Method for classifying and identifying surface defects of a seal

Abstract

This invention relates to the field of machine vision inspection, specifically to a method for classifying and identifying surface defects in sealing components. The method includes: first, acquiring a surface image and constructing a surface coordinate mapping and functional area distribution map; then, collecting surface responses at each detection location under different sampling states and extracting reflection response features, morphological response features, and mechanical response features; subsequently, eliminating normal response deviations based on the normal response spectrum of functional areas to generate abnormal evidence regions, and completing defect classification accordingly; when classification results conflict, performing targeted supplementary sampling and updating the abnormal evidence regions before reclassification. This invention can improve classification accuracy, reduce the false positive rate, and enhance detection stability.

Description

Technical Field

[0001] This invention relates to the field of machine vision inspection, specifically to a method for classifying and identifying surface defects in sealing components. Background Technology

[0002] The quality of seals directly affects the reliability of assembly seals, the lifespan of the entire machine, downtime maintenance costs, and production consistency. The manufacturing process requires consistent assessment of surface anomalies to prevent defective products from entering subsequent assemblies. Existing solutions often rely on single-shot imaging, fixed thresholds, or direct classification, which are insufficient in distinguishing structural zoning differences, localized reflections, morphological disturbances, and stress changes. This can easily lead to misjudging normal fluctuations as anomalies or causing classification conflicts between similar defects, increasing the burden of re-inspection and resulting in insufficient stability of online detection. Summary of the Invention

[0003] This invention provides a method for classifying and identifying surface defects in sealing components, which at least solves the problem of how to stably distinguish between normal response fluctuations and real anomalies in different structural regions and under different sampling conditions, and how to reliably re-determine when there are classification conflicts.

[0004] A method for classifying and identifying surface defects in a sealing component, the method comprising: Acquire a surface image of the seal to be tested, and construct a surface coordinate mapping and functional area distribution map based on the surface image; Based on surface coordinate mapping, the surface response at each detection location is collected under different sampling conditions, and the reflection response features, morphological response features and mechanical response features are extracted. Based on the normal response map of the functional area constructed from defect-free samples, the reflection response characteristics, morphological response characteristics and mechanical response characteristics are compared, and the response deviations falling into the corresponding range of the normal response map of the functional area are eliminated to generate abnormal evidence areas. Defects are classified based on the response features and morphological features of the anomalous evidence region to obtain the defect category and defect location. In the event of conflicting defect classification results, targeted supplementary sampling is performed, and the anomalous evidence region is updated based on the targeted supplementary sampling results before defect classification is performed again.

[0005] In one possible implementation, a surface coordinate mapping and functional area distribution map are constructed based on the surface image, including: extracting the boundary contour of the seal to be tested; establishing the correspondence between the surface position and the coordinate position according to the boundary contour to generate the surface coordinate mapping; and dividing the functional areas according to the structural positions corresponding to each coordinate position in the surface coordinate mapping to generate the functional area distribution map.

[0006] In one possible implementation, the surface response at each detection location is acquired under different sampling states, including: acquiring first and second surface images corresponding to different polarization states at each detection location under different polarization states; acquiring multi-directional surface images corresponding to different illumination directions at each detection location under different illumination directions; and acquiring surface images before loading, during loading, and after loading at each detection location under mechanical loading conditions.

[0007] In one possible implementation, the extraction of reflection response features, morphological response features, and mechanical response features includes: determining reflection response features based on the brightness difference between the first surface image and the second surface image; determining morphological response features based on the brightness distribution difference of the multi-directional surface images; and determining mechanical response features based on the displacement changes of the surface images before loading, during loading, and after loading.

[0008] In one possible implementation, the normal response map of the functional area is constructed based on multiple defect-free samples and includes the normal response range of reflection, normal response range of morphology and normal response range of each functional area at each detection location.

[0009] In one possible implementation, the reflection response characteristics, morphological response characteristics, and mechanical response characteristics are compared to eliminate response deviations that fall within the range corresponding to the normal response spectrum of the functional area. This includes: determining the functional area corresponding to each detection position based on the functional area distribution map and surface coordinate mapping; obtaining the normal reflection response range, normal morphological response range, and normal mechanical response range of the functional area at the detection position from the normal response spectrum of the functional area; comparing the reflection response characteristics, morphological response characteristics, and mechanical response characteristics with the normal reflection response range, normal morphological response range, and normal mechanical response range, respectively; and determining response deviations that exceed the corresponding normal response range as abnormal responses.

[0010] In one possible implementation, generating anomaly evidence regions includes: performing spatial connectivity analysis on the anomalous response to obtain candidate anomalous regions; eliminating isolated candidate anomalous regions based on the continuity between adjacent detection locations; and determining the remaining candidate anomalous regions as anomalous evidence regions.

[0011] In one possible implementation, defect classification is performed based on the response features and morphological features of the anomalous evidence region, including: determining at least two candidate defect categories and the classification confidence corresponding to the candidate defect categories based on the response features of the anomalous evidence region; and filtering the candidate defect categories based on the morphological features of the anomalous evidence region to determine the defect category; wherein the morphological features include region length, region width, and boundary continuity.

[0012] In one possible implementation, in the event of a conflict in defect classification results, targeted supplementary sampling is performed, including: determining a conflict in defect classification results when the difference in classification confidence between the two candidate defect categories with the highest classification confidence is less than or equal to a preset threshold; determining the sampling state corresponding to the targeted supplementary sampling based on the response features used to distinguish the two candidate defect categories; and performing supplementary sampling on the detection location corresponding to the abnormal evidence region in the sampling state.

[0013] In one possible implementation, after updating the anomalous evidence region based on the directional supplemental sampling results, the defect is reclassified, including: extracting supplemental response features based on the directional supplemental sampling results; fusing the supplemental response features with the reflection response features, morphological response features, and mechanical response features on which the anomalous evidence region was generated, and updating the anomalous evidence region; and reclassifying the defect based on the updated anomalous evidence region.

[0014] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: By constructing surface coordinate mapping and functional area distribution maps, unified positioning and zoning determination of different detection locations were achieved; by extracting reflection response features, morphological response features and mechanical response features under different sampling states, multidimensional characterization of abnormal information was achieved; by introducing normal response maps of functional areas, normal fluctuation elimination and abnormal evidence area extraction were achieved; and by using targeted supplementary sampling techniques, classification conflict resolution and stable result output were achieved. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the execution flow of the invention method; Figure 2 This is a schematic diagram of surface coordinate mapping and functional area distribution in a specific embodiment of the present invention; Figure 3 This is a comparison diagram of three types of response characteristics in the abnormal region in a specific embodiment of the invention; Figure 4 This is a comparison chart of the normal response range and the overall deviation value in a specific embodiment of the invention; Figure 5 A comparison chart of classification confidence levels before and after sampling is provided for a specific embodiment of the invention. Detailed Implementation

[0016] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0017] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0018] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0019] Surface defects typically refer to localized abnormal states formed on or near the outer surface of a material. These defects manifest as scratches, indentations, cracks, protrusions, depressions, deposits, edge damage, and residual local deformation. These anomalies not only differ in morphology but are also often accompanied by changes in reflectivity, boundary continuity, and stress response. Different types of surface defects may appear similar in images, while the same type of surface defect may exhibit different response characteristics in different structural regions. Therefore, accurate differentiation based solely on a single appearance is difficult. Based on this, this invention establishes a processing method for surface defect identification and classification, focusing on the response differences of the sealing surface at different detection locations and under different sampling conditions.

[0020] like Figure 1 As shown, a method for classifying and identifying surface defects in a sealing component is provided. The method includes: Acquire a surface image of the seal to be tested, and construct a surface coordinate mapping and functional area distribution map based on the surface image; In this embodiment, a surface image of the seal to be tested is first acquired, and a surface coordinate mapping and functional area distribution map are constructed based on the surface image. Specifically, an industrial camera can image the outer surface of the seal from a fixed viewing angle to obtain raw image data covering the area to be detected. Subsequently, based on the boundary position and structural distribution of the seal in the image, the surface position in the image is converted into a uniformly describable coordinate position, forming a surface coordinate mapping. Further, the surface of the seal is divided into zones according to the actual structural positions corresponding to different coordinate positions to obtain a functional area distribution map. In this way, a stable correspondence can be established between the pixel positions in the original image and the actual positions on the surface of the seal, providing a unified positional reference for subsequent response acquisition, feature extraction, and anomaly detection at different detection positions.

[0021] The process of constructing a surface coordinate mapping and functional area distribution map based on surface images includes: extracting the boundary contour of the seal to be tested; establishing the correspondence between surface position and coordinate position based on the boundary contour to generate a surface coordinate mapping; and dividing functional areas according to the structural position corresponding to each coordinate position in the surface coordinate mapping to generate a functional area distribution map.

[0022] In one embodiment, the process of constructing a surface coordinate map and functional area distribution map based on a surface image includes the following: First, the acquired surface image undergoes basic processing to eliminate interference from non-uniform brightness and local noise generated during the acquisition process on boundary recognition. This basic processing may include brightness normalization, local smoothing, and contrast correction.

[0023] After basic image processing, the boundary contour of the seal under test is extracted from the surface image. For annular seals, the inner and outer boundaries can be identified first, and a closed contour is formed based on the continuous distribution of boundary points. For strip-shaped or irregularly shaped seals, boundary points can be continuously tracked along the outer edge to form a boundary line consistent with the actual contour. When a local boundary is interrupted by reflections, shadows, or stains, the adjacent boundary points on both sides of the interruption can be used to complete the continuity, ensuring the integrity of the boundary contour. After the boundary contour is determined, the correspondence between the surface position and coordinate position is established based on the boundary contour, generating a surface coordinate mapping.

[0024] Specifically, the boundary contour can be used as a positional reference. Image regions within the boundary are sequentially numbered, and the surface region is unfolded according to a preset direction, so that each image position corresponds to a specific coordinate position. For annular seals, circumferential coordinates can be established along the circumference and radial coordinates along the radial direction; for strip seals, longitudinal coordinates can be established along the length and transverse coordinates along the width. In this way, any point in the surface image can be mapped to a corresponding position in a unified coordinate system. After the surface coordinate mapping is established, functional areas are divided according to the structural positions corresponding to each coordinate position in the surface coordinate mapping, generating a functional area distribution map.

[0025] The functional zones can be divided based on the actual structural purpose and surface morphology of the seal. For example, areas related to sealing contact can be designated as the sealing working area, areas near the edges as edge transition areas, and areas not directly involved in the sealing function as non-working areas. If the seal surface has fixed process boundaries, such as the area near the parting line, it can also be separately designated as the process influence area. After the functional zones are divided, each coordinate position in the functional zone distribution map carries a corresponding area attribute. In this way, when collecting surface responses at different detection positions, the detection position can be directly located based on the surface coordinate mapping, and the structural region to which the detection position belongs can be determined according to the functional zone distribution map, thus ensuring that subsequent processing is carried out with a unified position reference and a unified structural semantics.

[0026] Based on surface coordinate mapping, the surface response at each detection location is collected under different sampling conditions, and the reflection response features, morphological response features and mechanical response features are extracted. In this embodiment, based on surface coordinate mapping, corresponding data are collected at each detection location under different sampling states, forming multiple sets of surface response data at the same detection location. Specifically, the spatial order of the detection locations is first determined according to the surface coordinate mapping, and then the polarization state, illumination direction, and mechanical loading state are switched sequentially according to a preset sampling order, repeatedly collecting data at the same detection location. After collection, the image data corresponding to each detection location under different sampling states are organized accordingly, and reflection response features, morphological response features, and mechanical response features are extracted respectively. In this way, multiple types of features reflecting differences in surface reflection, local morphology, and stress changes can be obtained simultaneously at the same detection location, providing a unified data foundation for subsequent anomaly evidence identification.

[0027] The surface response at each detection location is acquired under different sampling conditions, including: acquiring first and second surface images corresponding to different polarization states at each detection location under different polarization states; acquiring multi-directional surface images corresponding to different illumination directions at each detection location under different illumination directions; and acquiring surface images before loading, during loading, and after loading at each detection location under mechanical loading conditions.

[0028] In one embodiment, when acquiring the surface response at each detection location under different sampling states, the current detection location is first determined based on surface coordinate mapping, and the imaging device is kept in correspondence with the current detection location. Subsequently, a first surface image and a second surface image corresponding to different polarization states are acquired at each detection location under different polarization states. This process can be accomplished by switching the polarizer direction or switching the analyzer direction, so that the first surface image and the second surface image correspond to the surface response at the same detection location under two polarization conditions.

[0029] To avoid the impact of detection position offset on subsequent difference analysis, the first and second surface images are preferably acquired consecutively within a short period, and position alignment is performed after acquisition. After completing polarization state acquisition, multi-directional surface images corresponding to different illumination directions are acquired for each detection position under different illumination directions. Specifically, multiple illumination units can be switched sequentially according to preset directions, so that light shines on the same detection position from different directions, and the image results under each illumination direction are recorded. For annular seals, multiple illumination directions can be set along the circumferential and radial directions; for strip seals or irregularly shaped seals, multiple illumination directions can be set along the length and width directions. Multi-directional surface images are used to reflect the brightness and shadow changes of the same detection position under different incident conditions. After acquiring images under different illumination directions, surface images before loading, during loading, and after loading are acquired for each detection position under mechanical loading. This process can be achieved by applying a small, recoverable external force to the seal by a loading mechanism, causing slight deformation of the seal surface without damaging the structure. The surface image before loading is used to record the initial state of the detection position, the surface image during loading is used to record the instantaneous state under the action of external force, and the surface image after loading is used to record the recovery state after the external force is removed.

[0030] To ensure comparability among the three images, the detection position remained constant during acquisition, and the imaging distance and sampling sequence were synchronously controlled. After acquisition, image data obtained from the same detection position under different sampling states were grouped and stored according to the detection position, ensuring that each detection position corresponds to a complete set of sampling data. This guarantees that subsequent extraction of reflection response features, morphological response features, and mechanical response features for the same detection position is based on associated images from the same location, avoiding judgment bias caused by mixing data from different locations.

[0031] Extracting reflection response features, morphological response features, and mechanical response features includes: determining reflection response features based on the brightness difference between the first surface image and the second surface image; determining morphological response features based on the brightness distribution difference of multi-directional surface images; and determining mechanical response features based on the displacement changes of the surface images before loading, during loading, and after loading.

[0032] In one embodiment, after grouping images at the same detection location, the reflection response characteristics are first determined based on the brightness difference between the first and second surface images. This process uses the same detection location as the center, performs brightness statistics on corresponding areas in the first and second surface images, and compares the magnitude and direction of brightness changes under two polarization conditions. If the brightness change at the same detection location is significant under both polarization conditions, it indicates that the detection location is more sensitive to polarization changes, and there is a large difference in surface reflection. If the brightness change is small, it indicates that the surface reflection at the detection location is relatively stable. To reduce the influence of local noise, a small neighborhood can be selected around the detection location, and the average or weighted brightness of the pixels within the neighborhood can be compared. The comparison result is then used as the reflection response characteristics.

[0033] Next, based on the differences in brightness distribution across surface images from multiple directions, morphological response features are determined. Specifically, the brightness distribution at the same detection location under different illumination directions can be compared one by one to identify changes in brightness, edge shadows, and the movement of locally highlighted areas. If significant brightness asymmetry, shadow extension, or edge protrusion occurs at the same detection location under different illumination directions, it can be concluded that there are significant surface morphological differences at that location. Furthermore, the order of brightness changes under multiple illumination directions can be organized to form morphological response features reflecting local protrusions, depressions, scratches, or edge transition states.

[0034] Subsequently, the mechanical response characteristics are determined based on the displacement changes in the surface images before, during, and after loading. This process first aligns the same detection location in the three images, then compares the relative displacement and deformation changes of corresponding areas in the images before and after loading. The surface image during loading primarily reflects the immediate changes under stress, while the surface image after loading primarily reflects the recovery after the external force is removed. If the same detection location still retains significant residual displacement after loading, or if the recovery process is inconsistent with the surrounding area, it indicates that the detection location has different response characteristics under stress compared to the surrounding area.

[0035] To ensure the stability of displacement changes, displacement changes in the surrounding neighborhood of the detection location can be included in the analysis to determine whether the changes in the local area are continuous. After extracting the three types of features, the reflection response features, morphological response features, and mechanical response features are correlated according to the detection location, forming a feature data set consistent with the surface coordinate mapping. Subsequently, when comparing normal responses in functional areas, the three types of features from the same detection location can be directly used for joint judgment, thus ensuring clear and consistent data input for subsequent abnormal evidence identification processes.

[0036] Based on the normal response map of the functional area constructed from defect-free samples, the reflection response characteristics, morphological response characteristics and mechanical response characteristics are compared, and the response deviations falling into the corresponding range of the normal response map of the functional area are eliminated to generate abnormal evidence areas. In this embodiment, after extracting the reflection response features, morphological response features, and mechanical response features, the three types of features corresponding to each detection location are compared based on the pre-established normal response map of the functional area of ​​the defect-free sample. During the comparison, the functional area distribution map and surface coordinate mapping are used as the basis for the location index. First, the functional area to which the current detection location belongs is determined, and then the normal response range of the corresponding location is called to judge each of the three types of features. Response deviations falling within the normal response range are discarded as normal fluctuations, while response deviations exceeding the normal response range are retained as abnormal responses. Subsequently, the abnormal responses of each detection location are spatially organized and analyzed for continuity to form an abnormal evidence region. In this way, normal structural differences, normal process fluctuations, and real abnormal changes can be distinguished, so that subsequent defect classification is based on the already screened abnormal regions.

[0037] The normal response map of the functional area is constructed based on multiple defect-free samples and includes the normal response range of reflection, normal response range of morphology and normal response range of each functional area at each detection location.

[0038] In one embodiment, the normal response map of the functional area is constructed based on multiple defect-free samples and includes the normal reflection response range, normal morphological response range, and normal mechanical response range of each functional area at each detection location. Specifically, multiple defect-free samples of the same model, structural specification, material type, or within the same production process window are first selected as modeling samples. For each defect-free sample, a surface image is acquired in the aforementioned manner, surface coordinate mapping and functional area distribution map are constructed, and reflection response features, morphological response features, and mechanical response features are extracted under the same sampling conditions. Subsequently, the multiple defect-free samples are aligned according to a unified surface coordinate mapping, so that the same coordinate position corresponds to the same physical region in different samples.

[0039] After completing the position alignment, the three types of feature data for each detection location are organized according to functional areas. For each detection location in each functional area, the distribution of reflection response features, morphological response features, and mechanical response features across multiple defect-free samples is statistically analyzed, and the corresponding normal response range is established accordingly. The normal response range can be determined based on the upper and lower limits of the sample data, the concentration range, or the stable fluctuation range after removing discrete points. If a defect-free sample shows a significant deviation at a particular detection location, and this deviation is inconsistent with that of the majority of samples, this deviation can be excluded as abnormal modeling noise to avoid including individual abnormal fluctuations in the normal spectrum.

[0040] After statistical analysis, an index relationship is established based on functional zones and detection locations. The normal response ranges for reflection, morphology, and mechanics are written into the corresponding data structures to form a functional zone normal response map. This way, different detection locations within the same functional zone have their own independent normal response boundaries, and the normal fluctuation differences between different functional zones can also be preserved. During subsequent detection, the corresponding normal response range can be directly retrieved for comparison based on the functional zone to which the current detection location belongs and its position in the surface coordinate mapping, without needing to reconstruct the baseline. By pre-establishing this location-based and zoned normal response map, the natural response differences between edge areas, transition areas, and working areas can be avoided from being mixed up, and the judgment boundary can be avoided from being too narrow or too wide due to relying on a single sample as a baseline.

[0041] The reflection response characteristics, morphological response characteristics, and mechanical response characteristics are compared, and response deviations falling within the corresponding range of the normal response spectrum of the functional area are eliminated. This includes: determining the functional area corresponding to each detection position based on the functional area distribution map and surface coordinate mapping; obtaining the normal reflection response range, normal morphological response range, and normal mechanical response range of the functional area at the detection position from the normal response spectrum of the functional area; comparing the reflection response characteristics, morphological response characteristics, and mechanical response characteristics with the normal reflection response range, normal morphological response range, and normal mechanical response range, respectively; and identifying response deviations exceeding the corresponding normal response range as abnormal responses.

[0042] In one embodiment, when comparing reflection response features, morphological response features, and mechanical response features, the functional area corresponding to each detection position is first determined based on the functional area distribution map and surface coordinate mapping. Specifically, starting from the coordinate position of the current detection position, the corresponding region label can be queried in the functional area distribution map to obtain the structural region to which the current detection position belongs.

[0043] After determining the functional area, the normal response ranges for reflection, morphology, and mechanics at the current detection location are read from the normal response map of that functional area. After reading, the reflection response features extracted at the current detection location are compared with the normal response ranges, the morphology response features with the normal response ranges, and the mechanics response features with the normal response ranges. During the comparison, each feature value can be checked individually to see if it falls within its corresponding range. Alternatively, multiple feature results obtained from continuous sampling can be processed first, and then the processed result can be compared with the normal response range. If the reflection response feature falls within the normal response range, it indicates that the current detection location exhibits normal fluctuations in reflection; if it exceeds the normal response range, the corresponding reflection response deviation is retained. The morphology response features and mechanics response features are processed using the same logic.

[0044] After completing the independent comparison of the three types of features, all comparison results for the current detection location are summarized. If one or more types of features exceed the normal response range, the response deviation exceeding the corresponding normal response range is determined as an abnormal response; if all three types of features fall within the corresponding normal response range, the current detection location does not retain an abnormal response. To ensure the stability of the judgment results, a data consistency check can be added before the comparison, such as checking whether all three types of features at the current detection location have been successfully extracted. If a certain type of feature is missing, a supplementary judgment can be made based on the same type of feature at adjacent detection locations, or the current detection location can be marked as a location to be reviewed. To avoid natural fluctuations at the functional area boundaries being directly judged as abnormal, detection locations near the functional area boundaries can also be reviewed by simultaneously referencing adjacent functional areas. That is, the normal response ranges of the current functional area and adjacent functional areas at the corresponding locations are read simultaneously, and the response deviation is retained as an abnormal response only when the current feature exceeds both reference ranges simultaneously. By using this comparison method that unfolds layer by layer according to functional area, detection location, and feature type, it can be ensured that abnormal responses come from real out-of-boundary changes, rather than from normal differences between different regions, occasional sampling fluctuations, or boundary transition changes.

[0045] The process of generating anomalous evidence regions includes: performing spatial connectivity analysis on anomalous responses to obtain candidate anomalous regions; eliminating isolated candidate anomalous regions based on the continuity between adjacent detection locations; and identifying the remaining candidate anomalous regions as anomalous evidence regions.

[0046] In one embodiment, after determining the abnormal response, an abnormal evidence region is further generated. Specifically, the abnormal responses retained at each detection location are first mapped back to their corresponding positions in the surface coordinate mapping, forming an abnormal response distribution result. Subsequently, spatial connectivity analysis is performed on the abnormal responses. Spatial connectivity analysis can be performed according to the adjacency relationship in the surface coordinate mapping, that is, determining whether abnormal responses exist simultaneously between the current detection location and its neighboring detection locations.

[0047] For annular seals, adjacent positions can be inspected simultaneously along the circumferential and radial directions; for strip-shaped or irregularly shaped seals, adjacent positions can be inspected along the length and width directions. If multiple adjacent detection positions simultaneously exhibit abnormal responses, and these abnormal responses are spatially continuous, these detection positions are merged into a single candidate abnormal region. If a detection position exhibits an abnormal response alone and does not form a continuous relationship with surrounding detection positions, it is first marked as an isolated candidate abnormal region.

[0048] After initial aggregation, isolated candidate anomaly regions are eliminated based on their continuity between adjacent detection locations. Continuity is assessed in two ways: firstly, by checking the spatial continuity length—whether the candidate anomaly region remains anomalous across multiple consecutive detection locations; and secondly, by checking the consistency of anomaly types—whether anomalies at adjacent detection locations originate from the same or similar feature deviations. If an isolated candidate anomaly region neither meets the minimum continuity length requirement nor maintains a consistent anomaly type with its surrounding regions, it is considered sporadic noise, local interference, or transient fluctuation and is eliminated from subsequent evidence. Candidate anomaly regions that are continuous with their surrounding regions are retained, and their boundaries are determined.

[0049] When determining the boundary, the outer boundary of the region can be formed according to the range of the outermost continuous anomaly detection positions, and the set of detection positions covered by the region can be recorded. After the elimination and boundary determination are completed, the remaining candidate anomaly regions are determined as anomaly evidence regions. After the anomaly evidence regions are formed, each anomaly evidence region not only corresponds to a set of spatially continuous detection positions, but also retains the source of the anomaly response that constitutes the region. When classifying defects in the future, the judgment can be made directly based on the response feature distribution and regional morphological features within the anomaly evidence region, without having to return to each individual detection position for point-by-point analysis. By spatially organizing and continuously filtering the anomaly responses, the interference of isolated noise points on the classification results can be effectively reduced, while ensuring that truly spatially distributed anomaly regions are stably preserved.

[0050] Defects are classified based on the response features and morphological features of the anomalous evidence region to obtain the defect category and defect location. In the event of conflicting defect classification results, targeted supplementary sampling is performed, and the anomalous evidence region is updated based on the targeted supplementary sampling results before defect classification is performed again.

[0051] In this embodiment, after forming an anomalous evidence region, defect classification is performed based on the response features and morphological features of the anomalous evidence region, and the defect category and location are output. Specifically, the reflection response features, morphological response features, and mechanical response features retained within the anomalous evidence region are first used to determine multiple candidate defect categories corresponding to the current anomalous evidence region. Then, the spatial morphology of the anomalous evidence region is combined to filter the candidate results, obtaining the final classification result corresponding to the current anomalous evidence region. When there is a conflict in the current classification result, the current result is not directly output; instead, targeted supplementary sampling is performed on the current anomalous evidence region. After supplementary sampling is completed, supplementary response features are extracted and used to update the current anomalous evidence region. The updated anomalous evidence region is then reclassified for defects. Through this processing method, the final output defect category can be based on the anomalous information after supplementary verification.

[0052] Defect classification based on response features and regional morphological features of anomalous evidence regions includes: determining at least two candidate defect categories and their corresponding classification confidence levels based on the response features of anomalous evidence regions; and filtering candidate defect categories based on the regional morphological features of anomalous evidence regions to determine the defect category. The regional morphological features include region length, region width, and boundary continuity.

[0053] In one embodiment, when classifying defects based on the response features and morphological features of anomalous evidence regions, a region-level feature set is first established for each anomalous evidence region. The region-level feature set includes two parts: response features and morphological features. Response features are derived from the reflection response features, morphological response features, and mechanical response features retained at each detection location within the anomalous evidence region. For the same anomalous evidence region, the three types of features corresponding to all detection locations within the region can be summarized to form a region response description that characterizes the overall change state of the anomalous evidence region. If the change directions of the three types of features at different detection locations within the anomalous evidence region are basically consistent, then the consistent change is taken as the main response direction of the anomalous evidence region; if there are local differences between different detection locations, then the difference distribution is retained and used as part of the internal structural changes within the region in subsequent classification.

[0054] The regional morphological characteristics are determined by the spatial coverage of the anomalous evidence region, including at least the region length, region width, and boundary continuity. Region length characterizes the coverage distance of the anomalous evidence region along the primary extension direction, region width characterizes the coverage range of the anomalous evidence region along the secondary direction, and boundary continuity characterizes whether there are continuous interruptions, abrupt changes, or local breaks at the outer boundary of the anomalous evidence region. After establishing the region-level feature set, at least two candidate defect categories are determined based on the response characteristics of the anomalous evidence region, and the classification confidence score corresponding to each candidate defect category is obtained. Specifically, a correspondence rule between defect categories and response characteristics can be pre-established. For example, anomalous evidence regions with prominent changes in reflection response but weak changes in morphological response can be preferentially classified as surface adhesion candidates; anomalous evidence regions with prominent changes in morphological response and obvious boundary changes can be preferentially classified as surface damage candidates; and anomalous evidence regions with prominent changes in mechanical response and obvious differences before and after loading can be preferentially classified as structural anomaly candidates.

[0055] Based on this, candidate defect categories are further filtered according to the regional morphological characteristics of the anomalous evidence region. If a candidate defect category typically corresponds to a long and narrow distribution, and the length of the current anomalous evidence region is significantly greater than its width, with strong boundary continuity, then that candidate defect category is retained. If a candidate defect category typically corresponds to a blocky distribution, and the current anomalous evidence region is long and narrow, then the priority of that candidate defect category is reduced. After filtering, the defect category corresponding to the current anomalous evidence region is determined, and the defect location is determined based on the coverage position of the anomalous evidence region in the surface coordinate mapping. The defect location can be the coordinate range corresponding to the anomalous evidence region, or the center position of the anomalous evidence region and the region boundary range. In this way, the classification result contains both category information and location information, facilitating subsequent output and verification.

[0056] In the event of a conflict in defect classification results, targeted supplementary sampling is performed, including: determining a conflict in defect classification results when the difference in classification confidence between the two candidate defect categories with the highest classification confidence is less than or equal to a preset threshold; determining the sampling state corresponding to the targeted supplementary sampling based on the response features used to distinguish the two candidate defect categories; and performing supplementary sampling at the detection location corresponding to the abnormal evidence region in the sampling state.

[0057] In one embodiment, when there is a conflict in the defect classification results, targeted supplementary sampling is performed on the current anomalous evidence region. Here, a conflict in defect classification results refers to the inability to reliably distinguish between candidate defect categories corresponding to the current anomalous evidence region using existing classification results.

[0058] In specific judgment, the two candidate defect categories with the highest classification confidence can be read first, and the difference in their corresponding classification confidence values ​​can be compared. When the difference in classification confidence is less than or equal to a preset threshold, it is determined that there is a conflict in the defect classification results in the current abnormal evidence area. The preset threshold can be pre-set according to the actual detection requirements or determined based on the statistical results of historical samples. If the difference in classification confidence is greater than the preset threshold, it means that the classification results of the current abnormal evidence area have a relatively clear distinguishing boundary, and there is no need to enter the supplementary sampling process. If the difference in classification confidence falls within the preset threshold range, it is necessary to determine the sampling state corresponding to the targeted supplementary sampling based on the response characteristics used to distinguish the two candidate defect categories.

[0059] In specific processing, we first analyze the main distinguishing criteria of the two candidate defect categories in historical samples. For example, if the difference between two candidate defect categories in reflection response characteristics is more obvious, then the polarization-related sampling state is preferred; if the difference between two candidate defect categories in morphology response characteristics is more obvious, then the sampling state with more obvious changes in illumination direction is preferred; if the difference between two candidate defect categories in mechanical response characteristics is more obvious, then the mechanical loading state is preferred as a supplementary sampling state.

[0060] After determining the sampling state, supplementary sampling is performed only on the detection positions corresponding to the current anomalous evidence area, without repeating sampling on all detection positions. During the supplementary sampling process, the correspondence between the detection position and the surface coordinate mapping is maintained, and the supplementary sampling process is kept consistent with the original sampling process in terms of imaging position, sampling order, and external conditions, only changing the key sampling state used to distinguish candidate defect categories. In this way, the new data obtained from supplementary sampling is directly comparable to the original data. By performing local sampling on the conflict area, the sampling range can be limited to the current anomalous evidence area, reducing unnecessary data duplication, while ensuring that supplementary sampling directly serves the current conflict judgment without changing the detection results outside the existing anomalous evidence area.

[0061] After updating the anomalous evidence region based on the directional supplemental sampling results, the defects are reclassified, including: extracting supplemental response features based on the directional supplemental sampling results; fusing the supplemental response features with the reflection response features, morphological response features, and mechanical response features on which the anomalous evidence region was generated, and updating the anomalous evidence region; and reclassifying the defects based on the updated anomalous evidence region.

[0062] In one embodiment, after directional supplemental sampling is completed, the abnormal evidence region is updated based on the directional supplemental sampling results, and the defect classification is re-performed after the update. Specifically, the directional supplemental sampling results are first processed using the same data organization process as the original sampling process to extract supplemental response features. The supplemental response features correspond to the supplemental sampling state. If the supplemental sampling state is polarization-related sampling, the supplemental response features mainly reflect the reflection differences under the new polarization conditions; if the supplemental sampling state is illumination direction-related sampling, the supplemental response features mainly reflect the morphological changes under the new illumination conditions; if the supplemental sampling state is mechanical loading-related sampling, the supplemental response features mainly reflect the displacement or recovery changes under the new loading conditions.

[0063] After extraction, the supplementary response features are fused with the reflection response features, morphological response features, and mechanical response features used when generating the anomalous evidence region. During fusion, the supplementary response features are first mapped to the detection positions corresponding to the current anomalous evidence region, and then a one-to-one correspondence is established between the detection positions and the original three types of features. If the supplementary response features and the original three types of features maintain the same direction of change, the supplementary response features are added to the region feature set as enhanced evidence for the current anomalous evidence region; if the supplementary response features are inconsistent with the original three types of features, the differences are retained and used as an important basis for reassessment.

[0064] Subsequently, the anomalous evidence region is updated based on the fused set of regional features. The update may include retaining or removing detection locations within the anomalous evidence region, expanding or shrinking the region's boundaries, and adjusting the main response direction within the region. If supplementary response features indicate that some detection locations in the original anomalous evidence region no longer meet the anomalous conditions, those locations are removed from the current anomalous evidence region; if supplementary response features indicate that the anomalous change extends to adjacent detection locations, those adjacent detection locations are included in the current anomalous evidence region. After the update, defects are reclassified based on the updated anomalous evidence region. During reclassification, candidate defect categories, classification confidence, defect types, and defect locations are recalculated based on the fused response features and the updated regional morphological features.

[0065] If the difference between the candidate defect category with the highest classification confidence and the second candidate defect category exceeds a preset threshold after reclassification, the updated defect category and defect location are output. If conflicts still exist after reclassification, the current abnormal evidence region is marked as a region to be reviewed, and all response features and supplementary response features corresponding to the current region are retained for further judgment. Through this update and reclassification process, the final classification result can be based on more comprehensive regional response information, and the current output result is consistent with the supplementary sampling process.

[0066] In one specific embodiment, fluororubber rotary shaft seals from the same batch were selected for verification. The seal has an outer diameter of 72 mm, an inner diameter of 52 mm, and a visible radial surface width of 8 mm. First, 60 samples, confirmed to be defect-free by manual inspection, were continuously drawn from the production line as normal samples; then, 24 samples with surface abnormalities were drawn as verification samples, with sample number A17 used for detailed description in this embodiment. The imaging device used a 12-megapixel industrial camera with an imaging resolution of 4096 x 3072. A circumferential unfolded length was established at the seal's mid-diameter, and discretized into 720 circumferential positions and 80 radial positions to obtain a 720 x 80 surface coordinate mapping. Each circumferential position corresponds to approximately 0.27 mm arc length, and each radial position corresponds to approximately 0.10 mm surface width. The functional areas are divided according to structural purpose into a sealing working area, a transition area, an edge area, and a process influence area.

[0067] like Figure 2 As shown, surface coordinate mapping was first established based on the boundary contour of sample A17, and then functional zones were divided according to radial position. The lower blue band in the figure represents the sealing working area, the middle green area represents the transition area, the upper yellow area represents the edge area, and the uppermost purple area represents the process influence area. The local area enclosed by the black border is the subsequently identified anomalous evidence area, with a circumferential position ranging from 338 to 347 and a radial position ranging from 14 to 21. This figure was obtained by unfolding the surface image of sample A17 and overlaying the functional zone markings with the final anomalous area boundary. As can be seen from the figure, the anomalous evidence area is located near the boundary between the sealing working area and the transition area, indicating that this area is both at the actual stress location and close to the structural change location, thus possessing high engineering judgment value.

[0068] On sample A17, polarization sampling, directional illumination sampling, and micro-mechanical loading sampling were performed sequentially according to surface coordinate mapping. Polarization sampling acquired two consecutive images at the same detection position; directional illumination sampling acquired four images sequentially along the circumferential positive direction, circumferential negative direction, radial inward direction, and radial outward direction; mechanical loading sampling acquired images before, during, and after loading under a compression displacement of 0.25 mm. Taking the center point of the anomalous evidence area, i.e., circumferential position 343 and radial position 18, as an example, the average brightness of the first polarization image was 184 grayscale values, and the average brightness of the second polarization image was 139 grayscale values, with a brightness difference of 45; the average brightness values ​​under the four illumination directions were 121, 167, 114, and 181, respectively, with a difference of 67 between the maximum and minimum values; the residual displacement after loading relative to before loading was 0.084 mm. The normal response spectrum of the functional area established from 60 normal samples showed that the upper limit of the normal reflection response at the same location within the sealed working area was 26, the upper limit of the normal morphological response was 33, and the upper limit of the normal mechanical response was 0.028 mm. Comparing the values ​​of sample A17 with the normal upper limit, the reflection deviation factor is 45 divided by 26, yielding 1.73; the morphology deviation factor is 67 divided by 33, yielding 2.03; and the mechanical deviation factor is 0.084 divided by 0.028, yielding 3.00. Averaging these three deviation factors gives a comprehensive deviation value of 2.25. This value is significantly higher than the upper limit of 0.97 for comprehensive deviation of normal samples at the same location, indicating that this location has clearly deviated from the normal fluctuation range.

[0069] like Figure 3 As shown, the detection results within the circumferential range of 336 to 350 and the radial range of 14 to 21 were averaged circumferentially to obtain normalized variation curves for reflection response, morphological response, and mechanical response characteristics. This graph is plotted by dividing each characteristic value by its corresponding upper limit of normal; therefore, a value greater than 1 indicates exceeding the normal range. In the graph, the blue line represents the reflection response characteristic, the orange line represents the morphological response characteristic, and the green line represents the mechanical response characteristic. All three curves reach their peaks near the circumferential position of 343, with the peak value for the reflection response characteristic being 1.73, the peak value for the morphological response characteristic being 2.03, and the peak value for the mechanical response characteristic being 3.00. The gray-marked areas correspond to local anomaly zones. As can be seen from the graph, the mechanical response characteristic rises the fastest and has the highest peak, indicating that this anomaly is not simply surface contamination, but more closely resembles actual surface damage with residual material deformation; simultaneously, the reflection response characteristic and the morphological response characteristic rise synchronously, indicating that this area has both abnormal surface brightness and local boundary morphological changes.

[0070] like Figure 4As shown, the average of the three types of deviation multiples is calculated point by point to obtain the comprehensive deviation value curve within the range of circumferential positions 336 to 350, and compared with the normal lower limit and normal upper limit obtained from the statistics of 60 normal samples. The two dashed lines in the figure represent the normal lower limit and normal upper limit, respectively; the light green band in the middle represents the normal fluctuation range; and the red line represents the comprehensive deviation value of sample A17. It can be seen that the deviation values ​​at seven consecutive circumferential positions from 340 to 346 are all higher than the normal upper limit, with circumferential position 343 reaching 2.25, circumferential position 344 reaching 2.15, and circumferential position 342 reaching 1.98. This indicates that the anomaly is not an isolated noise point, but a regional anomaly with obvious spatial continuity. According to the spatial connectivity analysis rules of this invention, a continuous circumferential length of seven positions and a continuous radial width of eight positions are retained and identified as anomaly evidence areas, rather than being discarded as occasional noise.

[0071] During the initial classification, based solely on the response and morphological characteristics of the original anomaly evidence area, the system identified two main candidate categories: scratches and indentations. The classification confidence score for scratches was 0.41, while that for indentations was 0.39, with a difference of 0.02. The preset conflict threshold was set to 0.05, therefore this result was classified as a conflict. Further analysis of the distinguishing factors revealed a more significant difference in recovery after mechanical loading. Therefore, targeted supplementary sampling was triggered, adding a single, repeated loading sample with the same loading amount only to the detection location corresponding to the anomaly evidence area. After supplementary sampling, the residual displacement at the center point increased from 0.084 mm to 0.091 mm, and continuous residual displacement increases were observed at circumferential positions 341 to 345, indicating that this area better matches the characteristics of localized stress concentration and inconsistent recovery caused by scratches. The supplementary response features are then fused with the original reflection response features, morphological response features, and mechanical response features to update the boundary of the anomalous evidence region. The circumferential range remains 338 to 347, while the radial range is slightly adjusted from 14 to 21 to 15 to 21.

[0072] like Figure 5As shown, the classification confidence levels for the same anomalous evidence area changed significantly before and after supplementary sampling. Before supplementary sampling, the classification confidence levels for scratches, indentations, bubbles, and contamination were 0.41, 0.39, 0.12, and 0.08, respectively; after supplementary sampling, the classification confidence levels for scratches, indentations, bubbles, and contamination became 0.76, 0.14, 0.06, and 0.04, respectively. The blue bars in the figure represent the area before supplementary sampling, and the red bars represent the area after supplementary sampling. It can be seen that after supplementary sampling, the classification confidence level for scratches increased by 0.35, while the classification confidence level for indentations decreased by 0.25. The classification conflict was eliminated, and sample A17 was ultimately classified as a scratch defect, with the defect locations being circumferential positions 338 to 347 and radial positions 15 to 21. Upon verification under a 100x microscope, sample A17 was found to have a shallow scratch approximately 2.46 mm long and 0.62 mm wide at the corresponding location, and the verification result was consistent with the output of this invention.

[0073] Furthermore, batch verification was performed on 24 anomalous samples. Without targeted supplemental sampling, 20 out of 23 samples were correctly identified, with a classification accuracy of 20 / 24 (83.33%). With targeted supplemental sampling enabled, all 23 out of 23 samples were correctly identified, with a classification accuracy of 23 / 24 (95.83%). Statistical analysis of 60 normal samples showed that without targeted supplemental sampling, 4 samples were falsely reported, resulting in a correct classification rate of 56 / 60 (93.33%). With targeted supplemental sampling enabled, the false report dropped to 1 sample, and the correct classification rate of 59 / 60 (98.33%). These results demonstrate that the present invention can not only achieve accurate and verifiable anomaly identification on individual samples but also reduce false reports and improve the final classification stability of conflicting samples in batch testing.

[0074] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0075] The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for classifying and identifying surface defects in sealing components, characterized in that, The method includes: Acquire a surface image of the seal to be tested, and construct a surface coordinate mapping and functional area distribution map based on the surface image; Based on the surface coordinate mapping, the surface response at each detection location is collected under different sampling states, and the reflection response features, morphological response features, and mechanical response features are extracted. Based on the normal response map of the functional area constructed from defect-free samples, the reflection response features, the morphological response features and the mechanical response features are compared, and response deviations falling within the corresponding range of the normal response map of the functional area are eliminated to generate abnormal evidence areas. Defects are classified based on the response features and morphological features of the anomalous evidence region to obtain the defect category and defect location. In the event of a conflict in the defect classification results, targeted supplementary sampling is performed, and the anomalous evidence region is updated based on the targeted supplementary sampling results before the defect classification is performed again.

2. The method according to claim 1, characterized in that, The construction of surface coordinate mapping and functional area distribution map based on the surface image includes: Extract the boundary contour of the seal to be tested; Establish the correspondence between surface position and coordinate position based on the boundary contour, and generate the surface coordinate mapping; Functional zones are divided according to the structural positions corresponding to each coordinate position in the surface coordinate mapping, and the functional zone distribution map is generated.

3. The method according to claim 1, characterized in that, The acquisition of surface responses at each detection location under different sampling states includes: Acquire first and second surface images corresponding to different polarization states at each detection location under different polarization states; Acquire multi-directional surface images of each detection location under different lighting directions; Under mechanical loading, surface images before loading, during loading, and after loading were acquired at each detection location.

4. The method according to claim 3, characterized in that, The extraction of reflection response features, morphological response features, and mechanical response features includes: The reflection response characteristics are determined based on the brightness difference between the first surface image and the second surface image; The morphological response features are determined based on the differences in brightness distribution of the multi-directional surface images; The mechanical response characteristics are determined based on the displacement changes of the surface images before loading, during loading, and after loading.

5. The method according to claim 1, characterized in that, The normal response map of the functional area is constructed based on multiple defect-free samples and includes the normal response range of reflection, normal response range of morphology and normal response range of mechanical properties of each functional area at each detection location.

6. The method according to claim 5, characterized in that, The comparison of the reflection response characteristics, the morphological response characteristics, and the mechanical response characteristics, and the elimination of response deviations falling within the range corresponding to the normal response spectrum of the functional region, includes: Based on the functional area distribution map and the surface coordinate mapping, the functional area corresponding to each detection position is determined; Obtain the normal response range of reflection, normal response range of morphology, and normal response range of mechanical properties of the functional area at the detection location from the normal response map of the functional area; The reflection response characteristics, the morphology response characteristics, and the mechanical response characteristics are compared with the normal reflection response range, the normal morphology response range, and the normal mechanical response range, respectively. Response deviations that exceed the corresponding normal response range are identified as abnormal responses.

7. The method according to claim 6, characterized in that, The region for generating abnormal evidence includes: Spatial connectivity analysis is performed on the abnormal response to obtain candidate abnormal regions; Based on the continuity of the candidate abnormal regions between adjacent detection positions, isolated candidate abnormal regions are eliminated; The remaining candidate anomaly regions are identified as the anomaly evidence regions.

8. The method according to claim 1, characterized in that, The defect classification based on the response features and morphological features of the anomalous evidence region includes: Based on the response characteristics of the anomalous evidence region, at least two candidate defect categories and the classification confidence levels corresponding to the candidate defect categories are determined. Based on the regional morphological features of the abnormal evidence region, the candidate defect categories are screened to determine the defect category; The morphological features of the region include region length, region width, and boundary continuity.

9. The method according to claim 8, characterized in that, In the event of conflicting defect classification results, targeted supplementary sampling is performed, including: If the difference in classification confidence between the two candidate defect categories with the highest classification confidence is less than or equal to a preset threshold, the defect classification result is determined to be conflicting. Based on the response features used to distinguish the two candidate defect categories, the sampling state corresponding to the targeted supplementary sampling is determined; In the sampling state, supplementary sampling is performed on the detection location corresponding to the abnormal evidence region.

10. The method according to claim 9, characterized in that, The process of reclassifying defects after updating the abnormal evidence region based on the targeted supplementary sampling results includes: Extract supplementary response features based on the targeted supplementary sampling results; The supplementary response features are fused with the reflection response features, morphological response features, and mechanical response features on which the abnormal evidence region is based, and the abnormal evidence region is updated. Defects are reclassified based on the updated abnormal evidence region.

Images