A method and system for intelligently identifying defects in tempered coated glass

Abstract

The application discloses a kind of toughened coating glass defect intelligent identification method and system, belong to machine vision detection and industrial artificial intelligence field, to solve the problem that multiple film layer defects and toughened stress lines are similar in different illumination angles Feature and the model adaptability problem caused by film system, thickness, curvature and batch variation, the present application is by collecting the same coordinate reflection bright field, transmission bright field, multi-angle dark field and orthogonal polarization image, form the multi-optical state tensor corrected by reference area, use batch condition scale bias correction and optical relationship node aggregation to form mechanism fusion characteristics, and combine class prototype, confidence set and controlled update output defect result and production line instruction, realize the technical effect of similar defect separation, cross-batch adaptation and limited incremental update.

Description

Technical Field

[0001] This invention relates to the fields of machine vision inspection and industrial artificial intelligence, and in particular to a method and system for intelligent identification of defects in tempered coated glass. Background Technology

[0002] Tempered glass production lines typically employ machine vision devices for online inspection of the glass surface and coating. Existing devices mostly acquire images under single bright, dark, or fixed polarization conditions, and then determine coating defects based on grayscale thresholds, color differences, or general image recognition models. With the combined application of coating and tempering processes, pinholes, scratches, discoloration, delamination, particle adhesion, and tempering stress patterns can produce similar local brightness, texture, or edge appearances under different lighting directions and polarization conditions.

[0003] Existing detection methods struggle to simultaneously utilize the differences between reflection, transmission, angular scattering, and polarization birefringence responses. Consequently, they are prone to misclassifying tempered steel stress patterns as film defects or overlooking defects with weak responses under a single optical condition. Furthermore, different film systems, glass thicknesses, surface curvatures, and production batches alter image brightness, color distribution, and local scattering responses. Batch updates to all model parameters may compromise the recognition capabilities of historical batches, making it difficult to stably utilize existing models for continuous production.

[0004] Therefore, there is a need for an intelligent identification method and system for tempered coated glass defects that can overcome the shortcomings of the existing technologies. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent method and system for identifying defects in tempered coated glass. Addressing the limitations of existing technologies in distinguishing between various types of film defects and tempered stress patterns, and the susceptibility of models to variations in film system, thickness, curvature, and production batches, this invention proposes a technical solution that involves acquiring reflected bright fields, transmitted bright fields, multi-incident angle dark fields, and orthogonally polarized images at the same glass coordinates. These images are then corrected and registered with a current defect-free reference area to form a multi-optical-state image tensor. Appearance features and optical response trajectories are generated from candidate regions. Channel scales and biases are generated using the reference area baseline and production conditions. Optical state nodes are aggregated according to the incident angle difference, reflection-transmission relationship, and polarization direction difference. This solution combines category prototypes, confidence sets, and controlled parameter updates to complete defect prediction and production line decisions. It possesses the technical advantages of separating surface scattering and absorption changes from polarization stress response, adapting to changes in production conditions, and limiting the degradation of historical recognition capabilities.

[0006] This invention provides a method for intelligent identification of defects in tempered coated glass, comprising: S1. Collect reflected bright field images, transmitted bright field images, dark field images at at least two incident angles, and two orthogonal polarization state images at the same coordinates of tempered coated glass, and associate them with film system identifier, glass thickness, surface curvature parameters, and production batch identifier to form a group of original images with multiple optical states. S2. Based on the defect-free reference area of ​​the current tempered coated glass, the original image group of the multiple optical states is corrected and coordinate registered to form the image tensor of the multiple optical states and the response baseline of the reference area. S3. Input the multi-optical-state image tensor into the trained candidate segmentation network to obtain candidate region masks, extract the appearance features of the candidate region masks and the optical response trajectories arranged according to optical states to form candidate-level feature groups; S4. Encode the reference area response baseline, film system identifier, glass thickness, surface curvature parameters and production batch identifier into a batch condition vector. Generate channel scale and channel offset from the batch condition vector to correct the candidate level feature group. Set the correction features of each optical state as nodes, set the incident angle difference, reflection and transmission relationship and polarization direction difference as node relationships and aggregate node features to form candidate level mechanism fusion features. S5. Compare the candidate-level mechanism fusion features with the defect category prototype and perform defect prediction. Output the defect location, defect category, defect size, severity and confidence set. Generate a pass instruction, review instruction or rejection instruction based on the confidence set and severity. S6. Input the manually confirmed sample, the reference area drift amount of the reference area response baseline, and the defect category prototype drift amount into the controlled update module. When the joint drift triggering condition is met, update the batch condition adaptation parameters, defect category prototype, and layered risk calibration parameters to form update parameters for subsequent tempered coated glass to be used in steps S4 and S5.

[0007] Optionally, S1 includes: The glass conveying direction coordinates are determined by the pulse count of the glass conveying encoder, and the coordinates perpendicular to the conveying direction are determined by the camera calibration parameters; When the coordinates enter each imaging station, the reflected bright field light source, the transmitted bright field light source, the dark field light source with at least two different incident angles and the orthogonal polarization imaging component are triggered sequentially. Transform the encoder coordinates of each exposure to the same glass coordinate system; The system reads the film system identifier and production batch identifier from the production execution system, reads the glass thickness from the online thickness gauge, and reads the surface curvature parameters from the surface profile sensor. After performing a consistency check according to the glass identifier and the acquisition timestamp, it writes the acquisition record of the corresponding image. When any production parameter is missing or the consistency check fails, it writes a parameter invalid mark, terminates the automatic processing of the corresponding glass in steps S2 to S5, generates a review instruction, and prohibits the corresponding glass sample from entering the controlled update in step S6.

[0008] Optionally, S2 includes: The exposure correction coefficient is determined by the ratio of the target response to the measured response of the calibration plate under various optical conditions. The pixel gain and offset are determined by the flat field image and the dark field image. The scale correction amount is determined by the color channel statistics or grayscale statistics of the defect-free reference area. The defect-free reference area is selected from a predetermined area that avoids glass edges and product markings. When the pixel gradient or outlier pixel ratio in the predetermined area exceeds the validity threshold determined by the quantile of the corresponding statistical measure of historical defect-free glass, a predetermined backup reference area is used instead. When the backup reference area still exceeds the validity threshold, the automatic processing of the corresponding glass in steps S3 to S5 is terminated, a review instruction is generated, and the corresponding glass sample is prohibited from entering the controlled update in step S6. The corrected images are resampled and registered based on the calibration transformation and glass coordinates of each imaging station; The luminance statistics, color statistics, texture statistics, and polarization response statistics of the effective reference area under each optical state are calculated respectively to form the response baseline of the reference area; When the correction residual of any image channel exceeds the channel threshold determined by the historical calibration residual quantile, the image channel is marked as invalid, the automatic processing of the corresponding glass in steps S3 to S5 is terminated, a review instruction is generated, and the corresponding glass sample is prohibited from entering the controlled update in step S6.

[0009] Optionally, S3 includes: The edge continuity, aspect ratio, color shift, local texture, and area are calculated from the pixels within the candidate region mask and its boundary neighborhood. The particle scattering morphology is represented by the dark field grayscale attenuation sequence and bright spot dispersion in multiple directions from the center to the boundary of the candidate region, which together form the appearance features. Using the corresponding optical state response of the defect-free reference area as a normalization benchmark, the reflection response difference, transmission response difference, normalized response of each dark field incident angle, texture energy difference along the glass transport direction and perpendicular to the glass transport direction, and orthogonal polarization response difference of the candidate region are calculated. The responses are arranged in the order of reflected bright field, transmitted bright field, dark field incident angle and orthogonal polarization state to form the optical response trajectory.

[0010] Optionally, S4 includes: The batch condition vector is generated as follows: the reference response difference is obtained by subtracting the reference region response baseline from the historical reference baselines stored in layers according to film system and glass thickness; The glass thickness and surface curvature parameters are numerically encoded, and the film system identifier and production batch identifier are categorized. The reference response difference, numerical code, and category code are input into the batch condition adapter, which outputs the channel scale and channel offset that correspond one-to-one with each channel of the appearance feature and optical response trajectory. The channel scale is multiplied by the corresponding channel and then the channel offset is superimposed to obtain the corrected appearance feature and the corrected response features of each optical state. Furthermore, the node relationships include: representing the surface scattering and absorption response by the difference between the reflected bright field correction response characteristics and the transmitted bright field correction response characteristics; obtaining the corrected angle response slope by dividing the difference in correction response characteristics of adjacent dark field incident angles by the corresponding incident angle difference; and representing the polarization stress response by the difference in correction response characteristics of two orthogonal polarization states. The feature difference, corrected angle response slope, and polarization stress response corresponding to each node relationship are used as relationship features. Relationship weights are determined according to the relationship type. The correction response features corresponding to the same candidate region are weighted and aggregated. The aggregation result is then combined with the correction appearance feature and input into the interactive fusion layer to obtain the candidate-level mechanism fusion feature.

[0011] Optionally, S5 includes: The distances between the candidate-level mechanism fusion features and the prototypes of film pinholes, scratches, spots, delamination, particle adhesion, tempered stress patterns, and defect-free categories are calculated respectively. The distances are input into the classification prediction head to obtain the category scores, and the pixel mask and area or length severity are obtained by the segmentation prediction head and the severity prediction head respectively. Inconsistency scores for each category were constructed using calibration samples stratified by film system and glass thickness, and confidence sets were determined according to the quantiles corresponding to the target coverage. A pass instruction is generated when the confidence set contains only the no-defect category; a rejection instruction is generated when the confidence set contains only one defect category and the severity reaches the rejection threshold of that category; and a review instruction is generated in other cases. The calibration sample level, quantile, and threshold used for the judgment are recorded. Furthermore, the training of the candidate segmentation network, batch condition adapter, relation weight, interactive fusion layer, classification prediction head, segmentation prediction head and severity prediction head includes: dividing production samples labeled with glass coordinates, defect category, pixel mask, area or length, severity, as well as film system identifier, glass thickness, surface curvature parameter and production batch identifier into training sample group and calibration sample group. The candidate segmentation loss, category prototype distance loss, defect classification loss, pixel segmentation loss and severity prediction loss are jointly optimized using the training sample group, so that the joint optimization covers the candidate segmentation network, batch conditional adapter, relation weights, interactive fusion layer and each prediction head; The batch category center is obtained by fusion characteristics of the same category in production batches. The batch category center is weighted and averaged according to the number of samples of the corresponding category in each batch to form a global defect category prototype. The batch category center is stored as a historical defect category prototype. The non-consistency score quantiles corresponding to the target coverage and the removal thresholds for each defect category are determined using the calibration sample set.

[0012] Optionally, S6 includes: After a production batch switch, the initial response baseline of the current production batch is established by the response statistics of the effective reference area of ​​the calibration glass that has been manually confirmed to be defect-free within the preset calibration window. The baseline version, calibration window and sample batch identifier are recorded. Calculate the reference area drift of the reference area response baseline relative to the initial response baseline of the current production batch, and the defect category prototype drift of the mechanism fusion feature of the manually confirmed sample relative to the corresponding defect category prototype. The lower limit and upper limit of adaptation trigger are determined by two quantiles of the drift amount of the reference area in the historical verification batch, and the upper limit of playback constraint is determined by the quantile of the drift amount of the prototype of the defect category. When the reference area drift is less than the adaptation trigger lower limit, the current parameters are retained. When the reference area drift is not less than the adaptation trigger lower limit and does not exceed the adaptation upper limit, and the defect category prototype drift does not exceed the playback constraint upper limit, the candidate segmentation network and the feature extraction parameters used to generate the appearance features and optical response trajectories are frozen. Only candidate updated versions of the batch condition adapter, defect category prototype and hierarchical risk calibration parameters are generated, and the playback constraints of the historical defect category prototype are added to the update loss. The weighted sum of classification error rate, segmentation loss and severity error is used as the validation loss. The validation loss of the candidate update version and the current version is calculated using samples from historical validation batches and manually confirmed samples from the current production batch. The candidate update version is deployed only when the validation loss of the current production batch is not greater than that of the current version and the increase in validation loss of historical validation batches does not exceed the replay tolerance determined by the historical replay validation results. Otherwise, the current version is retained, an update anomaly record is generated and a review instruction is issued. When any drift exceeds the corresponding upper limit, parameter updates are stopped and an update exception record and review instructions are generated.

[0013] On the other hand, the present invention also provides an intelligent identification system for defects in tempered coated glass, comprising: The system comprises the following modules: a multi-optical-state acquisition module for generating a set of original images in multiple optical states; a calibration and registration module for generating image tensors and reference area response baselines in multiple optical states; a candidate feature formation module for generating candidate region masks, appearance features, and optical response trajectories; a conditional calibration and mechanism fusion module for calibrating candidate-level feature sets based on batch condition vectors and generating candidate-level mechanism fusion features; a defect prediction and instruction module for outputting defect location, defect type, defect size, severity, and confidence set, and generating pass / fail instructions, review instructions, or rejection instructions; and a controlled update module for updating batch conditional adaptation parameters, defect type prototypes, and hierarchical risk calibration parameters based on reference area drift and defect type prototype drift.

[0014] The beneficial effects of this invention are: 1. By combining the reflected bright field, transmitted bright field, multi-incident angle dark field, and orthogonal polarization state under the same glass coordinate system, and correcting and registering the images of each state, the surface scattering, absorption, angular response, and polarization stress response of the candidate region can be made to be on a comparable coordinate and scale basis.

[0015] 2. By generating channel scale and bias from the response baseline of the defect-free reference area and information on film system, thickness, curvature, and batch, and forming mechanism fusion features according to the relationship of optical state nodes, the influence of production condition distribution changes on candidate characterization can be reduced, and the separability of film defect response and tempered stress texture polarization response can be enhanced.

[0016] 3. By comparing category prototypes, making confidence set decisions, and implementing controlled updates with drift triggers, historical prototype replays, candidate version verification, and rejected deployment branches, uncertain candidates can be directed to manual review and adapted to new production batches under the condition of limited update scope. Attached Figure Description

[0017] 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: Figure 1 This is a flowchart of a method for intelligent identification of defects in tempered coated glass according to the present invention.

[0018] Figure 2 This is a flowchart of the S4 batch condition correction and optical mechanism fusion of the present invention. Detailed Implementation

[0019] 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.

[0020] refer to Figures 1-2A method for intelligent identification of defects in tempered coated glass, comprising: S1. Collect reflected bright field images, transmitted bright field images, dark field images at at least two incident angles, and two orthogonal polarization state images at the same coordinates of tempered coated glass, and associate them with film system identifier, glass thickness, surface curvature parameters, and production batch identifier to form a group of original images with multiple optical states. S2. Based on the defect-free reference area of ​​the current tempered coated glass, the original image group of the multiple optical states is corrected and coordinate registered to form the image tensor of the multiple optical states and the response baseline of the reference area. S3. Input the multi-optical-state image tensor into the trained candidate segmentation network to obtain candidate region masks, extract the appearance features of the candidate region masks and the optical response trajectories arranged according to optical states to form candidate-level feature groups; S4. Encode the reference area response baseline, film system identifier, glass thickness, surface curvature parameters and production batch identifier into a batch condition vector. Generate channel scale and channel offset from the batch condition vector to correct the candidate level feature group. Set the correction features of each optical state as nodes, set the incident angle difference, reflection and transmission relationship and polarization direction difference as node relationships and aggregate node features to form candidate level mechanism fusion features. S5. Compare the candidate-level mechanism fusion features with the defect category prototype and perform defect prediction. Output the defect location, defect category, defect size, severity and confidence set. Generate a pass instruction, review instruction or rejection instruction based on the confidence set and severity. S6. Input the manually confirmed sample, the reference area drift amount of the reference area response baseline, and the defect category prototype drift amount into the controlled update module. When the joint drift triggering condition is met, update the batch condition adaptation parameters, defect category prototype, and layered risk calibration parameters to form update parameters for subsequent tempered coated glass to be used in steps S4 and S5.

[0021] In this specific embodiment, S1 includes: In this specific embodiment, the multi-optical state acquisition module is set in the continuous conveying and detection section of tempered coated glass. When the glass enters the detection section, the production execution system issues glass identification, film system identification, production batch identification and theoretical thickness range. The acquisition controller establishes a multi-optical state acquisition record with the glass identification. The record fields include glass coordinates, optical state identification, camera identification, exposure time, encoder count, acquisition timestamp, film system identification, glass thickness, surface curvature parameters, parameter validity markers and image storage address. The record is written to the acquisition cache with a joint index composed of glass identification and glass coordinates. The glass conveying encoder is installed at the end of the conveying roller shaft. In this specific embodiment, each pulse of the encoder corresponds to a conveying displacement. mm, data acquisition controller according to Calculate the first The transmission direction coordinates corresponding to each pulse ,in This is the current cumulative pulse count. This serves as a reference count latched when the leading edge of the glass passes through the inlet photoelectric switch. and The units are all mm. When the encoder count monotonicity is abnormal or the displacement of the adjacent sampling period exceeds the displacement corresponding to the upper limit of the conveying speed, the effective parameter mark of the current glass is set to 0 and a verification instruction is generated. Coordinates perpendicular to the conveying direction Determined by camera intrinsic parameters, lens distortion parameters, and calibration board control points, the calibration program fits the homography transformation of each imaging station using millimeter coordinates and pixel coordinates on the glass plane. ,in This provides a state index for six optical states: reflected bright field, transmitted bright field, first dark field, second dark field, first polarization, and second polarization. The results are obtained by screening no fewer than 20 control points through least squares and random sampling consistency. A dimensionless matrix is ​​used. The matrix and its version number are saved when the root mean square error of the reprojection is no more than 0.30 pixels. Otherwise, the imaging station is prohibited from participating in automatic detection. In this specific embodiment, a reflective bright field station, a transmittance bright field station, two dark field stations with incident angles of 30 degrees and 60 degrees, and two polarization exposure states with polarization directions differing by 90 degrees are arranged sequentially along the conveying direction. The data acquisition controller saves the calibrated distance of each station relative to the entrance photoelectric switch. ,in Corresponding to the six imaging states mentioned above, Obtained from workstation distance calibration and in mm; when Enter target glass coordinates An exposure pulse is emitted when the corresponding trigger tolerance range is reached. The trigger tolerance is 0.20 mm and is determined by the encoder resolution calibration result. The target sampling location number, The first generation of encoder coordinate tracking The coordinates of the glass plane conveying direction. The first generation of camera calibration coordinates The horizontal coordinates of the glass plane and The units for all images are mm; the images for all six states are in target glass coordinates. Naming allows exposure records from different physical workstations to point to the same glass area; Each industrial camera outputs a 12-bit linear grayscale or color image. In this specific embodiment, the image block size corresponding to each target glass coordinate is 2048 pixels by 2048 pixels. The reflected bright field, transmitted bright field, 30-degree dark field, 60-degree dark field, 0-degree polarization, and 90-degree polarization are respectively identified by status indicators. , , , , and The subscript indicates the dark field incident angle or polarization direction in degrees. The original image value retains the camera output quantization level. After acquisition, it is written into the same original image group index in the order of status identifier. If any status image is missing, it is not filled with other statuses. The online thickness gauge reads the thickness after the leading edge of the glass enters the acquisition section. The unit is mm. The surface profile sensor collects the height at three lateral positions on the glass and obtains the principal curvature by fitting a quadratic surface. The unit is Membrane system identification and production batch identification The data acquisition controller reads the glass work order from the production execution system. , , and The instrument calibration version, sampling timestamp, and glass identification are written into each optical state record, and the thickness and curvature values ​​are used only as production conditions for that piece of glass and not inferred from the image. Consistency verification is based on the following criteria: glass markings are equal; the absolute value of the target coordinate difference after six exposures does not exceed 0.30 mm; the coordinate deviation between the earliest and latest exposure timestamps after conversion according to the conveyor speed does not exceed 0.30 mm; and the recording times of the thickness gauge and contour sensor are within the time window of the glass passing through the acquisition segment. and Non-empty is the pass condition; the verification result is written to... Fields, This indicates that the record is complete and the coordinates are consistent. This indicates that there are missing images, misalignments, missing parameters, or conflicting identifiers. The data acquisition controller performs data acquisition in five states: waiting for glass, coordinate tracking, workstation triggering, record verification, and packaging completion. It reads and updates the encoder count every 1 ms. The status action field includes the status identifier to be triggered, target coordinates, next station identifier, exposure pulse number, and completed status bitmap. The status transition is only allowed when the current station has been completed and the trigger range of the next station has not been crossed. If the target range has been crossed but no exposure completion response has been received, the record verification failure branch will be entered immediately. When a duplicate exposure record is received at the same glass coordinate, the retention conditions are that the camera hardware trigger sequence number is continuous, the target coordinate deviation is minimal, and the exposure time is within the calibrated exposure window. Candidate records are first sorted in ascending order by target coordinate deviation, and then sorted in ascending order by timestamp. The first record is retained and the remaining records are written to the duplicate exposure log. If the sorted first record still does not meet the 0.30 mm coordinate tolerance, the replacement image is not selected and the state is regarded as missing. The determination of the valid parameter marker is performed once after six states are completed. The input fields are the six exposure completion positions, coordinate deviation, and time window. , , and The output actions only include two types: encapsulating the original image group or generating a verification instruction. The verification cycle does not exceed 10 ms and is completed before the glass reaches the processing queue of step S2, so as to avoid unclosed acquisition records from entering subsequent corrections. when When the acquisition controller writes the abnormal field, trigger count, and six status image addresses into the review queue, it terminates the automatic processing of the glass block in steps S2 to S5, sends a review instruction to the production line controller, and writes a prohibition on updating flag into the controlled update sample table; when At that time, according to , , , , and Fixed-order encapsulation of multiple optical state raw image groups ,Will The collected data is then transferred to step S2.

[0022] In this specific embodiment, S2 includes: Step S2 reads the output of step S1 and Multi-optical state raw image group And read the target response, measured response, flat field image, dark field image, and homography transformation of the calibration board from the most recent valid calibration version according to the optical status identifier. The calibration and registration module establishes processing records based on glass identification, status identification, and calibration version to ensure that the six states of the same piece of glass use calibration parameters within the same validity period. State pixels First, set the center 80% of the effective pixels of the same exposure flat field image. Calculate the effective flat field scale ,in Excluding saturated pixels, bad pixels, and non-finite pixels, and containing at least 10,000 pixels. The unit is quantization level and limited to 512 to 3584 quantization levels; then press Perform joint correction of dark level, flat field and exposure, where It is a 12-bit raw quantization level. and These represent the quantization levels for dark and flat areas under the same exposure settings. To calibrate the target response of the board Compared with the measured response The ratio, Quantization levels are used to prevent the denominator from being zero; the ratio term is dimensionless and multiplied by... back Still in the 0 to 4095 quantization range. It is dimensionless and limited to 0.5 to 2.0. Insufficient or out-of-range samples, non-positive denominator or When the range is exceeded, the status will be changed. Marked as invalid; The defect-free reference area is predefined according to the product pattern, avoiding the glass edge by 50 mm and the product marking area, and has an area of ​​not less than [missing information]. The four rectangular candidate regions are used, and the calibration and registration module calculates the distance from the center of each candidate region to the current target coordinates. Sobel gradient mean and the proportion of outlier pixels that deviate from the median by more than three times the absolute deviation of the median. ,in The unit is mm. The unit is per quantization level per pixel. A dimensionless ratio ranging from 0 to 1; The historical reference layer selector first checks whether there are at least 30 historical glasses within the same film system and the same 0.5 mm thickness layer that have been manually confirmed to be defect-free and valid in all six states. If this condition is met, the accurate layer is selected. If the accurate layer does not meet the sample threshold, adjacent layers within the same film system are searched in ascending order of thickness layer center distance. Only the first valid layer with a distance not exceeding 0.5 mm and at least 30 samples is allowed. If the center distances are the same, the first item is selected in ascending order of thickness layer center value. If no qualified adjacent layers are found or there are no film system identifiers, an equipment-level conservative reference layer is selected, established by at least 30 production calibration glasses under the same imaging station and the same calibration plate response version, and manually confirmed to be defect-free. The conservative reference layer sets the states... and The 90th percentile is used as the validity threshold, the 25th percentile of the median of each channel is used as the scale correction baseline, and the median and interquartile range of the reference area response components are used as the historical reference baseline. The selector sequentially records the source layer, membrane system identifier or equipment identifier, thickness interlayer distance, number of samples, sample batch range, calibration version, baseline version, and effective time, and writes the out-of-distribution marker. If none of the above three sources meet the threshold, steps S3 to S5 are immediately terminated, a glass-level review instruction is generated, and automatic acceptance, automatic rejection, and controlled update in step S6 are prohibited. When using an adjacent layer or a conservative reference layer at the equipment level, processing is allowed to continue, but the minimum glass-level instruction is limited to review, and the source layer and out-of-distribution marker are changed accordingly. , The transfer proceeds to steps S4 and S5, where The multi-optical state image tensor, subsequently formed by stacking the six verified state-corrected images in a fixed order in step S2, has a spatial size of 2048 by 2048, a state dimension of 6, and retains the quantization level for each pixel. The reference area response baseline vector is formed by arranging the state statistics of the effective reference area according to the state identifier and the statistical identifier, and the unit of each component is stored in a field dictionary; the source level and distribution external marker are first written into the processing record of step S2, and then... and When the data is generated, it is passed along with the associated field. The manual confirmation result is only used as the accumulation of new batch candidate samples in step S6 and is not used to directly update the model. Reference area validity threshold and Defect-free glass was manually verified by taking selected historical reference layers from either the precise layer or adjacent layers. and The 95th percentile is used, and the aforementioned 90th percentile is used under conservative source conditions at the equipment level; the threshold record includes source level, membrane system identifier or equipment identifier, thickness layer, number of samples, sample batch range, quantile, calibration version, baseline version, and effective time; the calibration and registration module first... Ascending order Ascending order The ascending order and the ascending reference area number form a unique candidate order, and the first one that satisfies this order will be selected. and The candidate is determined as the main reference area; if the first candidate does not meet the threshold, the subsequent candidates are examined in turn and the first valid candidate is used as the backup reference area. If the backup reference area is separated from the first candidate by less than 100 mm laterally, the next candidate is examined. If there is no valid backup reference area in the candidate order, an invalid reference area mark is written, the automatic processing of steps S3 to S5 is terminated, a review instruction is generated and the sample is prohibited from entering step S6. Calculate the median for each state and color channel within the valid reference area. and with Perform current glass dimensional correction, among which Index for grayscale or color channels. The unified historical reference layer selector has been determined and is consistent with... , Scale correction baseline with the same baseline version and calibration version: For the precise layer source, take the median of the corresponding channel of the same film system and the same thickness layer; for the adjacent layer backtracking, take the median of the corresponding channel of the selected adjacent layer of the same film system; for the equipment-level backtracking, take the 25th percentile of the median of the corresponding channel of the aforementioned production calibration glass; do not reselect the reference layer in this formula. and All units are in the quantile order, and the proportional term is dimensionless. For the corrected quantization level, the scale correction record saves the source hierarchy, number of samples, baseline version, and calibration version; when If the quantization level is 50 quantization levels below the lower limit of the calibrated dynamic range, the channel is directly marked as invalid. During coordinate registration, the state will be... The corrected image is obtained through The image is transformed to a glass coordinate grid based on the reflected bright field state, and then bilinear interpolation is used to generate a resampled image. In each calibration cycle, the registration residuals are calculated using no fewer than 30 common control points. ,in For state The Middle Each control point pixel coordinate, The coordinates are the coordinates corresponding to the baseline state. and They are respectively x and y coordinates and Control points respectively The transformed x and y coordinates, all in pixels. To control the number of points, The unit is pixels; Channel residual threshold Take the 99th percentile of the most recent 20 valid calibration residuals for this imaging station and limit it to the range of 0.30 to 0.80 pixels. When there are fewer than 30 control points or the effective coverage of resampling is less than 98%, the status will be changed. All channels are marked as invalid, residuals, thresholds and calibration versions are recorded, automatic processing from steps S3 to S5 is terminated and a review instruction is generated, and a prohibition on updating flag is written. The six state images that passed the verification were processed by... , , , , and Sequential stacking forms a multi-optical-state image tensor The tensor has a spatial size of 2048 by 2048, a state dimension of 6, and the color image retains three color channels inside the state. The grayscale image is saved as a single channel and recorded in the channel description table. There is no need to copy the channels to change the dimension. The above procedure still applies when all four predefined reference areas are valid. , , The unique sorting of the reference area number is determined, and the first sorting digit is selected. The calibration and registration module saves all four candidate sorting keys, validity results, and the final selected area number. When the first sorting digit is invalid, only the elements that satisfy the criteria are selected. , The candidate for the lateral separation condition is selected in the original sorting order. If the candidate set is empty, the output reference area is invalid and the full glass mean is not used as a substitute. The calibration and registration module calculates the mean luminance, standard deviation luminance, median of the Lab color channel, the first 16 texture components of the local binary mode histogram, and the mean and standard deviation of the polarization response difference between 0 degrees and 90 degrees for each state within the effective reference area. These statistics are then arranged into a reference area response baseline vector according to the state identifier and the statistical identifier. Each dimensional statistic retains its quantization level or color difference unit and is recorded in the field dictionary. Before entering the subsequent model, it is normalized according to the training set statistics. Final processing record saving , Valid reference area coordinates Version, Version, , , And all valid tags, the multi-optical-state image tensor is only transferred when all six states are valid. and reference area response baseline The objects are passed as pairs to steps S3 and S4.

[0023] In this specific embodiment, S3 includes: In step S3, the candidate feature formation module reads the multi-optical state image tensor output in step S2. The images of each state are standardized according to the channel mean and standard deviation saved during the training phase, and divided into inference blocks of 512 pixels by 512 pixels with an adjacent overlap of 64 pixels according to the glass coordinates. The inference record saves the glass identifier, the coordinates of the top left corner of the block, the channel range of the six optical states and the model version. The overlapping area prediction is weighted and fused according to the cosine weight of the distance from the block boundary when recovering the whole image. The candidate segmentation network employs a U-shaped convolutional network with a residual encoder. The input is a tensor concatenated from six optical states in state order, and the encoder output scale is equal to the input. , , and The decoder utilizes multi-scale features to recover spatial resolution and output candidate probability maps via skip connections. The training samples are production samples with glass coordinates and pixel mask annotations. Positive samples cover pinholes, scratches, spots, delamination, particle adhesion and tempered stress textures in the film layer, while normal areas are used as negative samples. During inference, candidate thresholds are used. right Binarization; The threshold is determined solely by validation samples outside the training sample group. The threshold search set is a discrete grid of 0.05 to 0.95 with a step size of 0.01. The maximum value is selected from the thresholds where the candidate recall rate is not less than 99% and the average number of false positive candidates per glass block is not higher than the baseline version. If multiple records have the same threshold due to numerical precision, the first one is selected in ascending order of model version and validation batch number. In this specific implementation, the validation sample and model version are obtained according to this rule, resulting in a value of 0.45. Subsequently, a closing operation with a radius of 2 pixels is performed, and isolated connected components with an area less than 9 pixels are deleted. Connected components in the overlapping area of ​​the inference block are merged according to the intersection-union ratio greater than 0.5. Regions below the threshold are not included in the subsequent defect classification, but their maximum probability and coordinates are still written to the low response log. During training, the defect pixel mask corresponding to each inference block is scaled to the decoding scale. The candidate segmentation loss is composed of equally weighted binary cross-entropy and Dice loss. The optimizer uses the initial learning rate. Furthermore, the validation loss is halved if it does not decrease for five consecutive rounds. The system is trained for a maximum of 100 rounds, retaining the model with the highest validation candidate recall and a false positive rate not exceeding that of the baseline version. The calibration sample group does not participate in network gradient updates or for determining... This is only used for the risk calibration parameters and rejection threshold in step S5; When adjacent inference blocks provide multiple masks for the same candidate, the masks are first restored to glass coordinates. Candidate edges are then established from high to low based on the intersection-union ratio (IU). Records with an IU exceeding 0.5 and a candidate center distance of no more than 2 mm are grouped into the same connected group. Within the group, the mask to be retained is determined by the maximum weighted probability, and pixels are merged. If the candidate edge does not reach the threshold, it remains an independent record, and the probability within the group is all lower than the threshold. Candidate regions are not generated and written to the low-response log at times; For the first Candidate region mask The candidate feature formation module calculates the area of ​​its 5-pixel outward-extending boundary neighborhood. and aspect ratio ,in For the number of mask pixels, and To obtain the horizontal and vertical millimeter-per-pixel scales for calibration, Units are , and The length and short sides of the smallest bounding rectangle are given in mm. mm represents the zero protection quantity in the same unit; Edge continuity According to the distance between the mask boundary and the adjacent boundary point not exceeding The number of mm points is calculated as a proportion of the total number of boundary points, with a value ranging from 0 to 1, and the larger the value, the more continuous the boundary. The color offset is composed of the difference between the median of the Lab channels of the candidate area and the effective reference area, and the unit is Lab color difference. The local texture uses the contrast, homogeneity and energy of the gray-level co-occurrence matrix and the local binary mode histogram. All texture quantities are normalized according to the corresponding quantities in the reference area with the same dimension ratio. Particle scattering morphology was read along four directions: 0°, 45°, 90°, and 135°, with the candidate center as the origin. and In the dark field state, the grayscale sequence extends 10 pixels from the center to the outer edge of the mask, and is processed in each direction by... Calculate the normalized dark field response, where For 30-degree or 60-degree dark field indexes, The pixel position along the ray. This corresponds to the median of the reference range. The values ​​are dimensionless; the exponential decay fitting coefficients and fitting residuals of the sequence, along with the dispersion of bright spot coordinates above the 99th percentile of the reference area, are written into the scattering feature field. Calculate the candidate region in both reflection and transmission states separately. and ,in and The average quantization level of the candidate region. and The average quantization level of the effective reference area of ​​the same glass. Quantization level and All are dimensionless and trimmed to ; Calculation of dark field state And cut to ,in To correct the dark field image in the candidate mask The average quantization level within; the difference in texture energy of the gray-level co-occurrence matrix along and perpendicular to the conveying direction. Two of these energies are dimensionless quantities ranging from 0 to 1; let and Polarization images corrected at 0 degrees and 90 degrees respectively. The average quantization level within, let and To calculate the average quantization level corresponding to the effective reference area, respectively and And cut to ; then calculate And cut to The aforementioned dark field and polarization inputs are both derived from the corresponding corrected state images and are expressed in quantization units. , , and All are dimensionless; The appearance features are composed of edge continuity, aspect ratio, color shift, local texture, area, four-directional dark field attenuation parameters, and bright spot dispersion in a fixed field order. ,Will , , , , , , and The optical response trajectory is formed in the order of incident angles of reflected bright field, transmitted bright field, dark field, and two orthogonal polarization states. Missing fields are not replaced with 0; instead, the candidate is marked as an invalid feature and a review instruction is generated. Candidate-level feature groups Also preserve candidate masks and glass coordinate bounding box The candidate feature table also includes a multi-scale decoding feature reference field that corresponds one-to-one with the candidate identifier. This field stores the storage address, content hash, tensor shape, source inference block coordinates, and candidate segmentation network version of the four scale features of the decoder in step S3. When the same candidate crosses inference blocks, alignment and fusion are first performed according to the glass coordinates before a unique reference is generated; all features of the same glass... The candidate centers are sorted according to their transport direction coordinates and written into the candidate feature table. The candidate feature table also stores the glass-level candidate count, candidate threshold, connected component filtering rule version, and low response log summary. When the candidate count is 0, an empty candidate state is written. No default candidate, default feature, or default confidence set is constructed, and the empty candidate state is directly passed to the glass-level instruction decision branch in step S5. No candidate-level record is generated in step S4. The non-empty candidate table, along with the multi-scale decoding feature reference, candidate segmentation network version, normalized parameter version, and reference area response baseline identifier, is passed to step S4.

[0024] In this specific embodiment, S4 includes: Step S4 reads the candidate feature table from step S3 and the reference area response baseline from step S2. Historical reference source hierarchy and distribution markers; thickness layers are divided with a width of 0.5 mm and the boundaries are included in the numerical upper thickness layer; historical reference baselines. Must be determined with step S2 , and The selected reference layer and baseline version must be consistent and cannot be independently changed in step S4: when the precise layer is available, read the layer of the same film system and thickness; when reverting to adjacent layers, read the adjacent layers of the same film system determined in step S2, and identify the film system. If no qualified adjacent layer is found or there is no qualified adjacent layer, the equipment-level conservative reference layer of the same imaging station determined in step S2 is read; the fields of the historical reference baseline table include source layer, film system identifier or equipment identifier, thickness layer, interlayer distance, median and interquartile range of each state statistic, number of samples, sample batch range, calibration version, baseline version and effective time; if the source layer, number of samples or version passed in is inconsistent with the table entries, or if the selected table entries are missing, the glass is transferred to the review and automatic qualification, automatic rejection and controlled update are terminated; when using adjacent layers or equipment-level conservative reference layers, the interlayer distance, source layer and distribution out-of-distribution marker are written into the condition vector and passed to step S5; Reference response difference Calculation, where For the reference area response baseline component index, The historical interquartile range of the corresponding component. Take 1% of the standard deviation of the training samples under this unit of measurement to make the units of the numerator and denominator consistent. Dimensionless and trimmed to The historical table is missing components or the interquartile range is 0 and there are no valid components. The glass will then be transferred for review; Glass thickness and surface curvature Encode the training samples by mean and standard deviation respectively. and Both are dimensionless and are membrane system identifiers. and production batch identification The model is encoded using 16-dimensional and 8-dimensional embedding vectors obtained through training. The model version simultaneously stores a membrane system category dictionary, a production batch category dictionary, and their respective reserved unknown category embeddings. During training, 5% of the category labels are deterministically masked as unknown labels to learn the reserved embeddings. At runtime, the model first searches according to the category dictionary version, and if no match is found... or Using the corresponding unknown category embedding of the same deployed model version, without mapping to any existing batch, and writing the category distribution out-of-label, original identifier, and dictionary version; batch condition vector Depend on , , The values ​​are obtained by splicing the membrane system embedding, batch embedding, thickness interlayer distance and reference baseline version identifier. When the value exceeds 5 times the standard deviation of the training distribution, the clipped value is retained and written into the out-of-distribution label. When the category or out-of-distribution label exists, the current complete deployed parameter group is still used to complete steps S4 and S5, but the minimum instruction for glass level is verification. The label is passed to step S5 along with the candidate record, and the manual confirmation result continues to be accumulated as a new batch of samples in step S6. The batch condition adapter is a two-layer fully connected network. The first layer uses 64 ReLU units, and the second layer outputs and appearance characteristics... and optical response trajectory The same parameter twice the total number of channels, the first half through Obtain the channel dimensions, the latter half via The channel bias is obtained, where For feature channel index, and For the adapter's dimensionless output, Limited to 0.5 to 1.5, Limited to -0.5 to 0.5; For candidate features that have been standardized to dimensionless according to training statistics implement ,in Covering by fixed index and All channels, For the corrected dimensionless features; scale, bias, adapter version and out-of-distribution markers are written together with candidate identifiers into the conditional correction record. When out-of-distribution markers are present, subsequent identification is still performed, but the lowest level of the final instruction is limited to a review instruction. For each candidate, six optical state nodes are established, with the node order as follows: , , , , and The node features are formed by splicing together the correction response features of the corresponding state, the shared correction appearance features, and the state embedding, whereby... and The nodes read the optical response trajectory in step S3 respectively. and The channel is obtained through the batch condition adapter. and The characteristics of the relationship between reflection and transmission are defined as follows: The slope of the corrected angle response is obtained by dividing the difference in the corrected response characteristics of adjacent dark field incident angles by the corresponding incident angle difference. The angle difference is in degrees. The standardized response, measured in units of one degree, is defined as the polarization stress response. ; Before entering node aggregation, relation features are standardized according to the mean and standard deviation of the same relation type in the training set. The original value, standardized value, relation type, and source state are preserved for the reflection-transmission difference, corrected angular response slope, and polarization stress response, respectively. The standard deviation is less than [value missing]. The relation dimension is deleted during training and written into the model field list. At runtime, the relation vector is constructed only according to the list to avoid invalid constant dimensions from entering the weight calculation. The node relationship type table encodes the reflection and transmission relationship, dark field angle relationship, and orthogonal polarization relationship as follows: , and Each relation uses an independent linear transformation. With relation vector Calculate bounded gate weights , where logic gate functions , The input is a dimensionless scalar-gated input formed by relation type, two node features, and standardized relation features. and For node indexing, The node is a dimensionless feature. For correspondence features, For nodes The set of adjacent nodes, This represents vector concatenation. Strictly defined as 0 to 1 and not subject to softmax normalization within its neighborhood, therefore even With only one paired neighbor, the gating weight is still determined by the relation type and It is jointly determined, but not always equal to 1; Aggregate node features by Calculation, where The self-node transformation matrix, relation weights and matrix parameters are obtained from joint training and saved with the model version; when a relation feature contains non-finite values ​​or the source state is marked as invalid, the single relation is not deleted and the reasoning continues, but the entire candidate is set as feature invalid and a review instruction is generated to prevent incomplete node sets from changing the meaning of relation weights; Optical response trajectory in step S3 Use a fixed base field sequence number. and They occupy channels 7 and 8 respectively; both are standardized using the mean and standard deviation of their respective training sets saved with the model version, and then sorted according to their corresponding channels. and Obtain the correction value and ; six Perform average pooling and max pooling sequentially on fixed nodes, and then successively connect the resulting vectors with the corrected appearance features, ... and The concatenated input is an interactive fusion layer containing 256 ReLU units and 128 linear output units. The two correction values ​​are each consumed only once at the concatenation position, without being repeatedly written to a single state node or a three-class relation vector. The concatenation record stores two fields (sequence number, source candidate identifier, standardized parameter version, batch condition adapter version, and interactive fusion layer model version) according to the candidate identifier. If any field, parameter, or version is inconsistent, the candidate is set as an invalid feature and a review instruction is generated. The interactive fusion layer forms a 128-dimensional candidate-level mechanism fusion feature. Its parameters share the same type version identifier as the candidate segmentation network, batch conditional adapter, relation weights, and each prediction head, and are updated by backpropagation of the joint loss during training. Candidate-level mechanism fusion records and save glass identifiers, candidate identifiers, glass coordinate bounding boxes, , , The relationship weight summary, historical reference baseline level and model version are retained as is, and the storage address, content hash, tensor shape and candidate segmentation network version of the multi-scale decoding features generated in step S3 are also retained as is. The results are written into the mechanism fusion feature table according to the candidate identifier and the valid records are passed to step S5. The out-of-distribution markers, invalid feature markers, abnormal relationship records and missing or inconsistent multi-scale decoding feature references are also passed to the instruction decision branch of step S5.

[0025] In this specific embodiment, S5 includes: Step S5 reads the candidate-level mechanism fusion features output in step S4 according to the glass identifier and candidate identifier. The multi-scale decoded feature references passed in their original form are verified to be consistent between the candidate segmentation network versions, candidate bounding boxes, and content hashes of both before allowing the two prediction heads to read. If the references are missing, the content hashes do not match, the candidate coordinates cannot be aligned, or the model versions are inconsistent, the candidate is set as an invalid feature and a verification instruction is generated, without outputting the automatic rejection result. The classification prediction head reads the prototypes of the current model version from the prototype table for seven categories: film pinholes, scratches, spots, delamination, particle adhesion, tempering stress lines, and no defects. ,in For category indexing, and All are 128-dimensional dimensionless vectors. The prototype table simultaneously stores the category name, model version, source batch range, sample count, and update time. Before a model version is qualified for deployment, it undergoes seven types of prototype integrity gates, requiring that all seven categories of records exist, the sample count is positive, the prototype is a finite 128-dimensional vector, and the versions are consistent. If any condition is not met, the candidate version cannot be deployed. At runtime, if the prototype cache of a deployed model is missing, corrupted, or the version is inconsistent, the entire glass will not perform distance, category score, softmax, or confidence set calculations, and the seven related fields will be marked as unavailable and a review instruction will be generated, prohibiting automatic qualification and automatic rejection. Mechanistic fusion features are calculated based on the dimension-wise mean stored in the training sample group before distance calculation. and standard deviation Normalization ,in The dimension index ranges from 1 to 128, and all quantities are dimensionless. For dimensionless zero-protection quantities, the normalized results are trimmed to... Category prototypes use the same group and generate; Only after passing the seven types of prototype integrity access control can the candidate be... With category Calculate the normalized Euclidean distance ,in For category prototype The A dimensionless component Dimension index from 1 to 128, The smaller the value, the closer the candidate is to the class prototype; the classification prediction head uses... Calculate the category score, where and The classification parameters are obtained through joint training. The non-negative distance weights, calibrated to 0.8 in this specific implementation, are used to obtain the class probabilities by applying a softmax function that includes all seven classes. A higher probability indicates stronger category support. No category may participate in normalization using default distance, default score, or absence. The segmentation prediction head reads four-scale tensors from the verified multi-scale decoded features based on the candidate identifier. After performing region alignment with the candidate glass coordinate bounding boxes, it performs step-by-step upsampling and skip fusion, outputting a pixel mask that corresponds one-to-one with the current candidate and uses the same glass coordinate system. The defect area was calculated in glass coordinates using a pixel threshold of 0.5. ,in for The number of pixels with a value of 1 that is a non-negative integer. Units are For scratch and stress texture categories, the mask skeleton is formed sequentially by adjacent points. Calculate the length of each pair of points. The unit is mm, where From 1 to The index of adjacent skeleton point pairs, and For pixel coordinate differences, empty skeleton or single-point skeleton is taken. and ; The severity prediction head reads the 128-dimensional normalized mechanism fusion features, classification probability vector, region pooling features of the segmentation prediction head output mask, and area or length normalized values ​​calculated from the mask, based on the candidate identifiers of the same model version. It then concatenates the four inputs in a fixed field order and outputs a dimensionless severity score from 0 to 1. No calculation is performed if any input is missing, not finite, or version inconsistent. Then it switches to the review branch. The training labels are determined by the normalized value of the area or length relative to the tolerance of the same type of process and the manual grade. At runtime, it also retains , and The fields for the rejection threshold table for area-type defects include membrane system, thickness layer, defect category, area threshold, severity threshold, version, and source sample. For length-type defects, the corresponding length threshold is saved. When the membrane system and thickness layer are not hit, no extrapolation is performed across layers and a review instruction is generated. During the training phase, samples labeled with glass coordinates, defect type, pixel mask, area or length, severity, and production conditions are grouped by production batch. 70% of the batches constitute the training sample group, 15% of the batches are used for verification, and the remaining 15% of the batches constitute the calibration sample group. Samples from the same production batch are not allowed to cross groups. Each sample saves the original annotator, reviewer, annotation version, and production conditions. Samples with conflicting labels are subject to manual arbitration before training. Samples that have not completed arbitration are not included in any data group. Joint losses Calculation, where For candidate segmentation loss, For class prototype distance loss, Cross-entropy for defect classification, For pixel segmentation loss, For severity error, each term is first divided by the mean of the first round loss of the training sample group to normalize it to dimensionless. The normalized baseline is saved according to the model version, and the weights are determined by grid search of the validation set. Production batches in the training sample group and categories The average value of the mechanism fusion features is used to obtain the batch category center. ,definition For training sample groups, the class A batch set with valid samples, then sorted by Form a global defect category prototype, in which For batch Medium category The number of valid samples, and All are 128-dimensional dimensionless vectors; when empty or Time not calculated , categorize The candidate model version is marked as prototype unavailable and prohibited from being deployed until valid training samples are added and the model version is regenerated; if this state occurs in the currently deployed version due to runtime integrity failure, the entire glass is promoted to the review according to the aforementioned integrity failure branch, and the seven distances, class scores, probabilities and confidence sets are all recorded as unavailable, without being padded with 0 or any default value; Each batch category center is stored as a historical defect category prototype. The storage record uses model version, production batch identifier, and category identifier as a composite key, and the fields include... , Membrane system, thickness layer, curvature range, generation time, and training sample summary; batch category centers with fewer than 20 samples are still written into the history table but marked as low sample centers, and do not participate independently in global prototype weighting but are incorporated into the statistics of adjacent batches in the same membrane system; The confidence set was constructed using calibration samples stratified by film system and 0.5 mm thickness layer, for the true class. calibration samples Define Inconsistent Scores ,in A value between 0 and 1, with larger values ​​indicating greater inconsistency between the model and the true class; the target coverage is set to 95%, and the 95th percentile is adjusted using a finite number of samples for each level and class. When there are fewer than 30 single-layer, single-type calibration samples, they are merged into adjacent thickness layers of the same film system and the actual layer level is recorded. The class probabilities of calibration samples are obtained through a single inference from the same model version frozen after training. The calibration process does not update the candidate segmentation network, adapter, relation weights, interaction fusion layer, or prediction head; each Record the actual number of samples, sorting position, quantile calculation rules, and model version. Recalculate when the stratified sample size changes by 20% or the model version changes. Not usable across model versions; For candidates Constructing confidence sets ,in Traverse the seven categories, Quantile records from candidate corresponding film systems and thickness layers, Includes all categories that satisfy the condition that the inconsistency score does not exceed the threshold; if If the result is empty, contains more than two categories, the calibration level used is merged, or step S4 has an out-of-distribution marker, then the candidate is determined to be uncertain and enters the review branch; When multiple candidate masks overlap in the glass coordinate system and are predicted to be of different categories, they are not directly covered according to category scores. Instead, each candidate's confidence set and mask are retained. If the cross-union ratio (CUI) is greater than 0.7, the conflicting candidates are merged into a single review record and all category probabilities are saved. If the CUI is not greater than 0.7, the independent defect records are retained. This conflict rule only changes the organization of the review record and does not change the physical size and severity of each candidate. The production line instruction generator uses data acquisition valid bits, correction valid bits, feature valid bits, and distribution external markers. The decision record consists of the comparison results of size, predicted defect category, physical size threshold and severity threshold. The decision rule version and model version are stored together. If any necessary field is missing, the default value is not used. Instead, the glass level instruction is promoted to a review instruction. Step S5 first reads the glass-level candidate count and empty candidate state passed in step S3; when the candidate count is 0, it does not proceed to the calculation of candidate-level distance, prediction head, or confidence set, nor does it construct a default candidate or default confidence set; when the acquisition valid bit, correction valid bit, and candidate formation processing state are all valid and there are no out-of-distribution markers, it writes the zero-candidate glass-level result and generates a qualified instruction; if any necessary state is missing, invalid, or has an out-of-distribution marker, a review instruction is generated; the zero-candidate result saves the candidate count, Connected component filtering rule version, low response log summary, valid acquisition bits, valid correction bits, candidate formation processing status, out-of-distribution markers, and instruction rule version; candidate count greater than 0 and all candidates in the same glass. A qualified instruction is generated when all categories contain only the defect-free category and there are no out-of-distribution anomalies in acquisition, correction, features, or distribution; when a candidate It contains only one defect category, and corresponds to or Reaching the physical size rejection threshold for this category and When the severity threshold is reached, a rejection instruction is generated; in other cases, a review instruction is generated. Multiple candidate instructions are merged into a glass-level instruction based on the priority of rejection over review and review over pass. After the production line controller returns the instruction reception sequence number, the defect prediction and instruction module writes the reception sequence number into the glass-level result; if no reception response is received within 100 ms, it is resent once. If there is still no response the second time, the glass is kept in the physical verification channel and the communication abnormality is recorded. Unconfirmed qualified instructions are not regarded as released results. Defect results are recorded, including glass coordinate bounding box, pixel mask address, and prediction category. , , Seven types of probability, seven types of prototype distance, confidence set Calibration sample level, target coverage, quantiles The physical size threshold, severity threshold, and model version are used to send glass-grade pass / fail instructions, review instructions, or rejection instructions to the production line controller; after manual review, the label is confirmed to match the corresponding label. Write the manually confirmed sample table as the controlled update input for step S6.

[0026] In this specific embodiment, S6 includes: The controlled update module in step S6 only reads the samples formed in step S5 and manually confirmed, and excludes records of invalid parameters in step S1, invalid reference area or channel in step S2, invalid features in step S3, and abnormal relationships in step S4; after the production batch is switched, a calibration window containing 30 manually confirmed defect-free calibration glasses is established, sorted by acquisition time, and the start and end times of the window, film system identifier, thickness layer, sample batch identifier, and current model version are saved; Response baseline of the effective reference area for each glass within the calibration window Calculate the median of each component to form the initial response baseline for the current production batch. And save the median absolute deviation by component. ,in Baseline component index and Corresponding Units are consistent; the initial response baseline record uses the batch identifier and incremental version number as a joint key. If the window has less than 30 blocks or the effective samples of any key optical state are less than 25 blocks, a new baseline is not established and the review mode is maintained. The drift of the current glass reference area is calculated. ,in The number of baseline components in the reference area response. , This is the standard deviation of the component in the historical defect-free sample. This ratio is half the calibrated resolution of the corresponding sensor and is strictly greater than 0. Both ratios are in the same unit as the corresponding baseline components, thus ensuring the denominator is strictly greater than 0. Each ratio is dimensionless. This is a non-negative dimensionless quantity, and the larger the value, the more significant the deviation of the current reference response from the initial baseline of the batch. This value, along with the glass identifier and baseline version, is written into the drift record, and the drift rolling window is formed by the 30 most recent valid glasses after the initial baseline was completed. ,in Glass labels and corresponding tags arranged in ascending order of collection time. A record set that is not a numerical matrix and has no dimensionless units; when Take the window if it contains at least 25 pieces of usable glass. The median is used as the batch trigger value. When there are fewer than 25 windows, Mark as unavailable, keep in review mode, and prohibit the creation of candidate update versions; For manual confirmation category samples Standardized features are generated based on the one-dimensional mean and standard deviation saved in step S5. And calculate the prototype offset. ,in For the current version category The prototype's first A dimensionless component Dimension index from 1 to 128, It is dimensionless; the median shift of the class shift is taken as the average shift of the same type of sample among the most recent 100 manually confirmed samples. ,definition The defect category prototype drift is obtained by weighting the effective category set with at least 10 samples of each category and then averaging the results based on the number of effective samples. , It is a non-negative dimensionless quantity, and the larger it is, the greater the deviation of the current batch features from the current category prototype; When empty Mark as unavailable without assigning 0, retain the current parameters and category prototype, generate insufficient sample record and review instruction, and prohibit the creation of candidate update versions until at least one category reaches 10 valid samples and then recalculate; Adapt to trigger lower limit and the upper limit of adaptation Take from historical verification batches respectively The 75th and 95th quantiles, and the upper limit of the replay constraint. Take from historical verification batches The 95th percentile threshold table stores information on the film system, thickness layers, historical batch range, quantiles, sample size, version, and effective date, meeting the requirements. and Only enabled; updates will stop and update exception records and review instructions will be generated when the precise level is not hit or the threshold order is abnormal. The controlled update module hits only one branch for each batch state according to a fixed priority: the first priority is... or If unavailable, no candidate version is created, no data is written to the training queue, and records of insufficient window or insufficient samples and verification instructions are generated respectively; the second priority is when both drifts are available but... or At this point, the out-of-limit exception branch is executed; only if the first two items are not hit, the following normal hold and candidate update branches are evaluated; when and When available, retain the current batch condition adaptation parameters, defect category prototypes, and stratified risk calibration parameters; do not create candidate versions or write update anomaly records; samples that have been manually confirmed and passed quality filtering will only continue to accumulate; when , and and Candidate update versions can be created when available, freezing the candidate segmentation network and all feature extraction parameters used to form appearance features and optical response trajectories. Only the two-layer parameters of the batch conditional adapter and the sufficient class of samples can be updated. and current membrane systems and thickness layers And a rejection threshold, candidate versions are not invoked in online inference before verification; the number of categories participating in prototype updates and with sufficient samples. It must be a positive integer. When not creating candidate update versions and generating insufficient sample records; Candidate updates are trained using the joint task loss of the current batch of manually confirmed training samples plus historical prototype replay constraints, and are only updated when... Press at time Calculate the playback item, where The number of categories that participate in the update and have sufficient samples. The prototype component index is from 1 to 128. and Candidate version and current version categories respectively The prototype's first A dimensionless component The dimensionless mean square error of all prototype components involved in the update category; the update loss is... The coefficient 0.2 is determined by a grid search that limits the increase in loss from older batches on historical verification batches; Verify loss Calculation, where For classification error rate, To normalize the segmentation loss, The severity mean absolute error is a dimensionless value divided by the severity range; all three terms are between 0 and 1, with smaller values ​​indicating better validation results. The controlled update module performs independent validation on the current production batch's independent validation set. and fixed historical independent verification batch samples The above calculates the candidate version and the current version , and None of them are involved in fitting candidate parameters, category prototypes, calibration quantiles, or elimination thresholds; Before candidate updates, a mutually exclusive set is pre-divided using a fixed hash algorithm and a fixed salt value: the current production batch is divided according to the last bit of the glass identifier hash, with the last bit 0 to 5 entering the update training set. The last digits 6 to 7 are included in the current batch of independent calibration sets. The last digits 8 to 9 are included in the current batch of independent validation sets. Historical production batches are categorized by the last digit of the batch identifier hash; batches ending in 0 to 5 are included in the historical playback training set. The last 6 to 7 are entered into the historical independent calibration set. The last digits 8 to 9 are entered into the fixed historical verification set. and define independent calibration sets. The current data uses glass identifiers as the mutually exclusive granularity; the same glass and all its candidates cannot cross [various categories]. , and The same current production batch is distributed across these three sets according to the design; historical data uses production batch identifiers as mutually exclusive granularities, and the same historical production batch cannot be distributed across different sets. , and The six-set zero-intersection check uses a joint key composed of the data source domain, production batch identifier, glass identifier, and candidate coordinates. A production batch spanning the current three sets is not considered abnormal. Each training batch is sorted by category and severity from... and Stratified sampling with a 1:1 head-to-tail ratio was used. Duplicate glass markers and candidate coordinate records were retained based on the latest verified version, according to the confirmation time. Dataset access control must be met before creating candidate versions. and Each set must contain at least 30 valid samples, and each category involved in the update must contain at least 10 valid samples in both sets. It must contain at least 20 valid samples, and each category involved in the update must contain at least 5 valid samples. The system must contain at least 70 valid samples, with at least 5 valid samples in each of the seven categories. Two validation sets must each contain at least 20 samples with valid masks and severity labels. 1:1 stratified sampling during the training phase can be performed on each category participating in the update. Candidate versions cannot be created or trained if any set is empty, or if the number or stratification coverage is insufficient. The current version is retained, a record of insufficient samples and a review instruction are generated, and manually confirmed samples continue to accumulate. Once the conditions are met, the data is re-divided according to the same salt value. Independent calibration sets are still checked separately according to the subsequent 30-sample-per-class gate. The six sets respectively store a fixed salt value version, a partitioning rule version, a list of source batches, a joint key caliber, and an index hash. Before updating, a pairwise zero intersection check with specified granularity is performed. If any non-designed intersection is not empty, the update is terminated, and a data partitioning anomaly record and a review instruction are generated. Batch condition adapter updates use learning rate The rules include a maximum of 20 rounds and stopping if the validation loss does not decrease for 3 consecutive rounds. The class prototype is recalculated after each round using the current sample and the replay sample. After network training stops and candidate versions are frozen, only the independent calibration set is used. Calculate the dimensionless stratification risk calibration parameters corresponding to the current membrane system, 0.5 mm thickness layer, and category. No training or validation set samples may participate in this calculation; the area threshold is determined solely by... The calibration observations form a candidate grid with units of The length threshold is determined solely by The calibration observations form a candidate grid in mm, and the severity threshold is a dimensionless value ranging only from 0 to 1. The calibration observations form a candidate grid. The three types of grids are searched separately, and no addition, subtraction or size comparison is performed on the original values ​​of different dimensions. Each precise level and category requires at least 30 valid calibration samples. If there are not enough, only the nearest adjacent thickness layers within the same membrane system are allowed to be merged sequentially, and a maximum of two adjacent layers can be merged. The actual level, merging order and number of samples are recorded. If there are still less than 30 samples after merging, the candidate version is disqualified from deployment, the current version is retained, and a record of insufficient samples and a review instruction are generated. Sort by inconsistency score in ascending order and then select the sorting position. Corresponding values, where The number of valid calibration samples for this level and category, and must be a positive integer. Sort the indices of dimensionless positive integers and take the values ​​not less than 1. The smallest positive integer, when this integer is greater than season Each physical quantity threshold grid is combined with the severity threshold grid to form a judgment combination. Among the combinations with a missed rejection rate of no more than 1%, the one with the lowest false rejection rate is selected. In case of a tie, the physical threshold and severity threshold with the smallest numerical value are selected in sequence. The candidate version saves the random seeds of each stage. Index hash, source batch, partition and merge rule version, number of samples, sort position, threshold search record and stop round, enable the same data and configuration to reproduce update results; Before each candidate version enters validation, a parameter range check is performed. The channel scale output must be maintained between 0.5 and 1.5, the channel bias output must be maintained between -0.5 and 0.5, and each category prototype must be a finite 128-dimensional vector with quantiles. The severity threshold must be between 0 and 1; if the check fails in either range, online verification is not performed, an updated exception record is generated directly, and the current version is retained. Only when and Deploy candidate update versions in a timely manner, among which The 95th percentile of the acceptable increase in loss in historical playback verification results, and the upper limit in this specific implementation is limited to 0.02, is used for all... and All are dimensionless; if any verification condition is not met, the candidate version is deleted from the deployment qualification, the current version is retained, and an update exception record and review instructions containing two sets of verification losses, replay tolerances and failure conditions are generated. The update state machine includes states such as waiting to trigger, building candidates, range checking, dual dataset verification, pending deployment, and deployed. Only one candidate version is allowed to be in the pending deployment state at any given time. The model registry performs a concurrency check by comparing the current version number with the candidate's parent version number. If the parent version is inconsistent, the candidate is discarded and recalculated from the latest deployed version to prevent updates from two production batches from overwriting each other. The first 50 windows after the new version is deployed will serve as monitoring windows for continuous calculation. , And the verification rate. If the verification rate increases by more than 10 percentage points compared to the baseline of the most recent 200 glass blocks of the same film system before deployment, or if any drift exceeds the corresponding upper limit, the model registry will be atomically switched back to the previous deployed version and a rollback record will be generated. The rollback version will not be automatically redeployed, and subsequent updates will still need to pass the complete triggering and verification conditions again. The aforementioned second-priority over-limit exception branch in or When the controlled update module determines that the current batch exceeds the restricted adaptation range, it stops parameter updates and maintains the current version. Only the results of manual review are allowed to continue to accumulate. The samples exceeding the limit are not written into the training queue. At the same time, update anomaly records and review instructions are generated. When the subsequent re-established calibration window and manually confirmed samples cause the two drifts to fall back into the trigger range, the candidate version still needs to be recreated from the current deployed version without restoring the previous failed version. After successful deployment, the batch condition adaptation parameter version, defect category prototype version, and stratified risk calibration parameter version in the model registry are updated atomically. The previously deployed version is retained as the rollback target, and the new version number, effective batch, training sample summary, and other relevant information are added. , , , The threshold version is written to the update audit log; steps S4 and S5 read the complete update parameter group only when the next piece of tempered coated glass begins processing, thereby avoiding the use of different versions for the same piece of glass during processing.

[0027] The above description is only a preferred embodiment 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.

[0028] This invention sequentially connects multi-optical state acquisition, current glass reference area correction, batch condition feature adaptation, and optical relationship fusion, so that the responses of the same candidate in reflection, transmission, dark field angle, and polarization direction jointly participate in identification, thereby forming a distinguishable candidate-level characterization for defects with similar image appearance but different optical mechanisms.

[0029] This invention changes the adaptation and fusion structure after candidate features enter the recognition network, and limits model updates to batch conditional adaptation parameters, defect category prototypes, and hierarchical risk calibration parameters; through joint drift triggering, historical category prototype playback, and updated version verification, the batch adaptation results are used for subsequent recognition only when the constraints are met.

Claims

1. A method for intelligent identification of defects in tempered coated glass, characterized in that, include: S1. Collect reflected bright field images, transmitted bright field images, dark field images at at least two incident angles, and two orthogonal polarization state images at the same coordinates of tempered coated glass, and associate them with film system identifier, glass thickness, surface curvature parameters, and production batch identifier to form a group of original images with multiple optical states. S2. Based on the defect-free reference area of ​​the current tempered coated glass, the original image group of multiple optical states is corrected and coordinate registered to form the image tensor of multiple optical states and the response baseline of the reference area. S3. Input the multi-optical-state image tensor into the trained candidate segmentation network to obtain candidate region masks, extract the appearance features of the candidate region masks and the optical response trajectories arranged according to optical states, and form candidate-level feature groups. S4. Encode the reference area response baseline, film system identifier, glass thickness, surface curvature parameters and production batch identifier into batch condition vectors. Generate channel scale and channel offset from the batch condition vectors to correct candidate level feature groups. Set the correction features of each optical state as nodes, set the incident angle difference, reflection and transmission relationship and polarization direction difference as node relationships and aggregate node features to form candidate level mechanism fusion features. S5. Compare the candidate-level mechanism fusion features with the defect category prototype and perform defect prediction. Output the defect location, defect category, defect size, severity and confidence set. Generate a pass instruction, review instruction or rejection instruction based on the confidence set and severity. S6. Input the manually confirmed sample, the reference area drift amount of the reference area response baseline, and the defect category prototype drift amount into the controlled update module. When the joint drift triggering condition is met, update the batch condition adaptation parameters, defect category prototype, and layered risk calibration parameters to form update parameters for subsequent tempered coated glass to be used in steps S4 and S5.

2. The intelligent defect identification method for tempered coated glass according to claim 1, characterized in that, S1 includes: The glass conveying direction coordinates are determined by the pulse count of the glass conveying encoder, and the coordinates perpendicular to the conveying direction are determined by the camera calibration parameters; when the coordinates enter each imaging station, the reflected bright field light source, the transmitted bright field light source, the dark field light source with at least two different incident angles, and the orthogonal polarization imaging component are triggered in sequence; the encoder coordinates of each exposure are transformed to the same glass coordinate system; The system reads the film system identifier and production batch identifier from the production execution system, reads the glass thickness from the online thickness gauge, and reads the surface curvature parameters from the surface profile sensor. After performing a consistency check according to the glass identifier and the acquisition timestamp, it writes the acquisition record of the corresponding image. When any production parameter is missing or the consistency check fails, it writes a parameter invalid mark, terminates the automatic processing of the corresponding glass in steps S2 to S5, generates a review instruction, and prohibits the corresponding glass sample from entering the controlled update in step S6.

3. The intelligent defect identification method for tempered coated glass according to claim 1, characterized in that, S2 includes: The exposure correction coefficient is determined by the ratio of the target response to the measured response of the calibration plate under various optical conditions. Pixel gain and bias are determined by flat-field and dark-field images. Scale correction is determined by the color channel statistics or grayscale statistics of the defect-free reference area. The defect-free reference area is selected from a predetermined region that avoids glass edges and product markings. When the pixel gradient or outlier pixel ratio within the predetermined region exceeds the validity threshold determined by the quantile of the corresponding statistics of historical defect-free glass, a predetermined backup reference area is used. When the backup reference area still exceeds the validity threshold, the automatic processing of steps S3 to S5 for the corresponding glass is terminated. The system generates a verification instruction and prohibits the corresponding glass sample from entering the controlled update of step S6; it resamples and registers the corrected image according to the calibration transformation of each imaging station and the glass coordinates; it calculates the brightness statistics, color statistics, texture statistics and polarization response statistics of the effective reference area under each optical state to form the reference area response baseline; when the correction residual of any image channel exceeds the channel threshold determined by the historical calibration residual quantile, the image channel is marked as invalid, the automatic processing of the corresponding glass in steps S3 to S5 is terminated, a verification instruction is generated and the corresponding glass sample is prohibited from entering the controlled update of step S6.

4. The intelligent defect identification method for tempered coated glass according to claim 1, characterized in that, S3 includes: calculating edge continuity, aspect ratio, color shift, local texture, and area from pixels within the candidate region mask and its boundary neighborhood; representing particle scattering morphology with dark field grayscale attenuation sequence and bright spot dispersion in multiple directions from the candidate region center to the boundary, thus forming the appearance feature; using the corresponding optical state response of the defect-free reference area as a normalization benchmark, calculating the reflection response difference, transmission response difference, normalized response of each dark field incident angle, texture energy difference along the glass transport direction and perpendicular to the glass transport direction, and orthogonal polarization response difference of the candidate region; arranging the responses in the order of reflected bright field, transmitted bright field, dark field incident angle, and orthogonal polarization state to form the optical response trajectory.

5. The intelligent identification method for defects in tempered coated glass according to claim 1, characterized in that, The batch condition vector in step S4 is generated as follows: the reference response baseline is subtracted from the historical reference baseline stored in layers according to film system and glass thickness to obtain the reference response difference; the glass thickness and surface curvature parameters are numerically encoded, and the film system identifier and production batch identifier are categorically encoded. The reference response difference, numerical code, and category code are input into the batch condition adapter, which outputs the channel scale and channel offset corresponding one-to-one with each channel of the appearance feature and optical response trajectory. The channel scale is multiplied by the corresponding channel and then the channel offset is superimposed to obtain the corrected appearance feature and the corrected response feature of each optical state.

6. The intelligent defect identification method for tempered coated glass according to claim 5, characterized in that, The node relationships in step S4 include: representing the surface scattering and absorption response by the difference between the reflected bright field correction response characteristics and the transmitted bright field correction response characteristics; obtaining the corrected angular response slope by dividing the difference in correction response characteristics of adjacent dark field incident angles by the corresponding incident angle difference; and representing the polarization stress response by the difference in correction response characteristics of two orthogonal polarization states. The feature difference, corrected angular response slope, and polarization stress response corresponding to each node relationship are used as relationship features. Relationship weights are determined according to the relationship type. The correction response features corresponding to the same candidate region are weighted and aggregated. The aggregation result is then combined with the corrected appearance features and input into the interactive fusion layer to obtain the candidate-level mechanism fusion features.

7. The intelligent identification method for defects in tempered coated glass according to claim 6, characterized in that, S5 include: The distances between the candidate-level mechanism fusion features and the prototypes of film pinholes, scratches, discoloration, film peeling, particle adhesion, tempered stress patterns, and defect-free categories are calculated respectively. The distances are input into the classification prediction head to obtain the category scores, and the pixel mask and area or length severity are obtained by the segmentation prediction head and severity prediction head respectively. Inconsistency scores for each category are constructed using calibration samples layered according to film system and glass thickness, and confidence sets are determined according to the quantiles corresponding to the target coverage. When the confidence set contains only the defect-free category, a pass instruction is generated. When the confidence set contains only one defect category and the severity reaches the rejection threshold of that category, a rejection instruction is generated. In other cases, a review instruction is generated, and the calibration sample layer, quantile, and threshold used for the judgment are recorded.

8. The intelligent defect identification method for tempered coated glass according to claim 7, characterized in that, S6 includes: After a production batch switch, the initial response baseline of the current production batch is established by the effective reference area response statistics of the calibration glass that has been manually confirmed to be defect-free within the preset calibration window. The baseline version, calibration window, and sample batch identifier are recorded. The reference area drift of the reference area response baseline relative to the initial response baseline of the current production batch and the defect category prototype drift of the mechanistic fusion characteristics of the manually confirmed sample relative to the corresponding defect category prototype are calculated respectively. The lower limit of adaptation trigger and the upper limit of adaptation are determined by the two quantiles of the reference area drift in the historical verification batches, and the upper limit of playback constraint is determined by the quantiles of the defect category prototype drift. When the reference area drift is less than the adaptation trigger lower limit, the current parameters are retained. When the reference area drift is not less than the adaptation trigger lower limit and does not exceed the adaptation upper limit, and the defect category prototype drift does not exceed the playback constraint upper limit, the candidate segmentation network and the feature extraction parameters used to generate the appearance features and optical response trajectories are frozen. Only candidate update versions of the batch condition adapter, defect category prototype, and hierarchical risk calibration parameters are generated, and the playback constraints of the historical defect category prototype are added to the update loss. The weighted sum of classification error rate, segmentation loss, and severity error is used as the verification loss. The verification loss of the candidate update version and the current version is calculated using historical verification batch samples and current production batch manually confirmed samples, respectively. The candidate update version is deployed only when the current production batch verification loss is not greater than the current version and the increase in historical verification batch verification loss does not exceed the playback tolerance determined by the historical playback verification results. Otherwise, the current version is retained, and update anomaly records and review instructions are generated. When any drift exceeds the corresponding upper limit, parameter updates are stopped and update anomaly records and review instructions are generated.

9. The intelligent defect identification method for tempered coated glass according to claim 7, characterized in that, The training of the candidate segmentation network, batch condition adapter, relation weights, interactive fusion layer, classification prediction head, segmentation prediction head, and severity prediction head includes: dividing production samples labeled with glass coordinates, defect category, pixel mask, area or length, severity, as well as film system identifier, glass thickness, surface curvature parameters, and production batch identifier into training sample group and calibration sample group; using the training sample group to jointly optimize the candidate segmentation loss, category prototype distance loss, defect classification loss, pixel segmentation loss, and severity prediction loss, so that the joint optimization covers the candidate segmentation network, batch condition adapter, relation weights, interactive fusion layer, and each prediction head; obtaining batch category centers according to the mechanistic fusion features of the same category in each production batch, and performing a weighted average of the batch category centers according to the number of samples of the corresponding category in each batch to form a global defect category prototype, and storing each batch category center as a historical defect category prototype; using the calibration sample group to determine the inconsistency score quantile corresponding to the target coverage and the elimination threshold of each defect category.

10. A smart defect recognition system for tempered coated glass, used to execute the smart defect recognition method for tempered coated glass according to any one of claims 1 to 9, characterized in that, include: A multi-optical-state acquisition module is used to generate a group of original images in multiple optical states; The calibration and registration module is used to generate multi-optical-state image tensors and reference area response baselines; The candidate feature formation module is used to form candidate region masks, appearance features, and optical response trajectories; the conditional correction and mechanism fusion module is used to correct candidate-level feature groups and form candidate-level mechanism fusion features based on batch condition vectors. The defect prediction and instruction module is used to output defect location, defect category, defect size, severity and confidence set, and generate pass instructions, review instructions or rejection instructions; the controlled update module is used to update batch condition adaptation parameters, defect category prototype and stratified risk calibration parameters based on reference area drift and defect category prototype drift.

Images