A visual recognition-based automobile appearance defect detection system

Abstract

The present application relates to the field of industrial vision technology, and especially relates to a kind of automobile appearance defect detection system based on visual identification, including image acquisition module, candidate body extraction module, feature extraction module and defect determination module.Image acquisition module applies non-contact pulse air flow disturbance in the process of continuously collecting images to the edge region to be detected, obtains the edge timing image group before disturbance, in disturbance and after disturbance removal;Candidate body extraction module extracts edge attached candidate body based on edge reference contour, and defines parent edge segment, connecting part, overhanging part and free end region;Feature extraction module takes parent edge segment as local motion reference, extracts local differential motion feature and unloading recovery motion feature;Defect determination module determines candidate body as sheet residual burr according to the state of connecting part attachment, overhanging part local flexure, free end relative migration and reverse recovery after unloading.This application realizes residual burr classification detection through time sequence dynamic identification.

Description

Technical Field

[0001] This invention relates to the field of industrial vision technology, and in particular to a vehicle exterior defect detection system based on visual recognition. Background Technology

[0002] Machine vision inspection technology acquires surface images of objects through image acquisition devices and uses image processing methods to analyze image information such as contours, edges, textures, grayscale, and spatial positions in the target area, thereby recognizing the surface condition of the object. With the increasing automation of automotive parts manufacturing and online quality inspection, machine vision inspection has been gradually applied to the appearance inspection of automotive structural components.

[0003] For automotive structural parts with cut edges, the appearance near the cut edges needs to be inspected after the corresponding processing is completed. Taking an integrated die-cast aluminum alloy automotive structural part as an example, after the gate, overflow groove, and other areas are removed, there may be protruding targets on the edge located outside the predetermined cut edge contour near the cut edge. By acquiring images of the cut edge area and comparing the actual edge contour with the predetermined cut edge contour, image information such as the position, shape, size, and spatial relationship between the protruding targets and the cut edge can be obtained.

[0004] In existing machine vision appearance inspection processes, image processing methods such as image segmentation, edge extraction, contour matching, feature analysis, and target recognition can be used to extract abnormal image regions within the area to be detected. Based on the positional relationship of the abnormal region relative to the baseline contour and its corresponding morphological features, protruding targets with sharp edges can be detected. For continuously acquired images, inter-frame matching or target association can also be used to obtain information on the positional changes of the detected target in different images.

[0005] In actual images of the cut edge area of ​​automotive structural components, edge targets of different natures may have similar image representations. For example, sheet-like residual burrs, cut edge debris located near the cut edge, and some rigid cut edge residues may all appear as protruding areas adjacent to the actual cut edge and extending outward from the reference contour in the image. The above-mentioned edge targets may overlap in terms of image features such as contour shape, protrusion direction, area, and adjacency relationship with the cut edge. This makes it difficult to distinguish between different types of protruding edge targets when directly determining the specific type based on the appearance of the edge targets in the acquired images.

[0006] Therefore, how to distinguish between sheet-like residual flash with similar appearance and other protruding edge targets in the image of the cut edge area of ​​automotive structural parts is a technical problem that needs to be solved in this field. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a vision-based automotive exterior defect detection system that uses time-series dynamic recognition to classify and detect residual burrs.

[0008] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows: A vision-based vehicle exterior defect detection system, comprising: The image acquisition module continuously acquires images of the edge to be detected and applies non-contact pulsed airflow disturbance during the acquisition process to obtain a time-series image group of the edge, which includes multiple frames of recovered images before, during and after the disturbance is removed. The candidate body extraction module, based on the tangent reference contour, determines the edge protrusion targets that connect to the outer side of the actual parent body tangent area as edge-attached candidate bodies, and defines the parent body tangent segment, connecting part, extension part and free end area of ​​the candidate body; The feature extraction module uses the cut edge segment of the parent body as a local motion reference to calculate the relative motion data of the connecting part, the extended part and the free end region across frames, and extracts the local differential motion features of the candidate body under airflow disturbance and the unloading and recovery motion features after the airflow is removed. The defect determination module makes a determination based on local differential motion characteristics and unloading recovery motion characteristics. When the following conditions are met: during airflow disturbance, the connecting part of the candidate body remains attached, the extended part produces local deflection constrained by the connecting part, and the free end undergoes relative migration; and after the airflow is removed, the free end produces recovery migration in the opposite direction of the disturbance, and the extended part returns to the front contour of the disturbance, the candidate body is determined to be a sheet-like residual flash.

[0009] Furthermore, the application direction of the non-contact pulsed airflow is configured to graze laterally along the cut edge or at a preset angle to the cut edge region to be detected, generating lateral disturbance force on the edge attachment candidate; the dynamic pressure parameter of the pulsed airflow is configured to be within a preset dynamic pressure range, the lower limit of the preset dynamic pressure range is greater than the first critical pressure corresponding to the preset pixel-level deformation displacement generated by the sheet-like residual flash, and the upper limit of the preset dynamic pressure range is less than the second critical pressure corresponding to the preset pixel-level deformation displacement generated by the normal rigid edge structure.

[0010] Furthermore, the acquisition process of the edge-cutting time-series image group includes: before the pulsed airflow establishes a substantial disturbance, acquiring at least one frame as an initial edge-cutting image to establish a static reference under no external force; during the peak window period of the continuous action of the pulsed airflow, continuously acquiring multiple frames of disturbance edge-cutting images to extract the maximum flexural displacement and deformation trajectory of the extended portion and free end region under airflow stress; after the pulsed airflow stops and the external force is removed, continuously acquiring multiple frames of recovery edge-cutting images, the duration of continuous acquisition being configured to cover at least one effective recovery stage of the free end region returning from the disturbance displacement state to the pre-disturbance position state, and continuous acquisition until it is determined that the offset of the free end region shows a decreasing trend and the outline of the extended portion shows a regression trend, fully recording the gradual recovery process of the free end region returning to the initial outline.

[0011] Furthermore, the process of determining the edge attachment candidate includes: registering and comparing the actual contour in the initial tangent image with the tangent reference contour; extracting the convex image region located outside the tangent reference contour and having a continuous pixel connection with the actual parent tangent region as the edge attachment candidate; and defining the local edge segment in the actual parent tangent region that directly intersects with the edge attachment candidate in space and forms a continuous structure as the parent tangent segment corresponding to the candidate.

[0012] Further, the process of defining the parent cut edge segment, connecting portion, extension portion, and free end region of the candidate body includes: establishing a spatial topological extension direction or distance field outward from the parent cut edge segment based on the extracted edge attachment candidate body; defining the root transition region of the edge attachment candidate body that is close to the parent cut edge segment and serves as the starting point for outward extension as the connecting portion; defining the main body region of the edge attachment candidate body that grows outward from the cut edge reference contour along the spatial topological extension direction as the extension portion, starting from the connecting portion; and defining the end boundary region of the edge attachment candidate body located on the side of the extension portion away from the connecting portion as the free end region.

[0013] Further, the extraction process of the local differential motion features and unloading recovery motion features includes: extracting the contour features of the parent tangent segment as a local coordinate reference between frames of the tangent time sequence image group, and performing relative position compensation and spatial alignment on each frame image; in the airflow disturbance stage image after position compensation and spatial alignment, determining the first relative displacement of the connecting part relative to the local coordinate reference, and the second relative displacement of the free end region relative to the connecting part, and determining the state in which the amplitude of the first relative displacement is less than a first set threshold and the amplitude of the second relative displacement is greater than the amplitude of the first relative displacement, forming a differential motion relationship between the connecting part and the free end region, as the local differential motion feature; in the recovery stage image after position compensation and spatial alignment, continuously extracting the temporal change data of the free end region and the extension relative to the local coordinate reference, and determining the recovery displacement vector of the free end region in the opposite direction to the displacement vector in the airflow disturbance stage, and the remaining offset amplitude of the free end region relative to the position before the disturbance decreasing frame by frame, while the temporal dynamic change of the extension contour approaching the corresponding contour in the initial tangent image frame by frame, as the unloading recovery motion feature.

[0014] Furthermore, the process in which the connecting portion of the candidate remains attached during airflow disturbance, the extended portion exhibits local deflection constrained by the connecting portion, and the free end undergoes relative migration includes: extracting the coordinate offset or contour overlap of the connecting portion relative to the parent body's cut edge segment in a cross-frame image; determining that the connecting portion remains attached when the coordinate offset is less than a preset holding threshold or the contour overlap is greater than a preset ratio; extracting the contour curvature data or morphological skeleton angle data of the extended portion in a cross-frame image; determining that the extended portion exhibits local deflection constrained by the connecting portion when the curvature change or angle change during airflow is greater than a preset deformation threshold; calculating the displacement vector of the free end region relative to the connecting portion; determining that the free end undergoes relative migration when the magnitude of the displacement vector is greater than a preset migration threshold.

[0015] Furthermore, the determination process for the extension to regress to the pre-disturbance contour includes: calculating the recovery displacement vector of the free end region in consecutive multi-frame recovery images; when the angle between the recovery displacement vector and the displacement vector obtained during the airflow disturbance stage falls within a preset reverse angle range, it is determined that the free end has generated a recovery migration opposite to the disturbance direction; extracting the contour features of the extension in consecutive multi-frame recovery images; when the image matching degree between the extension contour and the corresponding contour in the initial tangent image increases frame by frame, or the spatial distance between the contour boundaries decreases frame by frame, it is determined that the extension regresses to the pre-disturbance contour.

[0016] This invention relates to the field of industrial vision technology, specifically to a vision-based automotive exterior defect detection system. The system includes an image acquisition module, a candidate body extraction module, a feature extraction module, and a defect determination module. The image acquisition module applies non-contact pulsed airflow disturbance during continuous image acquisition of the cut edge area to be detected, obtaining a time-series image set of the cut edge before, during, and after the disturbance. The candidate body extraction module extracts edge-attached candidate bodies based on the cut edge reference contour and defines the parent cut edge segment, connecting part, extension part, and free end region. The feature extraction module uses the parent cut edge segment as a local motion reference to extract local differential motion features and unloading recovery motion features. The defect determination module determines the candidate body as sheet-like residual flash based on the attachment of the connecting part, the local deflection of the extension part, the relative migration of the free end, and the reverse recovery state after unloading. The specific advantages of this invention compared to the prior art are as follows: 1. By establishing a local coordinate reference based on the contour features of the parent body's cut edge segment, and performing relative position compensation and spatial alignment on the cut edge time sequence image, the movement of the connecting part relative to the parent body's cut edge segment and the movement of the free end relative to the connecting part can be characterized separately under the corresponding reference relationship, forming a differential displacement relationship between the connecting part and the free end. This can reflect the movement distribution state of the edge attachment candidate in different regions during airflow disturbance, providing image feature basis for subsequent identification of the state where the connecting end is constrained and the outer region undergoes relative movement; 2. By extracting the coordinate offset or contour overlap relationship of the connecting part relative to the tangent segment of the parent body, the contour curvature or skeleton angle change of the extended part, and the displacement vector of the free end relative to the connecting part, the connection retention, local deflection, and free end migration can be characterized by the image changes of different spatial regions. The three are established in the same airflow disturbance process, thereby obtaining the combined dynamic features that reflect the constraint state of the root of the candidate body and the stress deformation state of the extended area, providing a basis for the state determination of sheet-like edge targets. 3. By calculating the recovery displacement vector of the free end region in the continuous recovery images after the pulsed airflow is removed, and combining the continuous change relationship between the protrusion profile and the pre-disturbance profile, the free end migration generated during the disturbance stage can be established in time with the recovery process after the external force is removed. By utilizing the inverse relationship between the recovery migration direction and the disturbance migration direction, as well as the trend of the protrusion profile regressing to the initial state, a time-series image feature characterizing the recovery behavior of the candidate body after unloading can be formed, providing a basis for further determination of its connection state and deformation properties. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of the vision recognition-based automotive exterior defect detection system of the present invention; Figure 2 This is a schematic diagram of the visual recognition-based automotive exterior defect detection system of the present invention; Figure 3 This is a schematic diagram of the process for extracting local differential motion features and unloading and restoring motion features according to the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be further described in a non-limiting manner below with reference to the accompanying drawings and specific embodiments.

[0019] like Figures 1-3 As shown, the present invention provides a vehicle exterior defect detection system based on visual recognition, the technical solution of which is as follows: A vision-based vehicle exterior defect detection system, specifically as follows: Figure 1 and Figure 2 As shown, it includes: The image acquisition module continuously acquires images of the edge to be detected and applies non-contact pulsed airflow disturbance during the acquisition process to obtain a time-series image group of the edge, which includes multiple frames of recovered images before, during and after the disturbance is removed. The candidate body extraction module, based on the tangent reference contour, determines the edge protrusion targets that connect to the outer side of the actual parent body tangent area as edge-attached candidate bodies, and defines the parent body tangent segment, connecting part, extension part and free end area of ​​the candidate body; The feature extraction module uses the cut edge segment of the parent body as a local motion reference to calculate the relative motion data of the connecting part, the extended part and the free end region across frames, and extracts the local differential motion features of the candidate body under airflow disturbance and the unloading and recovery motion features after the airflow is removed. The defect determination module makes a determination based on local differential motion characteristics and unloading recovery motion characteristics. When the following conditions are met: during airflow disturbance, the connecting part of the candidate body remains attached, the extended part produces local deflection constrained by the connecting part, and the free end undergoes relative migration; and after the airflow is removed, the free end produces recovery migration in the opposite direction of the disturbance, and the extended part returns to the front contour of the disturbance, the candidate body is determined to be a sheet-like residual flash.

[0020] Example 1: Furthermore, the application direction of the non-contact pulsed airflow is configured to graze laterally along the cut edge or at a preset angle to the cut edge region to be detected, generating lateral disturbance force on the edge attachment candidate; the dynamic pressure parameter of the pulsed airflow is configured to be within a preset dynamic pressure range, the lower limit of the preset dynamic pressure range is greater than the first critical pressure corresponding to the preset pixel-level deformation displacement generated by the sheet-like residual flash, and the upper limit of the preset dynamic pressure range is less than the second critical pressure corresponding to the preset pixel-level deformation displacement generated by the normal rigid edge structure.

[0021] Furthermore, the non-contact pulsed airflow is used to apply transient lateral loads to edge attachment candidates located outside the cut edge without mechanical contact with the cut edge area to be detected. In specific implementation, an airflow jet assembly can be set on one side of the image acquisition device. The airflow jet assembly includes a compressed air source, a pressure regulating unit, a pulse control valve, and a nozzle facing the cut edge area to be detected. The pulse control valve controls the opening and closing of the nozzle according to the image acquisition sequence, so that the nozzle forms a pulsed airflow within a predetermined time window. The outlet direction of the nozzle can sweep along the transverse direction of the cut edge, even if the airflow direction has a component perpendicular to the local extension direction of the cut edge; or the nozzle axis can form a preset tilt angle relative to the local surface where the cut edge area to be detected is located, so that the airflow forms an action component on the surface of the edge attachment candidate facing the outside or above the cut edge reference contour.

[0022] In this embodiment, the intensity of the pulsed airflow is controlled by the gas dynamic pressure at the location of the cut edge to be detected. The gas dynamic pressure can be obtained by pre-calibrating the nozzle outlet pressure, the distance between the nozzle and the cut edge to be detected, and the corresponding airflow velocity. Alternatively, it can be calibrated directly or indirectly by a pressure detection element during the equipment commissioning phase. In the case of calibration using airflow velocity, the corresponding dynamic pressure can be determined based on the gas density and airflow velocity, and the nozzle output state under different gas supply pressures can be converted into the equivalent dynamic pressure at the location to be detected, thereby establishing the correspondence between the pulse control parameters and the actual gas dynamic pressure acting on the cut edge area.

[0023] To determine the preset dynamic pressure range, before the equipment is put into testing, a graded loading calibration can be performed using calibration samples corresponding to the material, edge thickness, and surface condition of the structural component to be tested. First, a calibration area with sheet-like residual flash is selected. While keeping the nozzle position, injection direction, and image acquisition parameters constant, the dynamic pressure of the pulsed airflow is gradually increased from low to high according to a preset step size. Images are continuously acquired during each level of dynamic pressure. A pre-set pixel displacement is used as the visually recognizable deformation condition; for example, the free end region relative to the connecting part is not less than [a certain value]. The relative displacement of each pixel is determined to achieve a preset pixel-level deformation displacement; where... Based on the camera resolution, imaging magnification, and edge positioning error, the first critical air pressure P1 is determined when the extended part or free end region of the sheet-like residual flash first generates the air flow pressure corresponding to the preset pixel-level deformation displacement, while the connecting part remains at the corresponding mother body cutting edge section.

[0024] Further selecting the structural component body area and the normal rigid edge structure as rigid reference areas, the airflow pressure is further increased under the same jet direction and image acquisition conditions. The positional changes of the structural component body contour or the normal rigid edge structure in the images before and after the disturbance are calculated respectively. When the structural component body area or the normal rigid edge structure first shows a contour offset or local deformation that reaches the preset pixel-level deformation displacement, the corresponding airflow pressure is determined as the second critical air pressure P2. Thus, the first critical air pressure P1 characterizes the minimum airflow effect required to produce a local flexural motion that can be recognized by the current vision system for the sheet-like residual flash edge, and the second critical air pressure P2 characterizes the airflow effect corresponding to the same order of magnitude of recognizable displacement of the structural component body or the normal rigid edge structure under the same imaging conditions.

[0025] After completing the above calibration, the pulse dynamic pressure Q during actual detection is set between the first critical pressure P1 and the second critical pressure P2, i.e., P1 < Q < P2. Preferably, in order to reduce the influence of air source fluctuations and local shape differences in different areas to be detected on the airflow state, a reserve margin can be set above the first critical pressure P1 and below the second critical pressure P2, so that the actual working dynamic pressure is within the stable working range determined by calibration. Within this working range, the sheet-like residual flash can form local deflection and free end migration that meet the requirements of image analysis under the action of airflow disturbance, while the structural body and normal rigid edge structure remain in a state below the preset pixel-level deformation displacement, so that the subsequent obtained time series images can be used to extract the local dynamic changes of the edge attachment candidate relative to the parent body tangent segment.

[0026] In one optional implementation, when there are differences in the burr thickness, overhang length, or local spatial posture of different cut edge regions, corresponding first critical pressure P1 and second critical pressure P2 can be established respectively. Alternatively, multiple sets of dynamic pressure intervals can be pre-stored according to the type of the region to be detected. During the detection process, the corresponding airflow control parameters are called according to the current position of the cut edge region or the predetermined cut edge contour. If a working interval that meets the predetermined margin requirement is not formed between the first critical pressure P1 and the second critical pressure P2 during the calibration process, the calibration can be performed again by changing the distance between the nozzle and the cut edge region to be detected, the injection tilt angle, the nozzle diameter, or the image acquisition magnification to obtain working conditions suitable for applying lateral disturbance to sheet-like edge targets.

[0027] Furthermore, the acquisition process of the edge-cutting time-series image group includes: before the pulsed airflow establishes a substantial disturbance, acquiring at least one frame as an initial edge-cutting image to establish a static reference under no external force; during the peak window period of the continuous action of the pulsed airflow, continuously acquiring multiple frames of disturbance edge-cutting images to extract the maximum flexural displacement and deformation trajectory of the extended portion and free end region under airflow stress; after the pulsed airflow stops and the external force is removed, continuously acquiring multiple frames of recovery edge-cutting images, the duration of continuous acquisition being configured to cover at least one effective recovery stage of the free end region returning from the disturbance displacement state to the pre-disturbance position state, and continuous acquisition until it is determined that the offset of the free end region shows a decreasing trend and the outline of the extended portion shows a regression trend, fully recording the gradual recovery process of the free end region returning to the initial outline.

[0028] Furthermore, the edge-cutting time-series image group is continuously acquired according to the application sequence of the pulsed airflow, and sequentially includes the initial edge-cutting image before disturbance, the disturbed edge-cutting image during the continuous action of the airflow, and the restored edge-cutting image after the airflow is removed. In specific implementation, the imaging area of ​​the image acquisition device covers the edge-cutting area to be detected, and the relative position between the image acquisition device and the structure to be detected and the exposure parameters remain unchanged during one detection process. After the image acquisition begins, the pulsed airflow is kept in a closed state, and at least one frame of the edge-cutting area to be detected is continuously acquired. When multiple frames are acquired, one frame with a complete outline and no instantaneous occlusion can be selected as the initial edge-cutting image, or the initial outline position before disturbance can be determined based on multiple frames of images, and this state can be used as a static reference when no external force is applied.

[0029] After completing the image acquisition before disturbance, the pulsed airflow is activated and image acquisition is maintained continuously. The pulsed airflow has an airflow establishment phase from activation to reaching a stable intensity. This phase can be pre-calibrated based on the nozzle outlet pressure, the distance between the nozzle and the area to be detected, and the pulse valve response time. The establishment time corresponding to the airflow reaching the preset working dynamic pressure from the activated state is determined, and the time interval during which the airflow reaches the preset working dynamic pressure and remains within the allowable fluctuation range is defined as the peak window period. The frame rate of the image acquisition device is set based on the airflow establishment time, the peak window duration, and the recovery time of edge attachment candidates after airflow removal. This ensures that the sampling time interval between adjacent images is less than the characteristic time scale of the dynamic process to be identified, guaranteeing that both the airflow disturbance and recovery phases can be captured. By acquiring multiple consecutive frames of images, when a pressure detection element is set, the peak window period can also be determined based on the air pressure signal corresponding to the area to be detected reaching a preset working dynamic pressure range. Within this peak window period, multiple frames of disturbed edge images are continuously acquired, so that at least two frames correspond to the state of continuous airflow application, and the acquisition time covers the process of the edge-attached candidate body shifting from the initial state to the formation of the maximum or near-maximum deflection state. The acquisition frame rate can be determined by the airflow loading test of the calibration sample before detection, and the image sampling interval within the peak window period is sufficient to characterize the process of the candidate body changing from the initial state to the maximum or near-maximum deflection state. When a pressure detection element is set, the airflow action stage of each disturbed edge image can be determined based on the correspondence between the pressure signal and the acquisition time of each frame image.

[0030] After the peak window period ends, the pulsed airflow is shut off to remove the external aerodynamic force acting on the edge to be detected, while image acquisition continues uninterrupted without interruption when the pulsed airflow is shut off. The first frame after the pulsed airflow stops or the first frame after the air pressure drops to a preset unloading threshold is used as the starting image for the recovery phase. Subsequently, multiple frames of recovered edge images are continuously acquired at the same sampling frame rate as the disturbance phase or a preset rate to record the temporal process of the edge attachment candidate changing from the disturbed position to the pre-disturbance position. The acquisition frame rate of the recovery phase is set according to the recovery time of the edge attachment candidate returning from the disturbed state to the pre-disturbance state, so that multiple frames of recovered edge images can be continuously acquired within at least one effective recovery phase, and the decreasing trend of the free end region offset and the regression trend of the protrusion contour can be determined based on the multiple frames. When the calibration results show that the candidate recovery process is fast, the image acquisition frame rate is increased accordingly to avoid missing the main recovery process due to excessively large sampling intervals between adjacent frames.

[0031] The acquisition duration of the recovery phase is determined based on the continuous image changes of the free end region and the extension. The position of the free end and the outline of the extension in the initial cut edge image are used as recovery references. In the continuously recovered cut edge images, the offset of the free end region relative to the initial position and the distance between the current outline of the extension and the initial outline are determined sequentially. When the offset of the free end region relative to the initial position in multiple consecutive recovered cut edge images generally shows a decreasing trend, and the distance between the current outline of the extension and the initial outline generally shows a decreasing trend, it is determined that at least one effective recovery phase has been covered. Continuous acquisition can continue until the above trend is maintained within a preset number of consecutive frames, or until the offset of the free end region and the deviation of the extension outline enter a preset recovery range and then it ends.

[0032] A single edge-cutting time-series image acquisition is completed in the following time sequence: acquisition in the no-airflow state—airflow establishment—continuous acquisition of the peak window—airflow removal—continuous acquisition of the recovery stage. Images of each stage are saved in the order of acquisition time, and the corresponding frame number and time relationship with the opening and closing times of the pulse airflow are retained. The initial edge-cutting image corresponds to the state without external force, the perturbed edge-cutting image covers the deflection process under the stable action of airflow, and the recovered edge-cutting image covers at least one effective recovery stage after the airflow is removed, in which the image returns from the perturbed position to the initial position.

[0033] Furthermore, the process of determining the edge attachment candidate includes: registering and comparing the actual contour in the initial tangent image with the tangent reference contour; extracting the convex image region located outside the tangent reference contour and having a continuous pixel connection with the actual parent tangent region as the edge attachment candidate; and defining the local edge segment in the actual parent tangent region that directly intersects with the edge attachment candidate in space and forms a continuous structure as the parent tangent segment corresponding to the candidate.

[0034] Furthermore, the edge attachment candidate is extracted based on the actual contour in the initial chopped edge image and the pre-determined chopped edge reference contour. In specific implementation, the chopped edge reference contour corresponding to the chopped edge area to be detected is first determined. The chopped edge reference contour can be obtained according to the design contour of the integrated die-cast aluminum alloy automotive structural part to be detected, or it can be obtained by extracting the chopped edge contour after image acquisition of a standard structural part that meets the chopped edge requirements. The chopped edge reference contour is pre-established according to the positional relationship with the chopped edge area to be detected, and at least covers the local cut edge range to be detected.

[0035] After obtaining the initial cut edge image, the actual parent body cut edge region and its actual contour are extracted from the initial cut edge image. Specifically, the structural component region can be determined based on the grayscale difference, brightness difference, or edge response between the surface of the structural component and the background, and the actual contour is obtained along the boundary between the structural component region and the background region. For images with local noise, reflection, or edge breaks, smoothing can be performed before contour extraction, and isolated edges with a length less than a preset value and discrete image regions not connected to the main body region of the structural component are removed to obtain a continuous actual parent body cut edge region and its boundary contour.

[0036] Subsequently, the actual contour in the initial chopped image is registered and compared with the chopped reference contour. During registration, stable structural edges, corner points, hole edges, or chopped contour feature points near the chopped area to be detected can be selected as registration references. Based on the positional offset and directional deviation between the corresponding references, the chopped reference contour is translated, rotated, or transformed under the condition of consistent scale, so that the chopped reference contour corresponds to the main body edge in the actual mother body chopped area that does not contain local protrusions. After registration, the chopped reference contour is mapped to the image coordinate range where the initial chopped image is located, so that the actual contour and the chopped reference contour can be compared under the same image position relationship.

[0037] After registration is completed, the actual parent body cutting edge region is compared along the cutting edge reference contour. Image regions that exceed the cutting edge reference contour and are located outside it are identified as protruding image regions to be screened. For each protruding image region to be screened, the pixel connectivity relationship between it and the actual parent body cutting edge region is further determined. When the protruding image region can be directly connected to the actual parent body cutting edge region through continuous foreground pixels, and there is no complete separation between the two by background pixels, it is determined that the protruding image region has a continuous pixel connectivity relationship with the actual parent body cutting edge region, and the protruding image region is extracted as the edge attachment candidate. For isolated regions that are completely outside the parent body cutting edge region and have no continuous pixel connection with it, they are not identified as edge attachment candidates in this step. In one specific implementation, the continuous pixel connectivity can be determined using a pixel neighborhood-based connectivity determination method, preferably using 8-neighborhood connectivity determination. That is, for a foreground pixel in a protruding image region, when there is a connectivity path formed by the foreground pixel through continuous foreground pixels adjacent in the horizontal, vertical, or diagonal direction, and the connectivity path extends to the foreground pixel in the actual parent cut edge region, it is determined that the protruding image region and the actual parent cut edge region have a continuous pixel connectivity relationship. For local false connections caused by isolated noise pixels, the foreground image can be denoised before connectivity determination, or the false connections can be excluded based on the number of continuously connected pixels and the size of the connection region. The determination parameters used are preset according to the image resolution, imaging magnification, and foreground pixel positioning error.

[0038] After determining the edge attachment candidate, the corresponding parent cut edge segment is determined along the intersection position between the edge attachment candidate and the actual parent cut edge region. Specifically, the intersection boundary formed by the direct connection between the edge attachment candidate and the actual parent cut edge region is obtained. On one side of the actual parent cut edge region, with the cut edge position corresponding to the intersection boundary as the center, edge segments of a set length are continuously taken on both sides along the local extension direction of the actual cut edge. The taken edge segments remain within the actual parent cut edge region and form a continuous structure with the intersection position of the edge attachment candidate. This local edge segment is defined as the parent cut edge segment corresponding to the edge attachment candidate. When there are multiple protruding image regions that are separated from each other in the same cut edge region to be detected, each edge attachment candidate and its corresponding parent cut edge segment are determined according to the above process. When there are multiple edge attachment candidates in the same cut edge region to be detected, the association relationship between each edge attachment candidate and its corresponding parent cut edge segment is established in the initial cut edge image, and an independent cross-frame corresponding identifier is determined for each edge attachment candidate. In subsequent perturbed and recovered edge images, each candidate body is associated across frames based on its positional relationship with the corresponding parent body edge segment, its region position, contour size, contour shape, and motion continuity in adjacent frames, so that the temporal change data of the connecting parts, extensions, and free ends of different candidate bodies remain independent of each other.

[0039] For candidate objects that remain connected to the corresponding parent cut edge segment, the cross-frame correspondence is determined primarily based on the parent cut edge segment and the original connection position. For candidate objects that separate from the parent cut edge segment and undergo overall migration under airflow disturbance, the corresponding target is determined in the current frame based on its position, motion direction, contour size, and contour shape in the previous frame. When multiple potentially corresponding candidate objects exist within the same search range, the continuity of positional changes, the degree of contour size changes, and the similarity of contour shapes between each candidate object and the target in the previous frame are comprehensively compared. The target with the highest comprehensive matching degree that meets the preset association conditions is selected to establish a cross-frame correspondence. When a unique corresponding target that meets the association conditions cannot be obtained, the corresponding time frame is marked as an invalid association state, and the motion data of other candidate objects are not used as the local differential motion features or unloading recovery motion features of that candidate object.

[0040] Further, the process of defining the parent cut edge segment, connecting portion, extension portion, and free end region of the candidate body includes: establishing a spatial topological extension direction or distance field outward from the parent cut edge segment based on the extracted edge attachment candidate body; defining the root transition region of the edge attachment candidate body that is close to the parent cut edge segment and serves as the starting point for outward extension as the connecting portion; defining the main body region of the edge attachment candidate body that grows outward from the cut edge reference contour along the spatial topological extension direction as the extension portion, starting from the connecting portion; and defining the end boundary region of the edge attachment candidate body located on the side of the extension portion away from the connecting portion as the free end region.

[0041] Furthermore, after determining the edge attachment candidate and its corresponding parent cut edge segment, the intersection position between the parent cut edge segment and the edge attachment candidate is taken as the starting position for the candidate to extend outward. A spatial topological relationship extending from the parent cut edge segment to the end of the candidate is established within the continuous image area corresponding to the edge attachment candidate. Specifically, the pixel position directly adjacent to the parent cut edge segment in the edge attachment candidate can be determined as the starting boundary. From the starting boundary, the candidate extends step by step along the continuous pixels inside the candidate to the side away from the cut edge reference contour. The topological extension order of the candidate from the inside to the outside is determined according to the distance between each position and the starting boundary along the continuous path inside the candidate.

[0042] In one specific implementation, a distance field of the edge attachment candidate can be established based on the aforementioned starting boundary. For each pixel position inside the candidate, the shortest continuous path length from that position to the starting boundary along the continuous region of the candidate is determined, and this path length is used as the distance value of the corresponding position. Positions closer to the parent body's cut edge have smaller distance values, and the distance values ​​gradually increase as they extend outward along the candidate. This yields the spatial topological extension direction from the parent body's cut edge to the far end of the candidate. For candidates with curved or inclined extensions, the path is determined along the continuous region of the candidate itself, rather than using a single horizontal or vertical direction in the image coordinates as the extension direction.

[0043] Based on the spatial topological extension direction or distance field, on the side of the candidate attached to the edge that is close to the parent body's cut edge segment, taking the starting boundary that directly intersects with the parent body's cut edge segment as a reference, a preset distance range is continuously extended into the candidate body. The root transition region within this range that is close to the parent body's cut edge segment and constitutes the starting point for the candidate body's outward extension is defined as the connecting part. The preset distance range can be determined according to the candidate body's width in the image, the imaging magnification, or a preset pixel distance, and the defined area continuously covers the main intersection position between the candidate body and the parent body's cut edge segment.

[0044] Taking the boundary of the connecting part away from the mother body's tangent segment as the starting position of the extension, it continues to extend in the direction of increasing distance value along the spatial topology extension direction. The main image region located outside the connecting part and continuously extending outward from the tangent reference contour in the edge attachment candidate is defined as the extension. In the case of local bending of the candidate, the extension path of the extension is determined according to the continuous change direction from a smaller distance value to a larger distance value in the distance field, so that the extension remains continuously connected to the connecting part.

[0045] Further, the side furthest from the connection part in the edge attachment candidate body is determined along the distance field, and continuous end boundary regions with distance values ​​reaching the maximum value or within the preset far end distance range are extracted in the far end boundary of the candidate body. This region is defined as the free end region. When the far end boundary is composed of multiple adjacent pixels, the preset length boundary continuously extending to both sides around the maximum distance position can be used as the free end region. Thus, the connection part, the extension part and the free end region are determined sequentially along the spatial topological order outward from the parent body cutting edge segment, and the three are kept to correspond continuously within the same edge attachment candidate body.

[0046] Further, the extraction process of the local differential motion features and unloading recovery motion features includes: extracting the contour features of the parent tangent segment as a local coordinate reference between frames of the tangent time sequence image group, and performing relative position compensation and spatial alignment on each frame image; in the airflow disturbance stage image after position compensation and spatial alignment, determining the first relative displacement of the connecting part relative to the local coordinate reference, and the second relative displacement of the free end region relative to the connecting part, and determining the state in which the amplitude of the first relative displacement is less than a first set threshold and the amplitude of the second relative displacement is greater than the amplitude of the first relative displacement, forming a differential motion relationship between the connecting part and the free end region, as the local differential motion feature; in the recovery stage image after position compensation and spatial alignment, continuously extracting the temporal change data of the free end region and the extension relative to the local coordinate reference, and determining the recovery displacement vector of the free end region in the opposite direction to the displacement vector in the airflow disturbance stage, and the remaining offset amplitude of the free end region relative to the position before the disturbance decreasing frame by frame, while the temporal dynamic change of the extension contour approaching the corresponding contour in the initial tangent image frame by frame, as the unloading recovery motion feature.

[0047] Furthermore, after obtaining the time-series image set of the edge trimming and defining the region of the edge attachment candidate, the parent edge trimming segment corresponding to the candidate is used as a cross-frame motion reference. Relative position compensation and spatial alignment are performed on each frame image before the disturbance, during the disturbance, and during the recovery phase. Specifically, the contour point set of the parent edge trimming segment is extracted in the initial edge trimming image, and stable edge segments, endpoints, or curvature change positions in the contour point set are selected as local contour features. The contour features of the corresponding parent edge trimming segment are extracted in each subsequent frame image. The translation amount and direction deviation of each frame relative to the initial edge trimming image are determined by the position matching between the corresponding contour points. Position compensation is performed on each frame image accordingly to make the parent edge trimming segments in different frames coincide or remain within the preset allowable deviation range. After spatial alignment is completed, the local coordinate reference is established by using the local extension direction of the parent edge trimming segment as the tangential direction and the direction towards the outside of the edge trimming reference contour as the normal direction.

[0048] In one specific implementation, during a single edge-trimming time-series image acquisition process, the set of contour points of the parent edge segment in the initial edge-trimming image is used as the reference contour point set, and the corresponding set of contour points of the parent edge segment in subsequent frames is used as the set of contour points to be registered. A two-dimensional rigid transformation including translation and rotation is used to spatially align the set of contour points to be registered. Based on the correspondence between the reference contour point set and the set of contour points to be registered, the translation amount and rotation direction deviation of the current frame relative to the initial edge-trimming image are determined by reducing the positional deviation between the corresponding contour points after spatial transformation. The correspondence between the contour points can be determined based on contour feature points, neighboring contour points, or local contour morphological features.

[0049] After spatial alignment is completed, the positional deviation between the transformed parent body cut edge contour and the corresponding parent body cut edge contour in the initial cut edge image is calculated. The registration residual is determined based on the average value, root mean square value, or other statistical quantities that can characterize the overall alignment degree of multiple corresponding contour points. When the registration residual does not exceed the preset registration allowable deviation, the current frame is determined to meet the spatial alignment requirements. The preset registration allowable deviation is pre-calibrated based on the repetitive positioning error of the same parent body cut edge continuous images under no airflow conditions and is set to be less than the corresponding displacement threshold used to determine the relative displacement of the connecting part and the relative migration of the free end, so that the positional error caused by image acquisition jitter and spatial registration does not constitute a relative displacement that meets the motion determination conditions.

[0050] In the airflow disturbance effect stage image after position compensation and spatial alignment, the cross-frame positions of the connecting part and the free end region are determined respectively. For the connecting part, the contour center position, feature point position, or geometric center of the corresponding contour region of the connecting part can be selected as the position representation point, and its position change relative to the local coordinate reference between the current disturbance image and the initial truncated image is calculated to obtain the first relative displacement of the connecting part. For the free end region, the end contour position corresponding to the free end region in the initial truncated image is determined in each disturbance truncated image, and the displacement change of the free end region relative to the connecting part position in the current frame is calculated to obtain the second relative displacement of the free end region relative to the connecting part.

[0051] The amplitudes of the first relative displacement and the second relative displacement obtained from the continuously disturbed edge-cutting image are compared. When the amplitude of the first relative displacement of the connecting part is less than a preset first threshold, and the amplitude of the second relative displacement of the free end region relative to the connecting part is greater than the amplitude of the first relative displacement, it is determined that the connecting part and the free end region have not undergone consistent overall migration during the airflow process, but have formed a differential motion relationship with a smaller amplitude of motion in the connecting part and a larger amplitude of relative motion in the free end region. This state is defined as the local differential motion feature. The first threshold can be predetermined based on the repeated positioning deviation of the mother body edge-cutting segment and the connecting part under no airflow conditions, the registration residual after spatial alignment of each frame image, and the allowable pixel position fluctuation under image acquisition conditions. The first threshold is made greater than the fluctuation range formed by the above-mentioned normal position error to avoid misjudging the position change caused by image acquisition jitter or spatial registration error as the actual relative motion of the connecting part.

[0052] After the pulsed airflow stops, the continuously acquired recovered edge images are positionally compensated and spatially aligned according to the same parent edge segment contour features. The position and contour of the free end region and the extension in each recovered image are continuously determined. The displacement direction of the free end region from the pre-disturbance position to the disturbed position during the airflow disturbance stage is taken as the disturbance displacement direction. During the recovery stage, the recovery displacement vector of the free end region relative to its disturbed position is calculated sequentially. When the direction of the recovery displacement vector in the continuously recovered images is reversed relative to the displacement direction during the disturbance stage, and the remaining offset amplitude of the free end region relative to the pre-disturbance position decreases frame by frame during the recovery process, it is determined that the free end region is in the process of returning to the pre-disturbance position.

[0053] Simultaneously, using the outline of the extended portion in the initial cut edge image as a reference, the corresponding outline of the extended portion is extracted in each recovered cut edge image, and corresponding boundary points are selected along the outline to determine the positional deviation between the current extended portion outline and the corresponding outline in the initial cut edge image. When the positional deviation of the outline in the continuously recovered images decreases frame by frame, making the extended portion outline approach the corresponding outline in the initial cut edge image frame by frame, and the free end region satisfies the above-mentioned conditions of direction reversal and frame-by-frame decrease in offset amplitude, the continuous temporal dynamic change is determined as the unloading recovery motion feature.

[0054] Furthermore, the process in which the connecting portion of the candidate remains attached during airflow disturbance, the extended portion undergoes local deflection constrained by the connecting portion, and the free end experiences relative migration includes: extracting the coordinate offset or contour overlap of the connecting portion relative to the parent body's cut edge segment in a cross-frame image; when the coordinate offset is less than a preset holding threshold or the contour overlap is greater than a preset ratio, it is determined that the connecting portion remains attached. The preset holding threshold is set to cover the positional fluctuation range corresponding to the repeated positioning deviation and spatial alignment error of the connecting portion relative to the parent body's cut edge segment under no airflow conditions; the preset ratio is based on the connecting portion under no airflow conditions. The contour overlap distribution between the extension and the initial connection is determined so that the overlap change caused by contour extraction and positional fluctuations under normal static conditions still meets the continuous attachment determination condition, while the determination condition is not met when the connection significantly detaches or migrates as a whole; the contour curvature data or morphological skeleton angle data of the extension in cross-frame images are extracted, and when the amount of curvature change or angle change during airflow is greater than a set deformation threshold, it is determined that the extension has produced local deflection constrained by the connection; the displacement vector of the free end region relative to the connection is calculated, and when the magnitude of the displacement vector is greater than a preset migration threshold, it is determined that the free end has undergone relative migration, specifically as follows. Figure 3 As shown.

[0055] Furthermore, in the airflow disturbance stage image after position compensation and spatial alignment, the connecting part, the extended part and the free end region are tracked across frames to determine the connection maintenance state, local deformation state and free end migration state of the candidate body under the action of airflow disturbance. The connecting part, the extended part and the free end region in each frame are all based on the corresponding region defined in the initial cut edge image as the initial position, and their position changes in the subsequent disturbance cut edge image are determined according to the local coordinate reference established by the parent body cut edge segment.

[0056] To determine whether the connecting part remains attached, the position of the connecting part relative to the parent body's cutting edge segment is extracted from each disturbed cutting edge image, and the coordinate offset between the current feature position of the connecting part and the corresponding feature position in the initial cutting edge image is calculated. The coordinate offset can be determined by the average position change of the connecting part contour center, the midpoint of the connecting boundary, or multiple corresponding contour feature points. As another determination method, the connecting part contour of the current frame can be extracted and compared with the corresponding connecting part contour in the initial cutting edge image. The ratio of the overlapping area of ​​the two to the area of ​​the initial connecting part region, or the ratio of the intersection area to the union area of ​​the two, is used as the contour overlap degree. When the coordinate offset of the connecting part is less than the preset holding threshold, or the contour overlap degree of the connecting part is greater than the set ratio, it is determined that the connecting part remains attached to the parent body's cutting edge segment. The preset holding threshold and the set ratio can be calibrated by continuously acquiring the same cutting edge region image under no airflow conditions before detection. The coordinate offset range and contour overlap range of the connecting part caused by image positioning error, small vibration of structural components, and contour extraction fluctuation are determined to ensure that the image position fluctuation generated under normal static conditions is within the holding determination range.

[0057] To determine local deflection of the extended portion, the contour of the extended portion is extracted from the initial cut edge image and each disturbed cut edge image. Continuous contour points are selected along the extension direction of the extended portion from the connecting part to the free end. The contour curvature at each position is determined based on the directional changes of adjacent contour points. Then, the change in contour curvature during the disturbance stage relative to the initial state is calculated. Alternatively, the extended portion image region can be refined to obtain a morphological skeleton distributed along the extension direction of the extended portion. The local tangent directions of the skeleton in the region near the connecting part and in the region extending towards the free end are determined, and the angular changes of the corresponding skeleton segments before and after the disturbance are calculated. When the curvature change exceeds a preset deformation threshold... If the change in the skeleton angle is greater than the corresponding deformation threshold, it is determined that the extension has a local deflection with the area where the connection is located as the constraint end during the airflow. The deformation threshold can be calibrated by the natural fluctuation range of the curvature of the extension contour or the skeleton angle in the continuous image under no airflow condition, under the condition of using the same contour preprocessing method, local calculation scale and curvature or skeleton angle calculation method as the actual detection, and determined in combination with the pre-defined minimum identifiable local deformation, so that the curvature change or angle change caused by contour extraction error and natural position fluctuation is lower than the deformation threshold, and the local deflection that reaches the predetermined identifiable level can meet the deformation determination condition.

[0058] To determine the relative migration of the free end, representative positions of the free end region are determined in the initial cut edge image and each disturbed cut edge image. The representative positions can be determined by the geometric center of the free end region, the center of the end contour, or the contour position farthest from the connector. Using the position of the connector as a relative reference, the relative position vector from the connector to the representative position of the free end region is calculated in the initial state and the current disturbed state, and the difference between the two is taken as the displacement vector of the free end region relative to the connector. When the magnitude of the displacement vector is greater than a preset migration threshold, it is determined that the free end region has migrated relative to the connector. The preset migration threshold can be pre-calibrated based on the spatial resolution of the image acquisition system and the position fluctuation range generated by repeated positioning of the free end region under no airflow conditions, and is set to be greater than the displacement range corresponding to the positioning fluctuation of the normal image.

[0059] During the same airflow disturbance phase, when the connecting part meets the condition for continuous attachment, the extending part meets the condition for local deflection, and the displacement vector amplitude of the free end region relative to the connecting part is greater than the preset migration threshold, it is determined that the current edge attachment candidate has a motion state in which the connecting part continuously maintains attachment, the extending part undergoes local deflection with the connecting part as a constraint, and the free end undergoes relative migration.

[0060] Furthermore, the determination process for the extension to regress to the pre-disturbance contour includes: calculating the recovery displacement vector of the free end region in consecutive multi-frame recovery images; when the angle between the recovery displacement vector and the displacement vector obtained during the airflow disturbance stage falls within a preset reverse angle range, it is determined that the free end has generated a recovery migration opposite to the disturbance direction; extracting the contour features of the extension in consecutive multi-frame recovery images; when the image matching degree between the extension contour and the corresponding contour in the initial tangent image increases frame by frame, or the spatial distance between the contour boundaries decreases frame by frame, it is determined that the extension regresses to the pre-disturbance contour.

[0061] Furthermore, after the pulsed airflow stops, according to the local coordinate reference established by the aforementioned mother body tangent segment, position compensation and spatial alignment are performed on multiple consecutive frames of recovered tangent images, and the positions of the free end region and the extension are continuously determined in each recovered tangent image. The representative position of the free end region in the initial tangent image is taken as the position before disturbance, and the representative position when the free end region reaches its maximum displacement during the airflow disturbance stage is taken as the disturbance position. The vector pointing from the position before disturbance to the disturbance position is determined as the displacement vector during the airflow disturbance stage. The representative position of the free end region is determined in the same way as the aforementioned disturbance stage, and can be the geometric center of the free end region, the center of the end contour, or the contour position farthest from the connection part.

[0062] For the recovery migration of the free end region, the current representative position of the free end region is determined in each recovery edge image, and a corresponding recovery displacement vector is formed with the disturbance position as the starting point and the current representative position as the ending point. The directional angle between each recovery displacement vector and the displacement vector of the airflow disturbance stage is calculated. When the directional angle falls within the preset reverse angle interval, it is determined that the current recovery displacement direction is opposite to the displacement direction of the airflow disturbance stage. When multiple consecutive recovery edge images satisfy the directional relationship, it is determined that the free end region has generated recovery migration opposite to the disturbance direction. The preset reverse angle interval is set with 180° corresponding to the two vector directions being opposite as the center of the allowable angle deviation. The allowable angle deviation can be determined by performing multiple airflow loading and unloading on the calibration sample and statistically analyzing the angular fluctuation range of the recovery direction relative to the disturbance direction of the free end during normal recovery. The same angle determination condition is used in subsequent detection.

[0063] Simultaneously, the protrusion contour is extracted from each recovered edge image and compared frame by frame with the corresponding protrusion contour in the initial edge image. In one embodiment, the protrusion region in the current recovered image is matched with the initial protrusion region in terms of position, and the degree of contour overlap or corresponding region overlap between the two is calculated. The result is used as the image matching degree. According to the acquisition time sequence of the recovered images, when the image matching degree of the adjacent subsequent frame is greater than the image matching degree of the previous frame, it is determined that the protrusion contour approaches the corresponding contour in the initial edge image frame by frame.

[0064] As another implementation, continuous boundary points can be selected along the protrusion contour in the initial cut edge image, and corresponding boundary positions can be determined on the corresponding protrusion contour in each restored cut edge image. The distance between the corresponding boundary points can be calculated, and the average or maximum distance between the corresponding boundary points can be used to characterize the spatial distance between the current protrusion contour and the initial contour. The images are compared in chronological order. When the spatial distance of the contour corresponding to the subsequent frame is less than the spatial distance of the contour corresponding to the previous frame, the protrusion contour is determined to approximate the corresponding contour in the initial cut edge image frame by frame. If there are local pixel fluctuations during the contour extraction process, the same parameters can be used to smooth the contour of each frame before frame-by-frame comparison to keep the contour extraction conditions consistent for each restored image.

[0065] After completing the above-mentioned recovery stage determination, the dynamic state of the recovery stage is matched with the dynamic state of the airflow disturbance stage. When the same edge attachment candidate satisfies the following conditions during the airflow disturbance: the connecting part continues to be attached, the extension part produces local deflection constrained by the connecting part, and the free end region migrates relative to the connecting part, and in the continuous recovery images after the airflow is removed, the free end region produces recovery migration opposite to the disturbance direction, and the image matching degree between the extension part contour and the corresponding contour in the initial cut edge image increases frame by frame, or the spatial distance between the extension part contour and the initial corresponding contour decreases frame by frame, the edge attachment candidate is determined to be a sheet-like residual burr.

[0066] Example 2: Furthermore, when the edge attachment candidate fails to maintain the connection and attachment relationship with the corresponding parent body cut edge segment during airflow disturbance, and undergoes an overall migration relative to the parent body cut edge segment in the same direction under the action of airflow without forming the local differential motion feature, the defect determination module determines the edge attachment candidate as loose cut edge debris.

[0067] Furthermore, during the airflow disturbance stage, the edge attachment candidates identified by the initial edge-cutting image are continuously matched across frames. Based on the connection status between the candidate and the parent edge-cutting segment, as well as the motion status of each region of the candidate, it is determined whether it belongs to loose edge-cutting debris. In specific implementation, the position, contour shape and corresponding parent edge-cutting segment of the edge attachment candidate in the initial edge-cutting image are used as the initial reference. The target region continuously corresponding to the contour position of the candidate is extracted in the continuously disturbed edge-cutting image after position compensation and spatial alignment. When the candidate is spatially separated from the parent edge-cutting segment due to airflow, its cross-frame correspondence is still maintained according to the positional proximity, contour size and motion continuity of the target region in the preceding and following frames, so as to obtain the continuous motion process of the candidate before and after separation.

[0068] To determine the connection and attachment relationship between the candidate body and the parent body's cut edge segment, the pixel connection status between the original connection position of the candidate body and the corresponding parent body cut edge segment is checked in each disturbed cut edge image. When the continuous foreground pixel connection between the two in the initial state is interrupted in the image after airflow disturbance, and a continuous background interval is formed between the candidate body and the parent body cut edge segment, it is determined that the candidate body has failed to maintain the connection and attachment relationship with the corresponding parent body cut edge segment. As another determination method, the coordinate offset or contour overlap of the connection part relative to the parent body cut edge segment can be used for judgment. When the coordinate offset is used for judgment, if the coordinate offset of the connection part reaches or exceeds the preset holding threshold, it is determined that the original connection and attachment relationship has been released. When the contour overlap is used for judgment, if the contour overlap between the connection part and the original intersection position is less than or equal to the set ratio, it is determined that the original connection and attachment relationship has been released.

[0069] After confirming that the connection and attachment relationship could not be maintained, the positional changes of the candidate object in the continuously disturbed edge images are extracted. The representative positions of the original connection region, the extension region, and the free end region of the candidate object are determined in the preceding and following adjacent frames, and the displacement vector of each region relative to the local coordinate reference of the parent body edge segment is calculated. When the displacement vectors of the above regions are all oriented towards the migration direction caused by the airflow, and the directional angle between the displacement vectors is within a pre-set range of consistent direction, it is determined that each region of the candidate object has a consistent migration state. The range of consistent direction can be pre-determined based on the airflow jet direction and the normal fluctuation range of the displacement direction of each region when the loose target is driven by the airflow during the calibration process.

[0070] Further compare the displacement changes of the connecting part, the extended part, and the free end region. If each region changes position relative to the parent body's tangent segment along with the candidate body, and the free end region does not produce the second relative displacement relative to the connecting part that meets the aforementioned criteria for determining local differential motion characteristics, and the extended part does not exhibit local deflection with the original connecting part region as the constraint end, then it is determined that the candidate body has not formed the aforementioned local differential motion characteristics. Specifically, the relative motion relationship between the connecting part and the free end region can be calculated according to the aforementioned method for extracting local differential motion characteristics. When the differential relationship where the first relative displacement is limited and the second relative displacement is significantly increased is not met, it is taken as the determination result that no local differential motion characteristics have been formed.

[0071] When the same edge-attached candidate fails to maintain the connection and attachment relationship with the corresponding parent body cut edge segment during airflow disturbance, and the connecting part, extension part and free end region of the candidate migrate in the same direction relative to the parent body cut edge segment, and the local differential motion feature is not formed inside the candidate, the edge-attached candidate is judged as loose cut edge debris.

[0072] Furthermore, when the connecting part maintains a corresponding connection relationship with the mother body's cut edge segment during airflow disturbance, and the free end region does not undergo relative migration to the connecting part to a preset visually identifiable level, while the extended part does not undergo identifiable local flexural deformation with the region where the connecting part is located as the constraint end, the defect determination module determines the edge attachment candidate as a rigid cut edge structure.

[0073] First, the positional relationship between the connecting part and the corresponding parent body cutting edge segment is determined in each disturbance cutting edge image. Then, according to the aforementioned determination method for the connecting part to maintain continuous attachment, the coordinate offset of the connecting part relative to the parent body cutting edge segment or the contour overlap between the connecting part contour and the corresponding contour in the initial state is calculated. When the coordinate offset of the connecting part in the continuous disturbance cutting edge images remains less than the preset holding threshold, or the contour overlap remains greater than the set ratio, it is determined that the connecting part maintains a continuous connection relationship with the corresponding parent body cutting edge segment during the airflow disturbance.

[0074] After confirming that the connection remains connected, the connection is used as a relative position reference to continuously determine the representative position of the free end region in each disturbance edge image. The displacement vector of the free end region relative to the connection is obtained according to the aforementioned calculation method for the relative migration of the free end. The magnitude of the displacement vector is compared with a preset migration threshold. When the magnitude of the displacement vector of the free end region relative to the connection does not reach the preset migration threshold during the airflow disturbance stage, it is determined that the free end region has not produced a relative migration that reaches the preset visually recognizable level. The preset migration threshold adopts the same calibration conditions as the aforementioned determination of sheet-like residual burrs. It can be determined based on the spatial resolution of the current image acquisition system, the repeated positioning fluctuation of the free end region in the absence of airflow, and the preset minimum recognizable pixel displacement, so that positional changes smaller than the threshold are classified as not reaching the preset visually recognizable level.

[0075] Simultaneously, contour curvature data or morphological skeleton angle data of the extended portion are extracted from the initial cut edge image and each disturbed cut edge image, and the amount of curvature change or skeleton angle change before and after the airflow action is calculated according to the aforementioned local deflection determination method; when using contour curvature data for local deflection determination, if the amount of curvature change is less than or equal to the corresponding set deformation threshold, it is determined that the extended portion has not produced local deflection deformation that reaches the preset visually recognizable level; when using morphological skeleton angle data for local deflection determination, if the amount of skeleton angle change is less than or equal to the corresponding set deformation threshold, it is determined that the extended portion has not produced local deflection deformation that reaches the preset visually recognizable level.

[0076] During the same airflow disturbance stage, when the connecting part of the edge-attached candidate body continues to maintain the connection relationship with the corresponding parent body tangent segment, the displacement vector amplitude of the free end region relative to the connecting part does not reach the preset migration threshold, and the deformation change determined according to the adopted local deflection determination method is less than or equal to the corresponding set deformation threshold, it is determined that the candidate body has not formed a free end relative migration and local deflection motion state corresponding to the sheet-like residual flash under the current pulse airflow action condition, and the edge-attached candidate body is determined to be a rigid tangent structure.

[0077] The above embodiments are preferred embodiments of the present invention and are intended to explain the technical solutions of the present invention, rather than to limit the invention. Within the protection scope of the present invention, any equivalent modifications or substitutions can be made. The scope of the present invention is defined in the claims.

Claims

1. A vision-based vehicle exterior defect detection system, characterized in that, include: The image acquisition module continuously acquires images of the edge to be detected and applies non-contact pulsed airflow disturbance during the acquisition process to obtain a time-series image group of the edge, which includes multiple frames of recovered images before, during and after the disturbance is removed. The candidate body extraction module, based on the tangent reference contour, determines the edge protrusion targets that connect to the outer side of the actual parent body tangent area as edge-attached candidate bodies, and defines the parent body tangent segment, connecting part, extension part and free end area of ​​the candidate body; The feature extraction module uses the cut edge segment of the parent body as a local motion reference to calculate the relative motion data of the connecting part, the extended part and the free end region across frames, and extracts the local differential motion features of the candidate body under airflow disturbance and the unloading and recovery motion features after the airflow is removed. The defect determination module makes a determination based on local differential motion characteristics and unloading recovery motion characteristics. When the following conditions are met: during airflow disturbance, the connecting part of the candidate body remains attached, the extended part produces local deflection constrained by the connecting part, and the free end undergoes relative migration; and after the airflow is removed, the free end produces recovery migration in the opposite direction of the disturbance, and the extended part returns to the front contour of the disturbance, the candidate body is determined to be a sheet-like residual flash.

2. The vehicle exterior defect detection system based on visual recognition according to claim 1, characterized in that: The application direction of the non-contact pulsed airflow is configured to sweep laterally along the cut edge or at a preset angle to the cut edge region to be detected, generating lateral disturbance force on the edge attachment candidate; the dynamic pressure parameter of the pulsed airflow is configured to be within a preset dynamic pressure range, the lower limit of the preset dynamic pressure range is greater than the first critical pressure corresponding to the preset pixel-level deformation displacement generated by the sheet-like residual flash, and the upper limit of the preset dynamic pressure range is less than the second critical pressure corresponding to the preset pixel-level deformation displacement generated by the normal rigid edge structure.

3. The vehicle exterior defect detection system based on vision recognition according to claim 1, characterized in that: The acquisition process of the tangential edge image set includes: before the pulsed airflow establishes a substantial disturbance, acquiring at least one frame as an initial tangential edge image to establish a static reference under no external force; during the peak window period of the continuous pulsed airflow, continuously acquiring multiple frames of disturbed tangential edge images to extract the maximum flexural displacement and deformation trajectory of the overhang and free end regions under airflow stress; after the pulsed airflow stops and the external force is removed, continuously acquiring multiple frames of recovered tangential edge images, the duration of which is configured to cover at least one effective recovery stage of the free end region returning from the disturbed displacement state to the pre-disturbance position state, and continuously acquiring until the offset of the free end region shows a decreasing trend and the overhang contour shows a regression trend, fully recording the gradual recovery process of the free end region returning to the initial contour.

4. The vehicle exterior defect detection system based on vision recognition according to claim 1, characterized in that, The process of determining the edge attachment candidate includes: registering and comparing the actual contour in the initial tangent image with the tangent reference contour; extracting the convex image region located outside the tangent reference contour and having a continuous pixel connection with the actual parent tangent region as the edge attachment candidate; and defining the local edge segment in the actual parent tangent region that directly intersects with the edge attachment candidate in space and forms a continuous structure as the parent tangent segment corresponding to the candidate.

5. The vehicle exterior defect detection system based on vision recognition according to claim 1, characterized in that, The process of defining the parent cut edge segment, connecting portion, extension portion, and free end region of the candidate body includes: establishing a spatial topological extension direction or distance field outward from the parent cut edge segment based on the extracted edge attachment candidate body; defining the root transition region of the edge attachment candidate body that is close to the parent cut edge segment and serves as the starting point for outward extension as the connecting portion; defining the main body region of the edge attachment candidate body that grows outward from the cut edge reference contour along the spatial topological extension direction as the extension portion, starting from the connecting portion; and defining the end boundary region of the edge attachment candidate body located on the side of the extension portion away from the connecting portion as the free end region.

6. The vehicle exterior defect detection system based on vision recognition according to claim 1, characterized in that, The extraction process of the local differential motion features and unloading recovery motion features includes: extracting the contour features of the parent tangent segment as a local coordinate reference between frames of the tangent time sequence image group, and performing relative position compensation and spatial alignment on each frame image; in the airflow disturbance stage image after position compensation and spatial alignment, determining the first relative displacement of the connecting part relative to the local coordinate reference, and the second relative displacement of the free end region relative to the connecting part, and determining the state in which the amplitude of the first relative displacement is less than a first set threshold and the amplitude of the second relative displacement is greater than the amplitude of the first relative displacement, forming a differential motion relationship between the connecting part and the free end region, as the local differential motion feature; in the recovery stage image after position compensation and spatial alignment, continuously extracting the temporal change data of the free end region and the extension relative to the local coordinate reference, and determining the temporal dynamic change of the recovery displacement vector of the free end region relative to the displacement vector of the airflow disturbance stage, and the remaining offset amplitude of the free end region relative to the position before the disturbance decreasing frame by frame, while the contour of the extension approaches the corresponding contour in the initial tangent image frame by frame, as the unloading recovery motion feature.

7. The vehicle exterior defect detection system based on visual recognition according to claim 1, characterized in that: The process in which the connecting part of the candidate remains attached during airflow disturbance, the extended part undergoes local deflection constrained by the connecting part, and the free end undergoes relative migration includes: extracting the coordinate offset or contour overlap of the connecting part relative to the parent body's cut edge segment in a cross-frame image; determining that the connecting part remains attached when the coordinate offset is less than a preset holding threshold or the contour overlap is greater than a preset ratio; extracting the contour curvature data or morphological skeleton angle data of the extended part in a cross-frame image; determining that the extended part undergoes local deflection constrained by the connecting part when the curvature change or angle change during airflow is greater than a preset deformation threshold; calculating the displacement vector of the free end region relative to the connecting part; determining that the free end undergoes relative migration when the magnitude of the displacement vector is greater than a preset migration threshold.

8. The vehicle exterior defect detection system based on vision recognition according to claim 1, characterized in that, The determination process for the extension to regress to the front contour of the disturbance includes: calculating the recovery displacement vector of the free end region in the recovered images of multiple consecutive frames; when the angle between the recovery displacement vector and the displacement vector obtained during the airflow disturbance stage falls within a preset reverse angle range, it is determined that the free end has generated a recovery migration opposite to the disturbance direction; extracting the contour features of the extension in the recovered images of multiple consecutive frames; when the image matching degree between the extension contour and the corresponding contour in the initial truncated image increases frame by frame, or the spatial distance between the contour boundaries decreases frame by frame, it is determined that the extension regresses to the front contour of the disturbance.

Images