is continuously distributed along the extension direction of the damaged skeleton object; Show candidate sub-regions The boundary; Indicates the fresh edge zone; Indicates the boundary position Fresh edge response intensity at the location; Indicates the length of the boundary element; This represents a stable term, used to avoid the boundary length denominator being zero; Indicates candidate subregion The average value of the substrate exposure response within the range is used to characterize the degree of substrate exposure after coating damage; Indicates candidate subregion The average value of the paint layer fracture response within the range is used to characterize the degree of damage to the continuity of the paint layer; Indicates candidate subregion The average value of the highlight-shadow reversal response within the area is used to characterize the change in reflection state caused by the newly formed damage edge; Indicates candidate subregion The average value of the constraint response of the repair traces within the region is used to characterize the degree to which the candidate subregion is affected by historical repair traces; Indicates candidate subregion The average value of the historical repair texture is used to characterize the continuity of textures left from old paint touch-ups, old sanding, or old disassembly / reassembly. , , , , and This represents the weighting coefficient, used to adjust the influence of fresh edge response, bare substrate response, paint layer fracture response, specular shadow reversal response, repair trace constraint response, and historical repair texture continuity response on the new judgment.

[0049] For example: If there are already traces of paint repair on the outer panel of the car door, and the current scratch falls into the repair area; at this time, the contour of the repair trace intersects with the damaged skeleton object to form an overlapping area; in this overlapping area, only the area that simultaneously satisfies the substrate exposure response or the paint layer fracture response and remains connected to the fresh edge bands on both sides of the damaged skeleton object is identified as the new damage domain. This can avoid mistakenly including the original paint repair boundary as a whole in the scope of this damage.

[0050] In this embodiment, the generation and merging of damage assessment units are performed by the damage assessment output module. A damage assessment unit refers to structured assessment data built around a single damage skeleton object, which carries the location, boundary range, and repair results of the damage on the corresponding vehicle body component. Specifically, during generation, the vehicle body component identifier, damage skeleton object identifier, boundary information of the newly added damage domain, and repair operation label are written into the same damage assessment unit. The vehicle body component identifier is used to indicate the component to which the damage belongs; the damage skeleton object identifier is used to indicate the unique damage object corresponding to the assessment unit; the boundary information of the newly added damage domain is used to limit the actual scope of the damage; and the repair operation label is used to characterize the treatment category corresponding to the newly added damage domain.

[0051] Boundary information is not a simple outline in the original image, but a boundary description result mapped to the component parameter coordinate grid. By writing the boundary information of the newly added damage domain into the damage assessment unit, the assessment results formed in different images can be compared on the same coordinate basis. After generating multiple damage assessment units, the damage assessment output module merges the multiple damage assessment units according to the coverage relationship in the component parameter coordinate grid, the coincidence relationship of the damage skeleton object, and the boundary relationship of the newly added damage domain. The coverage relationship refers to the relationship of the occupied area of ​​each damage assessment unit in the component parameter coordinate grid. The coincidence relationship refers to whether the projection of the damage skeleton object corresponding to different damage assessment units in the grid belongs to the same damage path. The boundary relationship refers to whether there is inclusion, connection, or local overlap between the boundaries of each newly added damage domain.

[0052] In the specific merging process, candidate damage assessment units are first screened under the same vehicle body component identifier. Then, damage assessment units with the same damage skeleton object identifier or overlapping damage skeleton object projections are aggregated. When multiple damage assessment units have continuous coverage areas in the component parameter coordinate grid and the boundaries of the newly added damage domains have a connection relationship, the multiple damage assessment units are merged into the same output unit. After merging, duplicate assessment results can be avoided because the same damage skeleton object comes from multiple images.

[0053] For example: A scratch on the same car door generates two damage assessment units in the frontal side view and the oblique side view respectively. Although the two come from different images, their body parts are identified in the same way. The projections of the damage skeleton objects in the component parameter coordinate grid overlap with each other, and the boundaries of the newly added damage domains also form a continuous connection relationship. Then the damage assessment output module merges the two damage assessment units into the same assessment result for subsequent automatic vehicle damage assessment result output.

[0054] Example 2: like Figure 3 As shown, this embodiment further improves the design based on embodiment one. The difference is that in actual operation of embodiment one, when there is strong reflection, water film adhesion, local stain residue, or transparent protective film edge on the vehicle surface, the overlapping area formed by the repair trace constraint contour and the damaged skeleton object is prone to generating pseudo-edge responses that are confused with fresh edge bands. This causes the newly added damage domain to locally expand outward, cross-regional concatenation, or be mistakenly merged into the historical repair area. It fails to further distinguish between the real newly added damage boundary and the pseudo-boundary of light reflection, pseudo-boundary of attachment, and pseudo-boundary of film edge. Based on this, the automatic vehicle damage assessment system based on image recognition also includes: a pseudo-edge verification module, which is used to extract the bright inversion boundary and the attachment texture boundary in the overlapping area, remove the boundary segments that are discontinuous with the damaged skeleton object, and correct the newly added damage domain.

[0055] In this embodiment, the pseudo-edge verification module receives the repair trace constraint contour, damage skeleton object, and newly added damage domain output from the new damage identification stage. The overlapping area refers to the region formed by the intersection of the repair trace constraint contour and the coverage of the damage skeleton object. The pseudo-edge verification module does not perform general de-reflection processing on the entire image, but only performs boundary verification within the overlapping area to avoid introducing irrelevant areas into the newly added damage domain correction process.

[0056] In practice, the highlight reversal boundary and the attached texture boundary are first extracted in the overlapping area. The highlight reversal boundary refers to the boundary of sudden change in brightness caused by strong reflection, water film or local bright reflection. The attached texture boundary refers to the texture transition boundary formed by stain residue, transparent film edge or surface attachment. Then, the above boundaries are checked for continuity with the damaged skeleton object: any boundary segment that is only connected to the highlight area, only extends along the attached texture distribution, or is inconsistent with the extension direction of the damaged skeleton object is judged as a pseudo edge boundary segment.

[0057] After identifying the pseudo-edge boundary segment, the pseudo-edge verification module removes the pseudo-edge boundary segment from the boundary of the newly added damage domain and recloses the removed boundary to obtain the corrected newly added damage domain. This process can prevent reflections, water films, stains or membrane edges from being mistakenly included in the newly added damage range. In this embodiment, the verification is limited to the specific object of "the boundary of the newly added damage domain in the overlapping area".

[0058] For example: If there are traces of previous paint touch-up on the outer panel of the car door, after the scratch falls into this area, the actual scratch boundary and the pseudo edge formed by the water film highlight may appear simultaneously in the overlapping area; after the pseudo edge verification module extracts the highlight inverted boundary, if the boundary segment does not continue to extend along the damage skeleton object, it is removed from the new damage domain boundary, and only the boundary segment that continuously corresponds to the damage skeleton object is retained. In this way, the corrected new damage domain is closer to the actual damage range caused by this accident.

[0059] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0060] The above description is merely a preferred embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present application, based on the technical solution and concept of the present application, should be covered within the scope of protection of the present application.

Claims

1. An automatic vehicle damage assessment system based on image recognition, characterized in that, The system includes: Coordinate Grid Module: Acquires the image set of the vehicle to be evaluated, performs vehicle localization, body component segmentation, viewpoint annotation and quality screening on each image, and constructs the component parameter coordinate grid of the corresponding body component based on the component seam line, rib line, headlight boundary tangent point, wheel arch boundary tangent point and handle positioning point; Damage characterization module: Extracts candidate damage bands within the component parameter coordinate grid, performs refinement and fracture completion on each candidate damage band, and obtains damage skeleton line, skeleton node sequence, and cross-sectional width sequence generated along the normal of the damage skeleton line; Cross-graph unification module: Based on the projection position of the damage skeleton line in the component parameter coordinate grid, the order preservation relationship of the skeleton node sequence, the continuity relationship of the cross-sectional width sequence, and the closure relationship of adjacent boundaries, cross-graph skeleton splicing is performed on candidate damage bands belonging to the same body component and mapped to the same component parameter coordinate grid to obtain damage skeleton objects. New damage identification module: Extracts local low-frequency undulation layer and high-frequency coating residual layer around the damaged skeleton object, generates repair trace constraint contour based on the grinding texture and disassembly offset zone in the local low-frequency undulation layer and the coating transition zone in the high-frequency coating residual layer, and determines the new damage domain based on the overlap area between the repair trace constraint contour and the damaged skeleton object. Damage assessment output module: Generates a damage assessment unit based on the damage skeleton object, the newly added damage domain, the corresponding body parts, and the repair operation mapping relationship, and outputs the automatic vehicle damage assessment result based on the damage assessment unit.

2. The automatic vehicle damage assessment system based on image recognition according to claim 1, characterized in that, The method for constructing the component parameter coordinate mesh includes: generating a first parameter line along the seam line of the body component, generating a second parameter line along the rib line of the body component, using the tangent point of the lamp body boundary, the tangent point of the wheel arch boundary, and the handle positioning point as anchor points, performing component interpolation on the first parameter line and the second parameter line to form a parameterized mesh unit that corresponds one-to-one with the corresponding body component.

3. The automatic vehicle damage assessment system based on image recognition according to claim 2, characterized in that, When extracting candidate damage bands, the damage response region is first determined within the component parameter coordinate grid, and then the damage response region is segmented into strip constraints to obtain the candidate damage bands. When refining the candidate damage bands, the central connected path is extracted as the damage skeleton line, and the damage skeleton line, the skeleton node sequence, and the cross-sectional width sequence are written into the candidate damage band description structure.

4. The image recognition-based automatic vehicle damage assessment system according to claim 3, characterized in that, When performing cross-graph skeleton stitching, the candidate damage zone description structure is used as the matching unit. A stitching relationship matrix is ​​constructed based on the overlapping relationship of the projection interval, the order preservation relationship of the skeleton node sequence, the continuity relationship of the cross-sectional width sequence, and the consistency relationship of the boundary closure direction. Multiple candidate damage zone description structures are merged into the damage skeleton object according to the stitching relationship matrix.

5. The image recognition-based automatic vehicle damage assessment system according to claim 4, characterized in that, When there are intersecting damage skeleton lines in the candidate damage band description structure, the bifurcation nodes corresponding to the intersection positions are extracted. Based on the valley values ​​of the cross-sectional width on both sides of the bifurcation nodes, the curvature of the skeleton turning point, and the local enclosing boundary, the intersecting damage skeleton lines are split and recombined. The results of the split and recombination are then rewritten into the candidate damage band description structure and the splicing relationship matrix.

6. The automatic vehicle damage assessment system based on image recognition according to claim 5, characterized in that, When extracting the local low-frequency undulation layer and the high-frequency coating residual layer, the coverage area of ​​the damaged skeleton object in the component parameter coordinate grid is used as a local window. The image within the local window is decomposed into two layers, and the decomposed local low-frequency undulation layer and high-frequency coating residual layer are mapped back to the component parameter coordinate grid.

7. The image recognition-based automatic vehicle damage assessment system according to claim 6, characterized in that, When generating the repair trace constraint profile, the grinding texture and disassembly offset texture are extracted from the local low-frequency undulation layer, the coating transition texture is extracted from the high-frequency coating residual layer, and the repair trace constraint profile is generated based on the closed bounding boundary of the grinding texture, the disassembly offset texture and the coating transition texture.

8. The automatic vehicle damage assessment system based on image recognition according to claim 7, characterized in that, When determining the new damage domain, the overlapping area is first established based on the repair trace constraint contour and the damage skeleton object. Then, the fresh edge band is constructed along both sides of the damage skeleton object. The overlapping area is then segmented based on the substrate exposure response, paint layer fracture response, and specular shadow reversal relationship. The area that is connected to the fresh edge band and is located within the coverage area of ​​the damage skeleton object is determined as the new damage domain.

9. The automatic vehicle damage assessment system based on image recognition according to claim 8, characterized in that, When generating the damage assessment unit, the vehicle body component identifier, the damage skeleton object identifier, the boundary information of the newly added damage domain, and the maintenance operation label are written into the same damage assessment unit. The N damage assessment units are merged according to the coverage relationship in the component parameter coordinate grid, the overlap relationship of the damage skeleton object, and the boundary relationship of the newly added damage domain, where N is an integer greater than 1.

10. The automatic vehicle damage assessment system based on image recognition according to claim 1, characterized in that, Also includes: The pseudo-edge verification module is used to extract the highlight inversion boundary and the attached texture boundary in the overlapping area, remove the boundary segments that are discontinuous with the damaged skeleton object, and correct the newly added damage domain.

Images