A production quality identification method based on AI visual detection

Abstract

The application discloses a production quality identification method based on AI visual detection, comprising the following steps: collecting and preprocessing product images of products to be detected on a production line; identifying suspected abnormal areas and extracting corresponding area features by using a YOLOv8-seg model; screening quality uncertain areas according to a suspected abnormal area feature set and a review confidence interval; identifying suspected quality problem types of each quality uncertain area; calling a visual inquiry template library to generate a visual inquiry strategy; controlling a visual acquisition device to perform multi-condition supplementary acquisition on the quality uncertain areas; identifying real quality defect areas and visual interference areas, and outputting a production quality identification result. The application utilizes AI visual detection and multi-condition supplementary acquisition to identify real quality defects and visual interference, and has the advantages of accurate identification, low misjudgment rate and reliable disposal.

Description

Technical Field

[0001] This invention relates to the field of production quality inspection, and in particular to a production quality identification method based on AI visual inspection. Background Technology

[0002] Current production quality inspection typically uses machine vision and deep learning models to automatically identify products on the production line. This involves acquiring product images through a fixed light source, fixed shooting angle, preset exposure parameters, and a single spectral channel. After processing such as noise reduction, localization, and segmentation, target detection or instance segmentation models are used to identify suspected abnormal areas such as scratches, stains, cracks, dents, and assembly deviations. Based on the area location, mask boundary, and defect confidence level output by the model, a quality judgment result is generated.

[0003] However, in actual production environments, factors such as product materials, surface textures, transparent structures, curved surface reflections, and assembly gaps can cause complex variations in image representation. Visual interferences such as reflections, shadows, and sudden brightness changes can easily be misjudged as real defects, while subtle defects such as shallow scratches and transparent bubbles may be missed due to poor acquisition conditions. Existing technologies lack typified verification mechanisms for areas of uncertain quality, and cannot proactively adjust the light source, polarization, viewing angle, exposure, focal length, or spectral channels for supplementary acquisition based on suspected problem types. It is also difficult to determine the persistence, location consistency, and illumination change characteristics of abnormal areas through multi-condition visual responses, resulting in insufficient detection confidence and low accuracy in product qualification status and handling instructions.

[0004] Therefore, how to provide a production quality identification method based on AI visual inspection is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a production quality identification method based on AI visual inspection. This invention utilizes AI visual inspection and multi-condition supplementary data acquisition to identify real quality defects and visual interference, and has the advantages of accurate identification, low misjudgment rate and reliable handling.

[0006] A production quality identification method based on AI visual inspection according to an embodiment of the present invention includes the following steps:

[0007] Collect product images of the products to be inspected on the production line under initial visual acquisition conditions, perform preprocessing on the product images, and generate initial visual data of the products;

[0008] Input the initial visual data of the product into the YOLOv8-seg model to identify suspected abnormal areas in the product image and extract the regional features of the suspected abnormal areas to generate a suspected abnormal area feature set.

[0009] Based on the feature set of suspected abnormal regions and the verification confidence interval, regions with uncertain quality are screened to generate a set of regions with uncertain quality.

[0010] Based on the regional characteristics of each quality uncertainty region in the set of quality uncertainty regions, identify the suspected quality problem types corresponding to each quality uncertainty region;

[0011] Based on the suspected quality problem type, the visual inquiry template library is called to determine the visual inquiry target, visual acquisition parameter combination, number of supplementary acquisitions and order of supplementary acquisitions for each quality uncertainty area, and a visual inquiry strategy is generated.

[0012] According to the visual inquiry strategy, the visual acquisition device is controlled to perform multi-condition supplementary acquisition on areas with uncertain quality, and a multi-condition visual response sequence is generated.

[0013] Based on multi-condition visual response sequence identification, the system identifies real quality defect areas and visual interference areas, and outputs production quality identification results.

[0014] Optionally, the initial visual acquisition conditions include fixed light source orientation, light source intensity, shooting angle, exposure parameters, focal length parameters, and spectral channels, and the preprocessing includes image denoising, distortion correction, brightness normalization, product area positioning, and background area removal.

[0015] Optionally, the generation of the suspected abnormal region feature set specifically includes:

[0016] The initial visual data of the product is processed by input size matching, pixel value normalization and tensor quantization to form a visual tensor to be identified;

[0017] The visual tensor to be identified is input into the YOLOv8-seg model, and feature extraction is performed through the backbone network to obtain multi-scale product visual features that include texture details, edge structure and regional semantics.

[0018] Multi-scale product visual features are fused using a feature fusion network to form a fused feature map;

[0019] The detection branch is used to predict candidate boxes for abnormal regions in the fused feature map, and candidate abnormal boxes and candidate abnormal confidence scores are obtained.

[0020] By segmenting the branch, anomaly region mask prediction is performed on the fused feature map to obtain the candidate anomaly mask corresponding to the candidate anomaly box.

[0021] Candidate anomaly boxes and candidate anomaly masks are filtered based on the confidence level of the candidate anomalies, and overlapping areas are removed from the filtered candidate anomaly boxes and candidate anomaly masks to identify suspected anomaly areas.

[0022] The location of the suspected abnormal region is determined based on the candidate abnormal bounding box corresponding to the suspected abnormal region, and the boundary and area of ​​the suspected abnormal region are determined based on the candidate abnormal mask corresponding to the suspected abnormal region.

[0023] The pixel distribution, grayscale gradient, edge direction, and region contour within the suspected abnormal region are statistically analyzed to obtain texture continuity, brightness distribution, edge morphology, region appearance, and grayscale variation. These are then correlated with the location, boundary, area, and initial defect confidence of the suspected abnormal region to generate a feature set of the suspected abnormal region.

[0024] Optionally, the generation of the set of regions with uncertain quality specifically includes:

[0025] Configure a review confidence interval, which includes a lower review confidence limit and a higher review confidence limit, wherein the lower review confidence limit is less than the higher review confidence limit;

[0026] The initial defect confidence level corresponding to each suspected abnormal region is compared with the lower limit of the review confidence level and the upper limit of the review confidence level. When the initial defect confidence level corresponding to the suspected abnormal region is greater than or equal to the lower limit of the review confidence level and less than or equal to the upper limit of the review confidence level, the corresponding suspected abnormal region is determined as a quality uncertainty region.

[0027] Extract the regional features corresponding to the quality uncertain regions from the feature set of suspected abnormal regions, associate the quality uncertain regions with the corresponding regional features, and generate a set of quality uncertain regions.

[0028] Optionally, the identification of suspected quality problem types specifically includes:

[0029] The regional features corresponding to each quality uncertainty region in the set of quality uncertainty regions are normalized to obtain the type discrimination features corresponding to each quality uncertainty region;

[0030] Configure the type discrimination conditions corresponding to reflective interference, shadow interference, shallow scratches, dents, transparent bubbles, stains, cracks and assembly deviations respectively;

[0031] The type discrimination features corresponding to each quality uncertainty region are matched with each type discrimination condition. The number of feature terms satisfied by the type discrimination features in each type discrimination condition is counted. The type matching strength of each quality uncertainty region relative to each suspected quality problem type is determined based on the number of feature terms.

[0032] The suspected quality problem type with the strongest type matching strength is identified as the suspected quality problem type in the corresponding quality uncertainty region, and the quality uncertainty region is associated with the corresponding suspected quality problem type.

[0033] Optionally, the generation of the visual inquiry strategy specifically includes:

[0034] A visual inquiry template library is established, which includes visual inquiry template items corresponding to reflective interference, shadow interference, shallow scratches, dents, transparent bubbles, stains, cracks and assembly deviations, respectively. Each visual inquiry template item records the target discrimination content, the category of visual acquisition parameters to be adjusted, the parameter value range, the parameter change direction and parameter combination constraints.

[0035] Read the suspected quality problem types associated with each quality uncertainty region in the quality uncertainty region set, and call the corresponding visual inquiry template items from the visual inquiry template library according to the suspected quality problem type to generate template call results;

[0036] Based on the target discrimination content in the template call results, determine the visual inquiry target corresponding to each quality uncertainty area;

[0037] Based on the visual inquiry target, configure the categories of visual acquisition parameters to be adjusted for each quality uncertainty region. Based on the regional characteristics of the quality uncertainty region and the categories of visual acquisition parameters to be adjusted, generate a combination of visual acquisition parameters. The categories of visual acquisition parameters to be adjusted include light source adjustment, polarization adjustment, imaging adjustment and spectral adjustment.

[0038] The number of supplementary acquisitions is determined according to the combination of visual acquisition parameters, and the order of supplementary acquisitions is determined according to the correspondence between the combination of visual acquisition parameters and the visual inquiry target.

[0039] By associating the location of the quality uncertainty area, the type of suspected quality problem, the visual inquiry target, the combination of visual acquisition parameters, the number of supplementary acquisitions, and the order of supplementary acquisitions, a visual inquiry strategy for each quality uncertainty area is generated.

[0040] Optionally, the generation of the multi-conditional visual response sequence specifically includes:

[0041] Read the visual interrogation strategy corresponding to each quality uncertainty area, parse the combination of visual acquisition parameters according to the number of supplementary acquisitions and the order of supplementary acquisitions, and generate a supplementary acquisition control sequence;

[0042] The target acquisition area is determined based on the location of the area with uncertain quality, and coordinate matching is performed on the target acquisition area to generate supplementary acquisition and positioning results;

[0043] According to the supplementary acquisition control sequence and supplementary acquisition positioning results, control the visual acquisition device to switch to the acquisition state corresponding to each combination of visual acquisition parameters;

[0044] Under the acquisition states corresponding to each combination of visual acquisition parameters, supplementary image acquisition is performed on the target acquisition area and supplementary visual images are generated.

[0045] Based on the location results of the supplementary acquisition and the location of areas with uncertain quality, regional visual response images are extracted from the supplementary visual images, and multi-condition visual response sequences are generated according to the order of supplementary acquisition.

[0046] Optionally, the output of the production quality identification result specifically includes:

[0047] Read the multi-condition visual response sequence corresponding to each region with uncertain quality, extract the region visual response image according to the combination of visual acquisition parameters and the supplementary acquisition order, and generate region response analysis data.

[0048] Based on the regional response analysis data, response stability features, boundary persistence features, position consistency features, texture destructive features, morphological change features, and illumination response change features are extracted. Based on the retention state, offset state, and change state of each response feature under different combinations of visual acquisition parameters, a quality discrimination feature set is constructed.

[0049] Based on the quality discrimination feature set, the real defect discrimination and visual interference discrimination are performed on each quality uncertainty area to determine the real quality defect area and the visual interference area.

[0050] Based on the suspected quality problem type, region boundary, center location, region area, boundary persistence characteristics, texture destructive characteristics, and morphological change characteristics corresponding to the actual quality defect region, determine the defect type, defect region location, and defect severity.

[0051] Based on the actual quality defect area, visual interference area, defect type, defect location, and defect severity, the detection confidence level, product qualification status, and quality handling instructions are determined, and the production quality identification results are output. The production quality identification results include product qualification status, defect location, defect type, defect severity, detection confidence level, and quality handling instructions.

[0052] The beneficial effects of this invention are:

[0053] This invention uses the YOLOv8-seg model to identify suspected abnormal areas in the initial visual data of products to be inspected on the production line. It then combines this with a verification confidence interval to filter areas of uncertain quality. This allows for the differentiation between obvious defect areas, obvious interference areas, and areas requiring further confirmation. Consequently, subsequent inspection resources are focused on areas where actual quality defects are easily confused with visual interference. This reduces the need for repeated data collection and analysis of all product areas, improving the targeting and processing efficiency of the visual inspection process on the production line.

[0054] This invention invokes a visual interrogation template library based on the suspected quality problem type to determine the visual interrogation target, visual acquisition parameter combination, number of supplementary acquisitions, and supplementary acquisition sequence. This transforms the detection process from a single identification under fixed visual conditions to proactive visual verification of areas with uncertain quality. For different situations such as reflective interference, shadow interference, shallow scratches, dents, transparent bubbles, stains, cracks, and assembly deviations, it can obtain the visual response of the same area with uncertain quality under various visual acquisition conditions by adjusting the combination of light source orientation parameters, light source intensity parameters, polarization angle parameters, shooting angle parameters, exposure parameters, focal length parameters, and spectral channel parameters. This improves the ability to distinguish between similar appearance anomalies.

[0055] This invention extracts response stability features, boundary persistence features, positional consistency features, texture destructive features, morphological change features, and illumination response change features based on multi-condition visual response sequences, and constructs a quality discrimination feature set to distinguish between true defects and visual interference in areas of quality uncertainty. By differentiating between persistent defect manifestations under multi-condition acquisition and interference manifestations that fade or migrate with changes in illumination, polarization, and viewing angle, it can reduce misjudgments caused by visual factors such as reflections and shadows, and reduce missed detections of defects such as shallow scratches, transparent bubbles, and assembly deviations. Furthermore, it outputs product qualification status, defect area location, defect type, defect severity, detection confidence level, and quality handling instructions, enabling production quality identification results to directly serve subsequent handling processes such as release, re-inspection, rework, and rejection, thereby improving the accuracy and stability of production line quality management. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a flowchart of a production quality identification method based on AI visual inspection proposed in this invention;

[0058] Figure 2 This is a flowchart of the visual inquiry strategy generation process for a production quality recognition method based on AI visual inspection proposed in this invention.

[0059] Figure 3 This is a flowchart illustrating the output of production quality recognition results for a production quality recognition method based on AI visual inspection proposed in this invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0061] refer to Figures 1-3 A production quality identification method based on AI visual inspection includes the following steps:

[0062] Collect product images of the products to be inspected on the production line under initial visual acquisition conditions, perform preprocessing on the product images, and generate initial visual data of the products;

[0063] Input the initial visual data of the product into the YOLOv8-seg model to identify suspected abnormal areas in the product image and extract the regional features of the suspected abnormal areas to generate a suspected abnormal area feature set.

[0064] Based on the feature set of suspected abnormal regions and the verification confidence interval, regions with uncertain quality are screened to generate a set of regions with uncertain quality.

[0065] Based on the regional characteristics of each quality uncertainty region in the set of quality uncertainty regions, identify the suspected quality problem types corresponding to each quality uncertainty region;

[0066] Based on the suspected quality problem type, the visual inquiry template library is called to determine the visual inquiry target, visual acquisition parameter combination, number of supplementary acquisitions and order of supplementary acquisitions for each quality uncertainty area, and a visual inquiry strategy is generated.

[0067] According to the visual inquiry strategy, the visual acquisition device is controlled to perform multi-condition supplementary acquisition on areas with uncertain quality, and a multi-condition visual response sequence is generated.

[0068] Based on multi-condition visual response sequence identification, the system identifies real quality defect areas and visual interference areas, and outputs production quality identification results.

[0069] In this embodiment, the initial visual acquisition conditions include fixed light source orientation, light source intensity, shooting angle, exposure parameters, focal length parameters, and spectral channels. Preprocessing includes image denoising, distortion correction, brightness normalization, product area positioning, and background area removal.

[0070] In this embodiment, the generation of the suspected abnormal region feature set specifically includes:

[0071] The initial visual data of the product is processed by input size matching, pixel value normalization and tensor quantization to form a visual tensor to be identified;

[0072] The visual tensor to be identified is input into the YOLOv8-seg model, and feature extraction is performed through the backbone network to obtain multi-scale product visual features containing texture details, edge structure, and region semantics. Specifically, the visual tensor to be identified is input into the continuous convolutional layers in the backbone network for local feature extraction to obtain shallow visual features containing edge changes and brightness changes; the shallow visual features are processed through cross-stage feature connection to obtain mid-level visual features containing texture direction and local morphology information; the mid-level visual features are processed through spatial pyramid pooling to obtain deep visual features containing abnormal region semantics and product structure semantics; the shallow visual features, mid-level visual features, and deep visual features are output according to their corresponding feature layers to form multi-scale product visual features.

[0073] A feature fusion network is used to fuse multi-scale product visual features to form a fused feature map for anomaly localization and segmentation. Specifically, the multi-scale product visual features are sorted according to feature layer resolution. Low-resolution product visual features are upsampled to ensure they have the same spatial size as adjacent high-resolution product visual features. Channel alignment is performed on the spatially matched product visual features of different scales to obtain scale-aligned features with the same number of channels. The scale-aligned features are then concatenated according to channel dimension to obtain cross-scale concatenated features. Convolution, normalization, and activation are performed on the cross-scale concatenated features to obtain intermediate fused features containing shallow texture information, edge structure information, and deep semantic information. The intermediate fused features are then downsampled and passed across layers to enable bidirectional fusion of high-resolution detail features and low-resolution semantic features, forming a fused feature map.

[0074] The detection branch predicts candidate bounding boxes for anomalies in the fused feature map, obtaining candidate anomaly boxes and candidate anomaly confidence scores. Specifically, the detection branch performs convolution operations on feature layers of different scales in the fused feature map to obtain the bounding box offset, anomaly category score, and anomaly presence score for each feature point in each feature layer. Based on the position of each feature point in the fused feature map, the downsampling ratio of the corresponding feature layer, and the bounding box offset, the position range of the candidate anomaly box in the product image coordinate system is determined. The candidate anomaly confidence score is determined based on the highest score between the anomaly presence score and the anomaly category score.

[0075] Anomaly region mask prediction is performed on the fused feature map through a segmentation branch to obtain candidate anomaly masks corresponding to candidate anomaly boxes. Specifically: the fused feature map is input into the segmentation branch, which performs convolution on the fused feature map to obtain prototype mask features of the anomaly region and mask coefficients corresponding to the candidate anomaly boxes; the prototype mask features of the anomaly region are weighted and combined according to the mask coefficients to obtain the initial anomaly mask corresponding to the candidate anomaly box; the initial anomaly mask is cropped according to the position range of the candidate anomaly box in the product image coordinate system to obtain a local anomaly mask within the coverage area of ​​the candidate anomaly box; the local anomaly mask is size-mapped to make the size of the local anomaly mask consistent with the coverage area of ​​the candidate anomaly box in the product image; the size-mapped local anomaly mask is binarized to obtain the candidate anomaly mask corresponding to the candidate anomaly box.

[0076] Candidate anomaly boxes and masks are screened based on their anomaly confidence levels. Overlapping regions are then removed from the screened candidate anomaly boxes and masks to identify suspected anomaly regions. Specifically: candidate anomaly boxes with confidence levels below a preset confidence threshold are deleted, and their corresponding candidate anomaly masks are simultaneously deleted; the remaining candidate anomaly boxes are sorted from highest to lowest confidence level to obtain a ranking result; the current candidate anomaly box is selected sequentially according to the ranking result, and the overlap between the current candidate anomaly box and the other candidate anomaly boxes is calculated. The overlap level is determined by the ratio between the overlapping area of ​​two candidate anomaly boxes and the merged coverage area; when the overlap level is greater than a preset overlap threshold, candidate anomaly boxes with lower confidence levels are deleted, and their corresponding candidate anomaly masks are simultaneously deleted; the coverage area of ​​the remaining candidate anomaly boxes after overlapping region removal is determined as the suspected anomaly region, the remaining candidate anomaly masks are used as the candidate anomaly masks corresponding to the suspected anomaly regions, and the remaining candidate anomaly confidence levels are used as the initial defect confidence levels for the suspected anomaly regions.

[0077] The location of a suspected abnormal region is determined based on the candidate abnormal bounding box corresponding to the suspected abnormal region. The boundary and area of ​​the suspected abnormal region are then determined based on the candidate abnormal mask corresponding to the suspected abnormal region. Specifically: the position range of the candidate abnormal bounding box corresponding to the suspected abnormal region in the product image coordinate system is read, and this position range is used as the location of the suspected abnormal region; the candidate abnormal mask corresponding to the suspected abnormal region is read, and the pixel regions representing abnormalities in the candidate abnormal mask are determined as the effective mask region; contour extraction is performed on the effective mask region to obtain the boundary of the suspected abnormal region; the number of pixels representing abnormalities within the effective mask region is counted to obtain the pixel area of ​​the suspected abnormal region; the pixel area is converted into the region area according to the pixel calibration relationship of the product image, where the pixel calibration relationship includes the actual area corresponding to a single pixel in the product image, determined by the calibration parameters of the visual acquisition device; the region area is obtained by multiplying the number of pixels representing abnormalities within the effective mask region by the actual area corresponding to a single pixel.

[0078] The pixel distribution, grayscale gradient, edge direction, and region contour within suspected anomaly regions are statistically analyzed to obtain texture continuity, brightness distribution, edge morphology, region topography, and grayscale changes. These are then correlated with the suspected anomaly region location, boundary, area, and initial defect confidence to generate a suspected anomaly region feature set. The suspected anomaly region is divided into multiple adjacent local texture blocks. The principal gradient direction of each local texture block is extracted, and the directional offset amplitude and the number of texture breakpoints between adjacent local texture blocks are statistically analyzed to obtain texture continuity. A texture breakpoint is defined as a point where the gradient principal direction offset amplitude between adjacent local texture blocks reaches a preset directional offset. The threshold location is determined; the brightness values ​​of each pixel within the suspected abnormal region are read, and the average brightness, extreme brightness, brightness gradient change, and brightness histogram distribution are statistically analyzed to obtain the brightness distribution; edge detection is performed on the suspected abnormal region, and the extension direction, edge length, number of breakpoints, degree of curvature, and closure state of the abnormal edge lines are extracted to obtain the edge morphology; the length-to-width ratio, main extension direction, compactness, convex-concave distribution, and hole distribution of the region are extracted based on the boundary of the suspected abnormal region to obtain the region morphology; grayscale sequences are extracted along the edge direction and main extension direction of the suspected abnormal region, and the grayscale abrupt change location, grayscale change amplitude, and grayscale duration range are statistically analyzed to obtain the grayscale change.

[0079] In this embodiment, the generation of the set of regions with uncertain mass specifically includes:

[0080] Configure a review confidence interval, which includes a lower review confidence limit and an upper review confidence limit, with the lower review confidence limit being less than the upper review confidence limit. Specifically, this involves: obtaining the initial defect confidence level and manual review labels for suspected abnormal areas in historical production quality inspection samples; the manual review labels include defect confirmation labels and interference confirmation labels; statistically analyzing the initial defect confidence level distribution corresponding to the defect confirmation labels and the initial defect confidence level distribution corresponding to the interference confirmation labels based on the manual review labels; determining the overlap confidence level range between the initial defect confidence level distribution corresponding to the defect confirmation labels and the initial defect confidence level distribution corresponding to the interference confirmation labels; configuring the lowest confidence level value of the overlap confidence level range as the lower review confidence limit, and configuring the highest confidence level value of the overlap confidence level range as the upper review confidence limit, thus forming the review confidence interval.

[0081] The initial defect confidence level corresponding to each suspected abnormal region is compared with the lower limit of the review confidence level and the upper limit of the review confidence level. When the initial defect confidence level corresponding to the suspected abnormal region is greater than or equal to the lower limit of the review confidence level and less than or equal to the upper limit of the review confidence level, the corresponding suspected abnormal region is determined as a quality uncertainty region.

[0082] Extract the regional features corresponding to the quality uncertain regions from the feature set of suspected abnormal regions, associate the quality uncertain regions with the corresponding regional features, and generate a set of quality uncertain regions.

[0083] In this embodiment, the identification of suspected quality problem types specifically includes:

[0084] The regional features corresponding to each quality uncertainty region in the set of quality uncertainty regions are normalized to obtain the type discrimination features corresponding to each quality uncertainty region;

[0085] Configure the type discrimination conditions corresponding to reflective interference, shadow interference, shallow scratches, dents, transparent bubbles, stains, cracks and assembly deviations respectively. The type discrimination conditions are used to limit the texture continuity range, brightness distribution range, edge shape range, area morphology range, gray scale change range, area range and initial defect confidence range corresponding to the suspected quality problem type.

[0086] Specifically, the following steps are taken: Establish a type discrimination condition table, configuring the texture continuity condition, brightness distribution condition, edge morphology condition, regional morphology condition, grayscale change condition, regional area condition, and initial defect confidence condition for each suspected quality problem type. For example, the type discrimination conditions for reflective interference include concentrated extreme brightness values, a sudden increase in grayscale change amplitude, unstable edge morphology, and unbroken texture continuity; the type discrimination conditions for shadow interference include continuously decreasing brightness values, consistent grayscale change direction, weakened edge morphology, and patchy extension of the regional morphology; the type discrimination conditions for shallow scratches include interrupted texture continuity, linear edge extension, grayscale change distributed along a linear area, and a narrow and elongated regional area; and the type discrimination conditions for depressions include localized transitions between light and dark in brightness distribution, a locally concave contour in the regional morphology, and grayscale changes expanding around the center of the region. The morphology of scattered and edge-shaped bubbles is closed or semi-closed; the morphology of transparent bubbles includes ring-shaped or dot-shaped brightness distribution, circular or elliptical regional morphology, closed edge morphology, and grayscale change transitioning from the boundary to the center; the morphology of stains includes uneven brightness distribution, irregular patchy regional morphology, blurred edge morphology, and local weakening of texture continuity; the morphology of cracks includes broken texture continuity, elongated branching edge morphology, grayscale change extending along the branching direction, and irregular linear regional morphology; the morphology of assembly deviations includes positional deviation between the boundary of the suspected abnormal area and the boundary of the product structure, structural misalignment of the edge morphology, distribution of regional morphology along the assembly connection area, and correspondence between the regional area and the range of assembly gap variation; the morphology of each suspected quality problem type is written into the morphology table.

[0087] The type discrimination features corresponding to each quality uncertainty region are matched with each type discrimination condition. The number of feature terms satisfied by the type discrimination features in each type discrimination condition is counted. Based on the number of feature terms, the type matching strength of each quality uncertainty region relative to each suspected quality problem type is determined. The type matching strength is calculated by the ratio of the number of feature terms satisfied by the type discrimination features to the total number of feature terms contained in the corresponding type discrimination condition.

[0088] The suspected quality problem type with the strongest type matching strength is identified as the suspected quality problem type in the corresponding quality uncertainty region, and the quality uncertainty region is associated with the corresponding suspected quality problem type.

[0089] In this embodiment, the generation of the visual inquiry strategy specifically includes:

[0090] Establish a visual inquiry template library, which includes visual inquiry template items corresponding to reflective interference, shadow interference, shallow scratches, dents, transparent bubbles, stains, cracks and assembly deviations. Each visual inquiry template item records the target discrimination content, the category of visual acquisition parameters to be adjusted, the parameter value range, the parameter change direction and parameter combination constraints.

[0091] Establishing a visual inquiry template library includes: acquiring historical area features, manual review labels, initial visual acquisition conditions, multi-condition supplementary acquisition records, and supplementary acquisition discrimination results for each suspected quality problem type in historical production quality inspection samples; classifying historical production quality inspection samples according to reflective interference, shadow interference, shallow scratches, dents, transparent bubbles, stains, cracks, and assembly deviations; extracting target discrimination content for each suspected quality problem type during multi-condition supplementary acquisition to distinguish between actual quality defect areas and visual interference areas; determining the corresponding categories of visual acquisition parameters to be adjusted based on the target discrimination content; and applying the calibration parameters of the visual acquisition equipment. The system configures the parameter value ranges for each category of visual acquisition parameters to be adjusted, based on the location and area distribution of uncertain historical quality areas in the production line inspection cycle, the parameter change direction based on the changing trend of the target discrimination content under different visual acquisition conditions, and the parameter combination constraints based on the equipment linkage relationship between light source orientation parameters, light source intensity parameters, polarization angle parameters, shooting angle parameters, exposure parameters, focal length parameters, and spectral channel parameters. The system writes the suspected quality problem type, target discrimination content, category of visual acquisition parameters to be adjusted, parameter value range, parameter change direction, and parameter combination constraints into the corresponding visual inquiry template item to generate a visual inquiry template library.

[0092] Read the suspected quality problem types associated with each quality uncertainty region in the quality uncertainty region set, and call the corresponding visual inquiry template items from the visual inquiry template library according to the suspected quality problem type to generate template call results;

[0093] Based on the target discrimination content in the template call result, the visual inquiry target corresponding to each quality uncertainty region is determined. The visual inquiry target is used to characterize the quality discrimination content that needs to be verified in the multi-condition supplementary acquisition process of the quality uncertainty region.

[0094] The determination of visual inquiry targets includes: when the suspected quality problem type is reflective interference, the fading state of extreme brightness values, the stable state of edge morphology, and the state of texture continuity are identified as visual inquiry targets; when the suspected quality problem type is shadow interference, the migration state of low-brightness areas with changes in light source orientation, the consistency of grayscale change direction, and the state of regional morphological extension are identified as visual inquiry targets; when the suspected quality problem type is shallow scratches, the persistent state of linear edges, the persistent state of texture breaks, and the state of grayscale changes along linear areas are identified as visual inquiry targets; when the suspected quality problem type is depression, the state of local light and dark transitions, the state of regional center morphological changes, and the persistent state of closed edges are identified as visual inquiry targets; when the suspected quality problem type is... When the bubble is transparent, the visual inquiry targets are determined by the following: the state of maintaining annular or dotted brightness changes, the state of maintaining closed edges, and the state of changing regional visibility under different spectral channels. When the suspected quality problem is a stain, the visual inquiry targets are determined by the state of maintaining irregular patchy areas, the state of maintaining blurred edges, and the state of changing brightness distribution under different spectral channels. When the suspected quality problem is a crack, the visual inquiry targets are determined by the state of maintaining slender branched edges, the state of maintaining broken texture continuity, and the state of extending grayscale changes along the branch direction. When the suspected quality problem is an assembly deviation, the visual inquiry targets are determined by the state of maintaining the positional deviation between the boundary of the suspected abnormal area and the boundary of the product structure, the state of structural misalignment, and the state of changing assembly gaps.

[0095] Based on the visual inquiry target, configure the categories of visual acquisition parameters to be adjusted for each quality uncertainty area. Based on the regional characteristics of the quality uncertainty area and the categories of visual acquisition parameters to be adjusted, generate a combination of visual acquisition parameters. These categories include light source adjustment, polarization adjustment, imaging adjustment, and spectral adjustment. Specifically, light source adjustment corresponds to light source orientation and intensity parameters; polarization adjustment corresponds to polarization angle parameters; imaging adjustment corresponds to shooting angle, exposure, and focal length parameters; and spectral adjustment corresponds to spectral channel parameters. When the visual inquiry target involves reflection, extreme brightness values, or shadow migration, select light source adjustment, polarization adjustment, and imaging adjustment. When the visual inquiry target involves linear edges, texture breaks, closed edges, locally recessed contours, or structural misalignment, select light source adjustment, polarization adjustment, and imaging adjustment. When the visual inquiry target involves the visibility of transparent areas, stain brightness distribution, changes in regional visibility under different spectral channels, or changes in brightness distribution, select polarization adjustment, imaging adjustment, and spectral adjustment.

[0096] The generation of visual acquisition parameter combinations includes: reading the parameter value range, parameter change direction, and parameter combination constraints corresponding to each category of visual acquisition parameters to be adjusted from the template call results; determining the light source intensity parameter value and exposure parameter value based on the brightness distribution of the quality uncertainty area; determining the light source orientation parameter value, polarization angle parameter value, and shooting angle parameter value based on the texture continuity and edge morphology of the quality uncertainty area; determining the focal length parameter value based on the area and morphology of the quality uncertainty area; determining the spectral channel parameter value based on the grayscale change and brightness distribution of the quality uncertainty area; combining multiple parameter values ​​that satisfy the parameter combination constraints; and deleting parameter combinations that cause overexposure, underexposure, occlusion, out of focus, or exceed the operating range of the visual acquisition device to obtain the visual acquisition parameter combination.

[0097] The number of supplementary acquisitions is determined according to the combination of visual acquisition parameters, and the order of supplementary acquisitions is determined according to the correspondence between the combination of visual acquisition parameters and the visual inquiry target.

[0098] Determining the number of supplementary acquisitions includes: assigning one supplementary acquisition to each set of visual acquisition parameters; when an area of ​​uncertain quality simultaneously meets the criteria for multiple suspected quality problem types, merging the visual acquisition parameter combinations corresponding to each of the multiple suspected quality problem types to determine the number of supplementary acquisitions; deleting duplicate combinations from the merged visual acquisition parameter combinations, ensuring that only one visual acquisition parameter combination with completely identical light source orientation parameters, light source intensity parameters, polarization angle parameters, shooting angle parameters, exposure parameters, focal length parameters, and spectral channel parameters is retained; the number of visual acquisition parameter combinations after deleting duplicate combinations is determined as the number of supplementary acquisitions.

[0099] The supplementary acquisition sequence is determined as follows: Visual acquisition parameter combinations used to change brightness extrema, shadow migration, and reflection states are arranged first to determine reflective and shadow interference; visual acquisition parameter combinations used to enhance texture breaks, linear edges, closed edges, and structural misalignment states are arranged in the middle to determine shallow scratches, dents, cracks, and assembly deviations; visual acquisition parameter combinations used to switch spectral channels, adjust focal length parameters, and refine exposure parameters are arranged last to determine the severity of transparent bubbles, stains, and defects; when the same area of ​​quality uncertainty corresponds to multiple visual inquiry targets, the supplementary acquisition sequence is generated according to the interference elimination priority, defect confirmation priority, and severity identification priority corresponding to the visual inquiry targets.

[0100] By associating the location of the quality uncertainty area, the type of suspected quality problem, the visual inquiry target, the combination of visual acquisition parameters, the number of supplementary acquisitions, and the order of supplementary acquisitions, a visual inquiry strategy for each quality uncertainty area is generated.

[0101] In this embodiment, the generation of the multi-condition visual response sequence specifically includes:

[0102] Read the visual inquiry strategy corresponding to each quality uncertainty area, parse the combination of visual acquisition parameters according to the number of supplementary acquisitions and the order of supplementary acquisitions, and generate a supplementary acquisition control sequence. The supplementary acquisition control sequence includes the location of the quality uncertainty area, the type of suspected quality problem, the visual inquiry target, the supplementary acquisition order identifier, the combination of visual acquisition parameters, the action content of the visual acquisition device, and the acquisition result storage identifier.

[0103] The target acquisition area is determined based on the location of the quality uncertainty area, and coordinate matching is performed on the target acquisition area to generate supplementary acquisition and positioning results. The coordinate matching converts the region boundary and center position of the quality uncertainty area in the product image coordinate system into the imaging position in the camera imaging coordinate system according to the calibration parameters of the vision acquisition device. Based on the calibration relationship between the camera imaging coordinate system and the production line bearing coordinate system, the alignment position of the vision acquisition device when supplementing acquisition of the target acquisition area is determined.

[0104] According to the supplementary acquisition control sequence and supplementary acquisition positioning results, the visual acquisition device is controlled to switch to the acquisition state corresponding to each combination of visual acquisition parameters. The acquisition state includes the light source orientation parameters and light source intensity parameters corresponding to the adjustable light source, the polarization angle parameters corresponding to the polarization component, the shooting angle adjustment component corresponding to the shooting angle parameters, the camera exposure parameters, the lens adjustment component corresponding to the focal length parameters, and the spectral channel parameters corresponding to the multispectral acquisition component.

[0105] Under the acquisition states corresponding to each combination of visual acquisition parameters, supplementary image acquisition is performed on the target acquisition area and supplementary visual images are generated; the supplementary visual images are stored according to the acquisition result storage identifier, and are associated with the record of the quality uncertainty area identifier, acquisition time, visual acquisition parameter combination, supplementary acquisition sequence identifier, suspected quality problem type, and visual inquiry target;

[0106] Based on the supplementary acquisition and positioning results and the location of areas with uncertain quality, regional visual response images are extracted from the supplementary visual images, and a multi-condition visual response sequence is generated according to the supplementary acquisition order. The multi-condition visual response sequence is obtained by associating and arranging the regional visual response images after scale unification, position registration, and brightness normalization with the corresponding visual acquisition parameter combinations, suspected quality problem types, and visual inquiry targets. Scale unification adjusts the image size of the regional visual response images according to the area and boundaries of the areas with uncertain quality. Position registration spatially aligns the regional visual response images under different combinations of visual acquisition parameters according to the center position and boundaries of the areas with uncertain quality. Brightness normalization is performed on the regional visual response images to ensure brightness consistency based on the mean brightness, extreme brightness values, and brightness distribution range in the supplementary visual images.

[0107] In this embodiment, the output of the production quality identification result specifically includes:

[0108] Read the multi-condition visual response sequence corresponding to each region with uncertain quality, extract the region visual response image according to the combination of visual acquisition parameters and the supplementary acquisition order, and generate region response analysis data.

[0109] Based on the regional response analysis data, response stability features, boundary persistence features, position consistency features, texture destructive features, morphological change features, and illumination response change features are extracted. Based on the retention state, offset state, and change state of each response feature under different combinations of visual acquisition parameters, a quality discrimination feature set is constructed.

[0110] The response stability characteristics are determined based on the brightness and grayscale variation retention states of the same uncertain quality region under different combinations of visual acquisition parameters; the boundary persistence characteristics are determined based on the continuous appearance, closure, and breakpoint changes of the region boundary in different region visual response images; the positional consistency characteristics are determined based on the offset states of the region center position and region boundary position in different region visual response images; the texture destructive characteristics are determined based on the texture continuity breakage, edge extension interruption, and grayscale abrupt change persistence states; the morphological change characteristics are determined based on the changes in the region's aspect ratio, main extension direction, convex-concave distribution, and hole distribution in different region visual response images; and the illumination response change characteristics are determined based on the brightness extreme value fading state, shadow migration state, and reflection change state when the light source orientation parameters, light source intensity parameters, polarization angle parameters, and exposure parameters change.

[0111] The quality discrimination feature set includes real defect discrimination items and visual interference discrimination items. The real defect discrimination items consist of features that characterize the persistence of regional state from response stability features, boundary persistence features, positional consistency features, texture destructive features, and morphological change features. The visual interference discrimination items consist of features that characterize the fading of extreme brightness values, shadow migration, reduced reflection, discontinuous boundaries, or inconsistent positions from illumination response change features, boundary persistence features, and positional consistency features.

[0112] Based on the quality discrimination feature set, the real defect discrimination and visual interference discrimination are performed on each quality uncertainty area to determine the real quality defect area and the visual interference area.

[0113] The criteria for identifying true defects include: when a region of uncertain quality meets the following conditions, such as stable response, persistent boundaries, consistent position, and at least one of the following conditions, such as persistent texture destruction, preserved morphology, and persistent grayscale abrupt changes, the region of uncertain quality is identified as a region of true quality defects. The criteria for identifying visual interference include: for a region of uncertain quality that has not been identified as a region of true quality defects, when a region of uncertain quality meets at least one of the following conditions, such as the brightness extremum fading with changes in the light source, the shadow area migrating with the orientation of the light source, the reflection area weakening with changes in the polarization angle, and the boundary being discontinuous or inconsistent in position, the region of uncertain quality is identified as a region of visual interference.

[0114] Based on the suspected quality problem type, region boundary, center location, region area, boundary persistence characteristics, texture damage characteristics, and morphological change characteristics corresponding to the actual quality defect region, the defect type, defect region location, and defect severity are determined. Specifically: the type representing product entity abnormality among the suspected quality problem types corresponding to the actual quality defect region is taken as the candidate defect type. Based on the consistency between the actual defect discrimination items in the quality discrimination feature set and the quality discrimination content represented by the visual inquiry target, the candidate defect type is determined as the defect type. The region boundary of the actual quality defect region in the product image coordinate system is converted into the defect region outline, and the center location of the actual quality defect region in the product image coordinate system is converted into the defect region coordinates. The defect region location is composed of the defect region outline and the defect region coordinates. The region area, boundary persistence, texture damage degree, and morphological change degree of the actual quality defect region are classified into levels, and the level classification results are used as the defect severity. The defect severity includes minor, medium, and severe levels.

[0115] Based on the actual quality defect area, visual interference area, defect type, defect location, and defect severity, the detection confidence level, product qualification status, and quality handling instructions are determined, and the production quality identification results are output. The production quality identification results include product qualification status, defect location, defect type, defect severity, detection confidence level, and quality handling instructions.

[0116] The detection confidence score is obtained by weighting the proportion of true defect criteria satisfied, the proportion of visual interference criteria excluded, and the effective acquisition ratio. The proportion of true defect criteria satisfied is the ratio of the number of satisfied true defect criteria to the total number of true defect criteria. The proportion of visual interference criteria excluded is the ratio of the number of unsatisfied visual interference criteria to the total number of visual interference criteria. The effective acquisition ratio is the ratio of the number of visual response images of the region that have completed scale unification, position registration, and brightness normalization processing to the number of supplementary acquisitions.

[0117] The product's qualification status is determined based on the number and severity of actual quality defect areas. When the number of actual quality defect areas is zero, the product is qualified. When the number of actual quality defect areas is greater than zero, and the severity of all actual quality defect areas is minor, the product is qualified and awaits re-inspection. When the number of actual quality defect areas is greater than zero, and the severity of any actual quality defect area is moderate or severe, the product is qualified and unqualified.

[0118] Quality handling instructions are determined based on the product's conformity status and the highest defect severity of the actual quality defect area; a release instruction is generated when the product's conformity status is conformity; a re-inspection instruction is generated when the product's conformity status is pending re-inspection; a rework instruction is generated when the product's conformity status is non-conforming and the highest defect severity is medium; and a rejection instruction is generated when the product's conformity status is non-conforming and the highest defect severity is severe.

[0119] Example 1: To verify the feasibility of this invention in practice, it was applied to the product appearance quality inspection scenario at a manufacturing company's production base in East China. During continuous production, some product surfaces exhibit similar image appearances due to metallic reflections, localized shadows, transparent bubbles, shallow scratches, and assembly deviations. Visual inspection under a fixed light source and fixed shooting angle easily misjudges reflective areas as defects and shallow scratches as ordinary brightness variations, leading to increased re-inspection pressure.

[0120] In this scenario, the visual acquisition device acquires images of the product to be inspected under initial visual acquisition conditions and inputs the preprocessed initial visual data into the YOLOv8-seg model to identify suspected abnormal areas. For quality uncertainty areas with confidence levels within the verification confidence interval, the system calls upon the visual interrogation template library based on the suspected quality problem type to form corresponding visual interrogation targets, visual acquisition parameter combinations, supplementary acquisition times, and supplementary acquisition order. The visual acquisition device then switches the light source orientation, light source intensity, polarization angle, shooting angle, exposure, focal length, and spectral channels around the quality uncertainty area to obtain a multi-condition visual response sequence. Based on response stability, boundary persistence, positional consistency, texture destructiveness, morphological changes, and illumination response changes, it identifies actual quality defect areas and visual interference areas.

[0121] On-site operation records show that, within the same production location, product batch, and continuous production period, the use of this invention reduces misjudgments caused by reflections and shadows, improves the consistency of verification of shallow scratches, transparent bubbles, and assembly deviations, and reduces re-inspections due to visual interference in manual re-inspection records. Production line quality logs show that the inspection results can simultaneously output product qualification status, defect location, defect type, defect severity, inspection confidence level, and quality handling instructions, making release, re-inspection, rework, and rejection processes clearer. This demonstrates that this invention can improve the accuracy of production quality identification and the reliability of handling.

[0122] Table 1. Performance Comparison of the Invention and Traditional Production Quality Identification Methods

[0123] Indicator Categories Traditional fixed visual inspection method Method of the present invention Overall judgment time (seconds / item) 2.36 2.05 Defect identification accuracy (%) 87.60 92.80 Visual interference misjudgment rate (%) 9.40 5.80 True defect miss rate (%) 6.10 3.70 Percentage of manual re-inspection (%) 18.50 13.20 Defect location deviation (pixels) 8.60 6.40 Consistency rate of quality handling (%) 89.30 94.10

[0124] The data in Table 1 comes from the continuous production line verification records of the appearance inspection station at the East China production base of a manufacturing enterprise. Traditional fixed visual inspection methods use fixed light sources, fixed shooting angles, and fixed exposure conditions for identification. The method of this invention uses quality uncertainty area screening, visual inquiry template library calling, and multi-condition supplementary collection for identification.

[0125] The overall judgment time was reduced from 2.36 seconds / piece to 2.05 seconds / piece. This is because the present invention does not repeatedly collect the entire product image, but only supplements the collection of quality uncertainty areas within the verification confidence interval, thereby reducing the time consumption caused by invalid verification areas.

[0126] The defect identification accuracy improved from 87.60% to 92.80%, primarily because this invention obtains multi-condition visual response sequences of the same quality uncertainty area under different light sources, polarization, viewing angles, exposures, focal lengths, and spectral channels through a combination of visual acquisition parameters. This allows defects such as shallow scratches, transparent bubbles, dents, and assembly deviations to be further confirmed through boundary persistence, positional consistency, and texture destructiveness. The visual interference misjudgment rate decreased from 9.40% to 5.80%. This is because reflections and shadows exhibit brightness extrema fading, shadow migration, or reflection weakening under different light source orientations, polarization angles, and exposure conditions. This invention can classify such changes into the visual interference discrimination item, reducing the number of times visual interference is misjudged as real quality defects.

[0127] The false defect detection rate decreased from 6.10% to 3.70%. This is because traditional methods are insufficient in identifying defects with weak textures, light grayscale variations, and indistinct small boundaries. This invention improves the stability of identifying low-significance defects by supplementing the sampling sequence to progressively verify texture breaks, boundary continuity, and morphological preservation. The proportion of manual re-inspection decreased from 18.50% to 13.20%, indicating more stable detection confidence output, allowing more products to directly enter the release, rework, or rejection process.

[0128] The defect location deviation decreased from 8.60 pixels to 6.40 pixels because the present invention performs scale unification, position registration, and brightness normalization processing in the supplementary visual image, ensuring spatial alignment of the regional visual response images under different combinations of visual acquisition parameters. The consistency rate of quality handling increased from 89.30% to 94.10% because the present invention simultaneously outputs the defect type, defect severity, detection confidence level, and quality handling instructions, making the correspondence between the detection results and the release, re-inspection, rework, and rejection rules clearer.

[0129] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A production quality identification method based on AI visual inspection, characterized in that, Includes the following steps: Collect product images of the products to be inspected on the production line under initial visual acquisition conditions, perform preprocessing on the product images, and generate initial visual data of the products; Input the initial visual data of the product into the YOLOv8-seg model to identify suspected abnormal areas in the product image and extract the regional features of the suspected abnormal areas to generate a suspected abnormal area feature set. Based on the feature set of suspected abnormal regions and the verification confidence interval, regions with uncertain quality are screened to generate a set of regions with uncertain quality. Based on the regional characteristics of each quality uncertainty region in the set of quality uncertainty regions, identify the suspected quality problem types corresponding to each quality uncertainty region; The identification of suspected quality problem types specifically includes: The regional features corresponding to each quality uncertainty region in the set of quality uncertainty regions are normalized to obtain the type discrimination features corresponding to each quality uncertainty region; Configure the type discrimination conditions corresponding to reflective interference, shadow interference, shallow scratches, dents, transparent bubbles, stains, cracks and assembly deviations respectively; The type discrimination features corresponding to each quality uncertainty region are matched with each type discrimination condition. The number of feature terms satisfied by the type discrimination features in each type discrimination condition is counted. The type matching strength of each quality uncertainty region relative to each suspected quality problem type is determined based on the number of feature terms. The suspected quality problem type with the strongest type matching strength is identified as the suspected quality problem type in the corresponding quality uncertainty region, and the quality uncertainty region is associated with the corresponding suspected quality problem type; Based on the suspected quality problem type, the visual inquiry template library is called to determine the visual inquiry target, visual acquisition parameter combination, number of supplementary acquisitions and order of supplementary acquisitions for each quality uncertainty area, and a visual inquiry strategy is generated. The generation of the visual inquiry strategy specifically includes: A visual inquiry template library is established, which includes visual inquiry template items corresponding to reflective interference, shadow interference, shallow scratches, dents, transparent bubbles, stains, cracks and assembly deviations, respectively. Each visual inquiry template item records the target discrimination content, the category of visual acquisition parameters to be adjusted, the parameter value range, the parameter change direction and parameter combination constraints. Read the suspected quality problem types associated with each quality uncertainty region in the quality uncertainty region set, and call the corresponding visual inquiry template items from the visual inquiry template library according to the suspected quality problem type to generate template call results; Based on the target discrimination content in the template call results, determine the visual inquiry target corresponding to each quality uncertainty area; Based on the visual inquiry target, configure the categories of visual acquisition parameters to be adjusted for each quality uncertainty region. Based on the regional characteristics of the quality uncertainty region and the categories of visual acquisition parameters to be adjusted, generate a combination of visual acquisition parameters. The categories of visual acquisition parameters to be adjusted include light source adjustment, polarization adjustment, imaging adjustment and spectral adjustment. The number of supplementary acquisitions is determined according to the combination of visual acquisition parameters, and the order of supplementary acquisitions is determined according to the correspondence between the combination of visual acquisition parameters and the visual inquiry target. By associating the location of the quality uncertainty area, the type of suspected quality problem, the visual inquiry target, the combination of visual acquisition parameters, the number of supplementary acquisitions and the order of supplementary acquisitions, a visual inquiry strategy for each quality uncertainty area is generated. According to the visual inquiry strategy, the visual acquisition device is controlled to perform multi-condition supplementary acquisition on areas with uncertain quality, and a multi-condition visual response sequence is generated. Based on multi-condition visual response sequence identification, real quality defect areas and visual interference areas are identified, and production quality identification results are output. The output of the production quality identification results specifically includes: Read the multi-condition visual response sequence corresponding to each region with uncertain quality, extract the region visual response image according to the combination of visual acquisition parameters and the supplementary acquisition order, and generate region response analysis data. Based on the regional response analysis data, response stability features, boundary persistence features, position consistency features, texture destructive features, morphological change features, and illumination response change features are extracted. Based on the retention state, offset state, and change state of each response feature under different combinations of visual acquisition parameters, a quality discrimination feature set is constructed. Based on the quality discrimination feature set, the real defect discrimination and visual interference discrimination are performed on each quality uncertainty area to determine the real quality defect area and the visual interference area. Based on the suspected quality problem type, region boundary, center location, region area, boundary persistence characteristics, texture destructive characteristics, and morphological change characteristics corresponding to the actual quality defect region, determine the defect type, defect region location, and defect severity. Based on the actual quality defect area, visual interference area, defect type, defect location, and defect severity, the detection confidence level, product qualification status, and quality handling instructions are determined, and the production quality identification results are output. The production quality identification results include product qualification status, defect location, defect type, defect severity, detection confidence level, and quality handling instructions.

2. The production quality identification method based on AI visual inspection according to claim 1, characterized in that, The initial visual acquisition conditions include fixed light source orientation, light source intensity, shooting angle, exposure parameters, focal length parameters, and spectral channels. The preprocessing includes image denoising, distortion correction, brightness normalization, product area positioning, and background area removal.

3. The production quality identification method based on AI visual inspection according to claim 1, characterized in that, The generation of the suspected abnormal region feature set specifically includes: The initial visual data of the product is processed by input size matching, pixel value normalization and tensor quantization to form a visual tensor to be identified; The visual tensor to be identified is input into the YOLOv8-seg model, and feature extraction is performed through the backbone network to obtain multi-scale product visual features that include texture details, edge structure and regional semantics. Multi-scale product visual features are fused using a feature fusion network to form a fused feature map; The detection branch is used to predict candidate boxes for abnormal regions in the fused feature map, and candidate abnormal boxes and candidate abnormal confidence scores are obtained. By segmenting the branch, anomaly region mask prediction is performed on the fused feature map to obtain the candidate anomaly mask corresponding to the candidate anomaly box. Candidate anomaly boxes and candidate anomaly masks are filtered based on the confidence level of the candidate anomalies, and overlapping areas are removed from the filtered candidate anomaly boxes and candidate anomaly masks to identify suspected anomaly areas. The location of the suspected abnormal region is determined based on the candidate abnormal bounding box corresponding to the suspected abnormal region, and the boundary and area of ​​the suspected abnormal region are determined based on the candidate abnormal mask corresponding to the suspected abnormal region. The pixel distribution, grayscale gradient, edge direction, and region contour within the suspected abnormal region are statistically analyzed to obtain texture continuity, brightness distribution, edge morphology, region appearance, and grayscale variation. These are then correlated with the location, boundary, area, and initial defect confidence of the suspected abnormal region to generate a feature set of the suspected abnormal region.

4. The production quality identification method based on AI visual inspection according to claim 1, characterized in that, The generation of the set of regions with mass uncertainty specifically includes: Configure a review confidence interval, which includes a lower review confidence limit and a higher review confidence limit, wherein the lower review confidence limit is less than the higher review confidence limit; The initial defect confidence level corresponding to each suspected abnormal region is compared with the lower limit of the review confidence level and the upper limit of the review confidence level. When the initial defect confidence level corresponding to the suspected abnormal region is greater than or equal to the lower limit of the review confidence level and less than or equal to the upper limit of the review confidence level, the corresponding suspected abnormal region is determined as a quality uncertainty region. Extract the regional features corresponding to the quality uncertain regions from the feature set of suspected abnormal regions, associate the quality uncertain regions with the corresponding regional features, and generate a set of quality uncertain regions.

5. The production quality identification method based on AI visual inspection according to claim 1, characterized in that, The generation of the multi-conditional visual response sequence specifically includes: Read the visual interrogation strategy corresponding to each quality uncertainty area, parse the combination of visual acquisition parameters according to the number of supplementary acquisitions and the order of supplementary acquisitions, and generate a supplementary acquisition control sequence; The target acquisition area is determined based on the location of the area with uncertain quality, and coordinate matching is performed on the target acquisition area to generate supplementary acquisition and positioning results; According to the supplementary acquisition control sequence and supplementary acquisition positioning results, control the visual acquisition device to switch to the acquisition state corresponding to each combination of visual acquisition parameters; Under the acquisition states corresponding to each combination of visual acquisition parameters, supplementary image acquisition is performed on the target acquisition area and supplementary visual images are generated. Based on the location results of the supplementary acquisition and the location of areas with uncertain quality, regional visual response images are extracted from the supplementary visual images, and multi-condition visual response sequences are generated according to the order of supplementary acquisition.

Images