Computer vision based holographic page defect detection method

Abstract

The application discloses a holographic page defect detection method based on computer vision and belongs to the field of holographic storage medium quality detection and computer vision image processing. In order to solve the problem that local scattering, scratches, pollution or microstructure damage of a holographic page are difficult to be stably found in an early stage and may cause data block loss or distortion of a recovered image, the application realizes the technical effects of automatic detection of holographic page defects and affected data block ranges and re-inspection prompt through multi-imaging image registration, page grid fusion features, partition patch memory library, page grid token map propagation and conformal calibration threshold determination.

Description

Technical Field

[0001] This invention relates to the fields of holographic storage medium quality inspection and computer vision image processing, and particularly to a method for detecting defects in holographic pages based on computer vision. Background Technology

[0002] Holographic storage achieves high-density information storage through page-based data recording and reading. However, the surface state, scattering characteristics, and readout reconstruction quality of holographic pages can change during exposure, packaging, transportation, and long-term service. Existing detection methods typically rely on manual sampling, observation of single surface images, or a posteriori judgment of readout results, making it difficult to simultaneously cover multiple types of information, such as intra-page coding units, data block boundaries, scattering response, and reconstruction residuals.

[0003] However, existing technologies have at least the following shortcomings: First, manual or single-image detection is difficult to cover high-density page-type data areas, and small scratches, contamination, and microstructural damage are easily missed in the early stages; second, uneven exposure, differences in background scattering, and differences in position within the page can easily lead to abnormal score shifts, resulting in false alarms or unstable thresholds; third, the impact of local defects on adjacent page grids, data blocks in the same group, and error correction code groups lacks structured propagation analysis, making it difficult to output the range of affected data blocks and re-inspection suggestions in a timely manner.

[0004] Therefore, a holographic page defect detection method is needed to address the shortcomings of the existing technology. Summary of the Invention

[0005] One objective of this invention is to propose a holographic page defect detection method based on computer vision. Addressing the problem in existing technologies that local scattering, scratches, contamination, or microstructural damage to holographic pages is difficult to detect reliably in its early stages, and may be amplified into data block loss or image distortion during subsequent reading, this invention proposes a method for acquiring surface reflection images, oblique incidence scattering images, and readout reconstructed page images, and performing page coordinate registration. Based on a self-supervised visual encoder, local grayscale uniformity, frequency domain scattering response, and readout reconstruction residuals, page grid fusion features are formed. This is further combined with a partitioned patch memory, page grid token map propagation, and conformal calibration threshold to generate a defect probability map and defect mask. This invention has the technical effect of automatically detecting holographic page defects and outputting the affected data block range and re-inspection suggestions.

[0006] This invention provides a computer vision-based method for detecting defects in holographic pages, comprising: S1, acquiring surface reflection, oblique incident scattering, and readout reconstruction images of the holographic page to be inspected, completing page coordinate registration based on the periodic features of synchronization markers or coding arrays, dividing the page into grids according to coding units or data blocks, and obtaining a set of page grid images and position information; S2, inputting each page grid into a trained self-supervised visual encoder, extracting multi-scale patch features, and calculating local gray-level uniformity values, frequency domain scattering response values, and readout reconstruction residual values ​​to form page grid fusion features; S3, based on the partitioned patch memory of normal holographic pages, calculating the distance between the page grid fusion features and the normal patch features of the corresponding region to obtain an initial anomaly score; S4, constructing coding units or data blocks... The process involves generating a page token that includes visual patch features, scattering response, frequency domain outliers, and readout reconstruction residuals. A coding unit graph is established based on spatial adjacency, error correction grouping, and the influence range of point diffusion. A data block risk vector is obtained through graph attention propagation. S5: Based on the normal outlier score distribution within the page partition and the preset target coverage quantile, a conformal calibration threshold is determined. The initial outlier score, data block risk vector, scattering change, and readout reconstruction residual are fused to generate a defect probability map and a defect mask. The page is then classified into confirmed defects, suspected re-inspection areas, and normal areas according to the conformal calibration threshold. S6: Based on the defect mask and location information, the defect coordinates, defect type, defect area, defect severity value, affected data block range, and re-inspection recommendations are output.

[0007] Optionally, S1 includes:

[0008] The surface reflection image, the oblique incidence scattering image, and the readout reconstructed page image are acquired respectively under the same page positioning reference.

[0009] Extract the marker center point from the synchronization marker, or extract the horizontal and vertical periodic vectors from the periodic features of the coding array;

[0010] The page coordinate transformation matrix is ​​obtained based on the center point of the mark or the horizontal periodic vector and the vertical periodic vector;

[0011] The three types of images are mapped to a unified page coordinate system using the page coordinate transformation matrix, and page grid index, page grid boundary and page grid position information are generated based on the coding unit spacing or data block boundary.

[0012] Optionally, S2 includes:

[0013] Each page area is cropped into a page patch that includes the page body and the preset neighboring border;

[0014] The page patch is input into the trained self-supervised visual encoder, and the patch representations output by different coding layers are extracted and pooled to obtain visual patch features;

[0015] The ratio of the standard deviation of gray level within a page grid area to the gray level mean plus a preset constant is used as the local gray level uniformity value.

[0016] A frequency domain transformation is performed on the grid region in the oblique incidence scattering image, and the proportion of the energy of the preset frequency band to the total energy of the grid is used as the frequency domain scattering response value.

[0017] The pixel difference statistics between the readout reconstructed page image and the normal reference reconstructed page image within the same page grid area are used as the readout reconstruction residual value;

[0018] The visual patch features, the local grayscale uniformity value, the frequency domain scattering response value, and the readout reconstruction residual value are normalized and spliced ​​together to obtain the page fusion feature;

[0019] Furthermore, the self-supervised visual encoder obtained through training is acquired using normal holographic pages and holographic pages containing labeled defective regions as training samples;

[0020] During training, occlusion recovery, adjacent page grid consistency constraints, and cross-imaging mode feature alignment are performed on the page grid patches to map the patch representations of the same page grid in the surface reflection image, oblique incidence scattering image, and readout reconstructed page image to the same feature space.

[0021] After training is complete, the encoder parameters are frozen, and the page fusion features of the normal holographic page are written into the partition patch memory.

[0022] Optionally, S3 includes:

[0023] The normal holographic page is divided into multiple partitions according to its position within the page, and the normal page grid fusion features are saved for each partition to form the partition patch memory library;

[0024] For the holographic page to be inspected, retrieve a preset number of neighboring normal page fusion features within the corresponding partition;

[0025] An initial anomaly score is generated based on the mean distance, nearest neighbor distance, or weighted distance between the feature distances between the page fusion feature to be inspected and the neighboring normal page fusion features;

[0026] The initial anomaly score is bound to the page index and then used to construct subsequent page tokens and determine defects.

[0027] Optionally, S4 includes:

[0028] Using the page index as the node identifier, the visual patch feature, the scattering response, the frequency domain outlier, the readout reconstruction residual, and the initial outlier score are concatenated into a page token vector;

[0029] Spatial adjacency edges are generated according to the contact relationship of the page grid boundaries, group association edges are generated according to the grouping of the same data block or the same error correction code, and optical influence edges are generated according to the page grid coverage determined by the point spread function radius.

[0030] Assign trainable attention weights to each connected edge according to the edge type, and update the lattice token vector through graph attention propagation;

[0031] The updated page token vector is mapped to the probability value of being affected by defects corresponding to each data block, and the data block risk vector is composed of the probability values ​​of being affected by defects.

[0032] Furthermore, the graph attention propagation uses a set of attention parameters corresponding to the edge type to calculate attention coefficients for spatially adjacent edges, grouped associated edges, and optically influential edges, respectively.

[0033] The attention coefficients are normalized within the range of incoming edges of the same node, and the normalized attention coefficients are used to perform weighted aggregation of adjacent page token vectors to obtain the updated page token vector.

[0034] When the probability values ​​of multiple page token vectors within the same error correction code group that are affected by defects reach the intra-group propagation threshold, the data blocks within the error correction code group that are not covered by the defect mask but have a group association edge with the defective page are added to the candidate list of affected data blocks. The candidate list of affected data blocks serves as the input for generating the range of affected data blocks and re-inspection recommendations.

[0035] Optionally, S5 includes:

[0036] Based on the normal holographic page calibration set, the inconsistency score is calculated in each page partition, consisting of the initial anomaly score, the page grid component corresponding to the data block risk vector, the scattering change, and the readout reconstruction residual value.

[0037] According to the preset target coverage, the quantiles are obtained from the empirical distribution of the non-consistency scores, and the quantiles are used as the conformal calibration thresholds corresponding to the partitions within the page.

[0038] The initial anomaly score, the page component corresponding to the data block risk vector, the scattering change, and the readout reconstruction residual value are normalized and fused according to a preset weight to obtain a defect probability map.

[0039] Pages in the defect probability map that reach the conformal calibration threshold are marked as identified defect areas, pages that do not reach the conformal calibration threshold but reach the re-inspection threshold are marked as suspected re-inspection areas, and the remaining pages are marked as normal areas. The defect mask is generated from the identified defect areas and the suspected re-inspection areas, wherein the re-inspection threshold is obtained by multiplying the conformal calibration threshold by a preset scaling factor.

[0040] Furthermore, the conformal calibration threshold is also partitioned and corrected according to the background exposure distribution within the page before being applied to the holographic page to be inspected;

[0041] The partition correction includes calculating the average background grayscale value and the scattering baseline value of each partition within a page, and inputting the average background grayscale value and the scattering baseline value into a threshold correction mapping table to obtain the threshold correction coefficient;

[0042] The corrected conformal calibration threshold is determined based on the conformal calibration threshold and the threshold correction coefficient.

[0043] The revised conformal calibration threshold is used to replace the conformal calibration threshold of the corresponding page partition in the defect probability map threshold determination and re-inspection threshold calculation.

[0044] Optionally, S6 includes:

[0045] The connected pages in the defect mask are merged, and the defect coordinates are calculated inversely based on the merged page boundaries and page coordinate transformation matrix.

[0046] The defect type is determined based on the linear brightness and darkness variation in the surface reflection image, the scattering variation in the oblique incidence scattering image, the reconstruction residual distribution in the readout reconstruction page image, and the page grid token vector. The defect type includes local scattering, scratches, contamination, and microstructure damage.

[0047] Convert the merged page area into the defect area;

[0048] The defect severity value is calculated based on a weighted combination of the defect probability map mean, defect area, readout reconstruction residual mean, and number of affected data blocks.

[0049] The data blocks that intersect with the defect mask and the data blocks whose probability of being affected by defects after graph attention propagation reach the data block risk threshold are identified as the affected data block range, and re-inspection suggestions are output for suspected re-inspection areas.

[0050] The beneficial effects of this invention are:

[0051] 1. By integrating surface reflection, oblique incidence scattering, and readout reconstruction images in the same page coordinate system, and dividing the page into grids according to coding units or data blocks, the defect detection object can correspond to the actual data area of ​​the holographic page, reducing the risk of missed detection caused by insufficient coverage of manual sampling inspection.

[0052] 2. Multi-scale patch features are extracted by a self-supervised visual encoder, and page fusion features are established by combining local gray-scale uniformity, frequency domain scattering response and readout reconstruction residual. Then, the initial anomaly score is calculated using the normal holographic page partition patch memory. This can identify abnormal states such as local scattering, scratches, contamination and microstructure damage from the perspective of normal sample distribution.

[0053] 3. By using page-based optical residual tokenization, coding unit graph attention propagation, and conformal calibration threshold determination, the risk vector of data blocks can be determined by combining spatial adjacency, error correction grouping, and the influence range of optical point diffusion. The data blocks are then classified according to identified defects, suspected re-inspection areas, and normal areas, thereby improving the stability of minor defect determination and reducing false alarms caused by uneven exposure. Attached Figure Description

[0054] 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:

[0055] Figure 1 This is a flowchart of a computer vision-based holographic page defect detection method.

[0056] Figure 2 This is a flowchart of step S4, page tokenization and graph attention propagation, of the present invention. Detailed Implementation

[0057] 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.

[0058] refer to Figures 1-2 A computer vision-based holographic page defect detection method includes: S1, acquiring surface reflection, oblique incident scattering, and readout reconstruction images of the holographic page to be inspected, completing page coordinate registration based on the periodic features of the synchronization marker or coding array, dividing the page into grids according to coding units or data blocks, and obtaining a set of page grid images and position information; S2, inputting each page grid into a trained self-supervised visual encoder, extracting multi-scale patch features, and calculating local gray-level uniformity values, frequency domain scattering response values, and readout reconstruction residual values ​​to form page grid fusion features; S3, based on the partitioned patch memory of normal holographic pages, calculating the distance between the page grid fusion features and the normal patch features of the corresponding region to obtain an initial anomaly score; S4, constructing coding units or data blocks into a set including... Page tokens containing visual patch features, scattering response, frequency domain outliers, and readout reconstruction residuals are used to establish a coding unit graph based on spatial adjacency, error correction grouping, and point diffusion influence range. A data block risk vector is obtained through graph attention propagation. S5: Based on the normal outlier score distribution within the page partition and the preset target coverage quantile, a conformal calibration threshold is determined. The initial outlier score, data block risk vector, scattering change, and readout reconstruction residual are fused to generate a defect probability map and a defect mask. Pages are then classified into confirmed defects, suspected re-inspection areas, and normal areas according to the conformal calibration threshold. S6: Based on the defect mask and location information, the defect coordinates, defect type, defect area, defect severity value, affected data block range, and re-inspection recommendations are output.

[0059] In this specific embodiment, S1 includes:

[0060] The detection electronic device initiates three-channel imaging acquisition using the same holographic page positioning reference, and the imaging control module records the surface reflection image as... The oblique incidence scattering image is denoted as Reading out the reconstructed page image is denoted as All three types of images are written to the page image cache and bound to the page number, exposure sequence number, acquisition time, imaging magnification, and positioning reference number. In this specific embodiment, the original image of a single page is resampled into a normalized grayscale array of 4096 by 4096 pixels, with a grayscale value range of [missing information]. The acquisition time difference of the three types of images on the same page is controlled within a mechanical positioning and holding cycle. Image groups that exceed this cycle are marked as waiting to be re-acquired and are not included in the page grid division.

[0061] The page registration module first... Detection of the set of synchronous marker centers ,exist and The set of synchronous marker centers with the same name was detected separately. and When the synchronization marker is obscured by a local defect, the module instead extracts the horizontal periodic vector from the periodic features of the encoding array. and vertical periodic vector ,in The period vector represents the imaging channel for reflection, scattering, or readout reconstruction. It is jointly determined by the one-dimensional autocorrelation peak and the two-dimensional phase correlation peak of the bright and dark fringes within the page. The peak confidence is written into the registration record. The higher the peak confidence, the more reliable the period feature. Below the period confidence threshold... Stop using periodic features to obtain the page coordinate transformation matrix and trigger same-page resampling;

[0062] For channels with at least four synchronization markers of the same name The page registration module follows Find the page coordinate transformation matrix ,in For channel The first in The homogeneous coordinates of the center point of each marker To unify the corresponding coordinates in the page coordinate system. For pixel distance, when there is a lack of valid synchronization markers, follow the principle of pixel distance. and The affine transformation matrix is ​​fitted and solved jointly with page boundary corner constraints. The solution is then presented as follows: Calculate the registration residuals;

[0063] In this specific implementation, the registration residual threshold The registration parameter table is determined by the 95th percentile of the registration residual of the normal holographic page positioning sample, and the unit is pixels. Time will pass Image mapped to a unified page coordinate system ,when The page number, failed channel, and residual value of the image group are retained and the re-sampling flag is output. Pages that fail to register are not included in the subsequent defect probability calculation.

[0064] The page generation module reads the encoded array spacing from the initialization parameters of the detection system. and Or a data block boundary table, in which To unify the center spacing of adjacent coding units in the horizontal direction under the page coordinate system To unify the center spacing of adjacent coding units in the vertical direction under the page coordinate system, both are obtained by combining the physical spacing of the coding array in the detection system initialization parameters with the imaging magnification conversion. The initialization parameters are generated by the holographic page writing controller during writing and written to the media metadata or archive index file along with the page number. Fields include the physical spacing of the coding array, data block boundaries, error correction code group identifier, writing batch, and coding version. At the start of detection, the electronic device loads the data according to the page number and checks it against the imaging magnification conversion coefficient in the factory calibration file. After passing the check, the data is processed accordingly. Generate page boundaries ,in For page grid indexing, if a data block boundary table is used, adjacent coding units are merged into a set of page grids corresponding to the same data block and the data block identifier is recorded. The page grid position record includes the page grid index, unified page coordinate polygon, center point coordinates, data block identifier, error correction code group identifier, and three types of image validity flags.

[0065] Before acquisition, the imaging control module reads the imaging parameter table. The table fields include reflected light exposure time, oblique incidence angle, scattered illumination intensity, readout reconstruction exposure time, camera gain, dark field correction version, and flat field correction version. In this specific embodiment, the reflected light and readout reconstruction channels use the same pixel coordinate origin. The oblique incidence scattering channel saves the angle compensation amount after the support calibration. If the saturation pixel ratio of any channel exceeds... Or the noise in the dark field exceeds the noise threshold. If so, the image in that channel is marked as invalid and a re-sampling on the same page is triggered;

[0066] Preprocessing module , and Dark field subtraction, flat field normalization, and bad pixel interpolation are performed separately. The flat field normalization rule is as follows: ,in Original grayscale This is a dark field image. This is a flat field diagram. To normalize the grayscale and prevent zeroing, the corrected amount As input for synchronous marker detection and periodic feature extraction;

[0067] When both the synchronization marker center point and the periodic feature are available, the page registration module prioritizes using the synchronization marker for solution. The affine matrix obtained by utilizing periodic characteristics. Perform a consistency check, if Above the matrix consistency threshold If the registration is low, the corresponding image group will be marked as low-confidence and added to the resampling queue. If the two are consistent, the center point, period vector, matrix consistency value, and residual will be marked. Write it together into the page coordinate registration record;

[0068] After the page grid is divided, the quality control module will determine the appropriate page layout based on each page layout. Internal effective pixel ratio, three-way imaging effective flags, and registration residual generation page grid effectiveness flags ,when When the page position is retained but the defect detection patch is not clipped, Add the page to the set of valid pages. The number of valid page sets, the number of missing page sets, and the reasons for missing tests are written into the detection task cache for use by the modal availability mask of S2 and the output confidence flag of S6.

[0069] The page image cropping module is based on from , and Cropping the corresponding area and generating a page grid image record for each page grid. ,in , and These are preprocessed images of the surface reflection image, oblique incidence scattering image, and readout reconstructed page image, respectively, after dark field subtraction, flat field normalization, bad pixel interpolation, and mapping to a unified page coordinate system using the corresponding page coordinate transformation matrix. The apostrophe indicates that the corresponding image has completed preprocessing and page coordinate registration. , and There are three types of imaging patches, and the page coordinate transformation matrix recorded is used for subsequent defect coordinate inverse calculation. A set of page grid images is formed, and the page grid position record table is written together with the page grid image set into the detection task cache and used as input to S2.

[0070] In this specific embodiment, S2 includes:

[0071] The feature extraction module reads the set of page images written by S1 and performs a feature extraction on each page. Expanding outwards from the main boundary of the page grid In this specific embodiment, the spacing between encoding units is trimmed with page grid patches. Setting the value to 0.25 ensures the patch includes both the main page structure and neighboring sidebands. The patch is scaled to 128x128 pixels while maintaining the three-way imaging order. For page patches lacking a particular imaging path, a zero matrix is ​​written to the corresponding channel and a modal availability mask is applied. Centered at zero;

[0072] Self-supervised visual encoder Employing a multi-scale convolutional visual coding network with a three-modal shared backbone, and surface reflection patches Oblique incidence scattering patch and read refactoring patch First, each patch representation is fed into the same weight-shared encoding backbone to obtain its own patch representation. Then, cross-imaging mode alignment and fusion are performed at the feature layer. The holographic page containing labeled defective regions is only used to remove defective region samples, construct a validation set, and check the encoder's ability to separate anomalous textures; it is not used as a separate feature. The supervised classification label of the main training loss is still occlusion recovery, adjacent page grid consistency and cross-imaging mode feature alignment. The training objective variables are the occluded patch pixels, the representation distance of adjacent page grids and the cross-modal representation distance of the same page grid. After the loss converges, the encoded backbone parameters are frozen and the model version is output.

[0073] The training management module represents the self-supervised training process as a training state vector. ,in The current parameters of the encoder. To cover up and restore the damage, For cross-modal alignment loss, For the consistency loss of adjacent page cells, To verify the stability of the normal page distance on the set; the decision variables for training are: The objective function is ,in Cross-modal alignment loss The weight, Consistency loss between adjacent pages The weights, both of which are non-negative training hyperparameters, are determined by the validation set and bound to the encoder model version, with the constraint that the reconstruction error on the validation set does not exceed [a certain value]. And the normal page spacing stability is not lower than If the constraints are not met, the system will revert to the previous model version and continue collecting normal samples, without releasing the new encoder version.

[0074] During the reasoning stage, Output the first image patch for each path. Patch representation of each coding layer , and The feature extraction module performs average pooling and max pooling on each modality representation, and then applies the modality availability mask. Multi-scale visual patch features are obtained by concatenating feature layers. After pooling The training set mean and standard deviation are normalized, and the normalization parameters are derived from normal lattice samples during model training and stored together with the model version number.

[0075] Local grayscale calculation module on surface reflection patch Read the normalized grayscale set within the main area of ​​the page grid. ,according to Calculate the local grayscale uniformity value ,in The grayscale mean is... The standard deviation of grayscale The zero-prevention amount is a normalized grayscale value of 0.01 in the same unit. The value is dimensionless, and the larger the value, the more uneven the grayscale of the page grid. If the proportion of effective pixels within the main body of the page grid is lower than... Then Write as missing test and in The grayscale item in the middle record is unavailable;

[0076] Frequency domain scattering calculation module for oblique incidence scattering patch Performing a two-dimensional Fourier transform on the main region yields The defect-sensitive frequency band, calibrated by the normal holographic page scattering sample, is denoted as... ,according to Calculate the frequency domain scattering response value ,in It is a full-band collection. To prevent zero quantity of the same unit energy, Cut to And write the page grid feature record, frequency band The image magnification and illumination angle are read as fields in the frequency domain parameter table;

[0077] Read the residual calculation module to read the reconstructed patch. Reference patch in the same page grid area as normal reference reconstruction pages ,according to Calculate the readout reconstruction residual value ,in This represents the number of effective pixels per page. and All Range of grayscale values, For dimensionless average residuals, refer to the patch. It is obtained by transforming the normal read samples or the target page image at the write end of the same encoding array into the same coordinate system.

[0078] The reference reconstruction page library is jointly established by the target page image at the write end and the normally reconstructed page confirmed by stable readout. Reference record fields include page number, encoding version, write batch, imaging magnification, exposure time, camera gain, flat correction version, and page coordinate transformation version. Before calculating the readout residual, the module prioritizes selecting a reference patch with the same page number, encoding version, and write batch, and performs the same channel normalization on exposure and gain differences. If the reference patch is unavailable or the registration residual between the reference page and the current page exceeds [a certain threshold], [the module will proceed with the next step]. Then Marked as missing test and by Set this contribution to zero, and do not replace it with the residual value from other pages;

[0079] After the page patch is trimmed, the patch standardization module follows... Perform channel normalization for each patch, where For the first Page number in the first... Normalization patch on the road imaging channel, This is the unnormalized grayscale matrix of the imaging patch for this path. Indicates the reflection, scattering, or readout reconstruction imaging channel. Indicates page grid index, and The first The mean and standard deviation of grayscale values ​​for the imaging channel in the normal holographic page calibration set. and From the normal holographic page calibration set, To prevent zero grayscale values ​​within the same unit, the normalized patch is then trimmed to... And linearly mapped to To avoid exposure drift directly amplifying abnormal scores;

[0080] train At that time, the training management module records the page grid samples as ,in This only indicates the defect area cleaning flag and the validation set layering flag. To identify page grid positions and imaging batches, random block masking is used for occlusion recovery. Cross-imaging alignment uses three separately encoded representations of the same page grid as positive samples and normal texture representations of different page grids as negative samples. During the training phase, optimization is not performed directly based on defect type labels. The model version is only released after the classification output is confirmed to be stable on the validation set after training.

[0081] The encoder output also includes page grid characterization confidence. This confidence level is obtained by weighting the occlusion recovery error, cross-imaging alignment similarity, and the proportion of effective pixels in the patch. If the feature extraction module marks the page as a low-confidence page, it retains the computable pages. , and And reduce visual patch features during fusion. The weight, Determined by the 10th percentile of the low-quality samples collected in the validation set;

[0082] After pooling, the multi-scale patch features are compressed using the trained linear projection. Will Mapped to fixed-dimensional visual patch features In this specific implementation, the dimension is 128, and the projection matrix is... and Model version binding: If the model version is inconsistent with the S3 partition patch memory version, the same version of the feature will be recalculated instead of using different versions of the feature.

[0083] Frequency domain scattering response External scattering variation During calculation, the partition is read first. Normal scattering baseline and quantile scale Press again get, The larger the value, the more significant the scattering enhancement relative to the normal baseline. This value is subsequently incorporated into both the S4 token construction and the S5 defect probability fusion.

[0084] Reading the reconstructed residuals also calculates a summary of the spatial distribution of the residuals and a residual plot. Divided into a central region and a side region, the module records the mean of the central residual, the mean of the side residual, and the proportion of the area of ​​the maximum connected residual, respectively, to distinguish between the main defects of the page grid and the boundary registration error. If the side residual is significantly higher than the central residual and the registration residual is close to the threshold, a boundary instability flag is written into the page grid fusion feature.

[0085] Before fusion, the parameter management module reads the feature normalization table. The table fields include feature name, mean of normal samples, standard deviation of normal samples, upper and lower clipping boundaries, applicable model version, and update time. All normalization outputs are written to... Scope, missing test items passed It participates in the stitching but does not participate in the corresponding item mean statistics, ensuring that the distance calculation of S3, the graph propagation of S4 and the fusion score of S5 use the same scale;

[0086] The fusion module will , , and Normalized to the quantile range of the training set or calibration set respectively. , , and , and according to Forming page grid fusion features The missing items were provided by Controlling for zero contribution, normalized upper and lower bounds, modal weights, and model version number are written into the page fusion feature table, and the normal holographic page... At the same time, the partition patch memory used by S3 is written according to the page partitioning.

[0087] In this specific embodiment, S3 includes:

[0088] The memory bank construction module divides normal holographic pages, which have been manually verified or confirmed by historical stable readouts, into categories based on their position within the page. In this specific implementation, a 4x4 partition is used, and for each partition... Establish a partition patch memory The memory records include page grid fusion features, page grid center coordinates, page number, imaging batch, coded material batch, model version number, and quality indicator. Only page grid fusion features from the three-way imaging in S2 that are valid and do not fall into the labeled defect area are written into the memory. ;

[0089] To avoid abnormal samples contaminating the normal memory, the construction module first calculates the local density of candidate normal features within the same partition, and then... Obtain candidate features To the nearest feature set median distance, delete Candidate records above the 97th percentile of the partition's history are selected, and a nearest neighbor index is created for the retained records. The index key includes the partition number, model version number, and coded material batch.

[0090] When the holographic page to be inspected enters S3, the anomaly scoring module reads the page grid fusion features. And page grid position records, according to the partition where the page grid center point is located. Query The number of searches in this specific embodiment Take 20, when the corresponding partition has fewer than 20 valid normal records. When a record is entered, the batch record of the same material is read from the adjacent partition to complete it, and the memory confidence flag of that page is written as pending review;

[0091] For the retrieved first Normal page fusion feature The anomaly scoring module first applies segment weights to the whitened visual feature segment, local grayscale uniformity segment, frequency domain scattering response segment, and readout reconstruction residual segment according to the feature metric configuration table. These segment weights are calibrated by the normal page validation set and sum to 1. Then, according to... Calculate normalized weighted Feature distance ensures that visual features and scalar features reside in the same dimensionless metric space, where For the first The fusion features of the page to be inspected and the retrieved page number Normalized weighted average between normal page fusion features Feature distance, , , and These are the distance weights for the visual patch feature segment, the local gray-level uniformity segment, the frequency domain scattering response segment, and the readout reconstruction residual segment, respectively. All are non-negative and their sum is [value missing]. , , , and The pages to be checked are respectively The normalized results of the corresponding visual patch features, local gray-level uniformity values, frequency domain scattering response values, and readout reconstruction residual values ​​are as follows: , , and The first The normalized results of four types of features corresponding to each neighboring normal page cell;

[0092] The initial anomaly score is determined by both the weighted distance and the nearest neighbor distance, calculated according to the following rules: ,in For the first The initial anomaly score for each page to be inspected. This is the fusion coefficient between the weighted average distance and the nearest neighbor distance, and in this specific implementation, it is taken as 0.6. For the first Normalized weights corresponding to the distances to neighboring normal page cells The number of adjacent normal pages retrieved. The larger the value, the more the page grid deviates from the normal patch distribution. The normalized initial outlier score is calculated according to... Write, For the first The partition where each page is located normal distance percentile, To cut to The normalized initial outlier score within the range, making Maintains the same numerical domain as other fused inputs of S5;

[0093] Partition Patch Memory It is maintained using a versioned append-only update method. Each time a new normal holographic page sample is added, its imaging parameter version is first verified. Model version and page partitioning parameters can only be appended if all three are consistent with the current memory version. After appending, the distance scale within the partition must be recalculated. The 95th percentile distance and the number of valid samples are updated and written to the version log of the memory.

[0094] During the search, if the material batch or imaging magnification of the page to be searched is different from the specified batch, the search results will be affected. The main versions are inconsistent. The anomaly scoring module uses a candidate order that prioritizes candidates from the same batch, followed by adjacent batches, and multiplies the nearest neighbor distance across batches by a batch correction factor. , The feature drift calibration is obtained from the same normal page under different batch imaging conditions. If the correction coefficient cannot be obtained, the low confidence flag of the memory bank is output and the page is sent to S5 suspected re-inspection judgment.

[0095] The rating module will With page grid index Partition number The initial anomaly score table is generated by binding the nearest neighbor record identifier, the memory confidence flag, and the model version number. If the visual patch feature is unavailable, then the page number... The initial anomaly score is replaced by the median initial anomaly score of the valid pages within the same data block, and the replacement source is recorded. The initial anomaly score table also stores the nearest neighbor distance, weighted distance, partition normalized baseline, retrieval timestamp, anomaly source description, page status description, number of valid modalities, number of samples in the memory bank, reason for low confidence, page verification flag, and output sequence number. It is then sent to the page token construction module of S4 and the conformal calibration module of S5.

[0096] In this specific embodiment, S4 includes:

[0097] The graph construction module reads the page grid fusion feature table generated by S2 and the initial anomaly score table generated by S3, using the page grid index. Construct a page token as a node identifier. ,in For frequency domain outliers, according to From frequency domain scattering response value Subtract partitions Normal scattering baseline After normalization, and From the normal holographic page calibration set;

[0098] Coding unit diagram The set of nodes in For all valid page cells, the edge set Generated by three types of rules: spatial adjacency edges are generated if two page grid boundaries share an edge or corner; group-related edges are generated if two page grids belong to the same data block or the same error correction code group; and group-related edges are generated if the center distance between two page grids is [not specified]. Not exceeding the radius of the point spread function The corresponding page grid coverage area then generates an optical influence edge. Read from the point spread function calibration record of the same optical system;

[0099] Each edge Write edge type Center distance Data block relationship flag, error correction code block flag, and optical influence weight ,in for Dimensionless values ​​of range, spatial adjacency edges Set to 1, group the associated edges and attach the data block identifier of the same group. The generation threshold of the three types of edges and the point spread function radius are stored as fields in the graph construction parameter table as the imaging magnification.

[0100] The graph attention propagation module uses a set of attention parameters corresponding to the edge type. , for the First, calculate the attention. Then at the same target node Within the inbound range Normalization yields the attention coefficients, where For edge type The corresponding set of attention parameters, For attention head index, For the first Each attention node For nodes Unnormalized attention score To normalize the attention coefficient, edge type In the Trainable attention vectors on each attention head For the first The linear projection matrix of each attention head. and Target nodes and adjacent nodes Page token vector, For the edge Corresponding edge type Edge type embedding, For the edge Optical influence weight, For nodes The set of adjacent nodes, Larger values ​​indicate a larger node For nodes The greater the contribution of the defect propagation;

[0101] When a node is updated, the propagation module follows the procedure. Obtain the updated page token vector ,in For gated activation functions, and The training sample consists of a holographic page labeled with a defect mask and affected data blocks, forming the projection matrix obtained during training. If a node has no incoming edges, a self-loop edge is added and... Retain the node's own page token;

[0102] The data block mapping module will belong to the same data block After updating, the page token set is averaged and entered into the risk mapping header, then... Compute data blocks Probability of being affected by defects ,in For data blocks Inside The average vector, and For training parameters, The range is Furthermore, the larger the value, the higher the risk of the data block being affected by defects;

[0103] When the intra-group propagation threshold is reached within the same error correction code group of The quantity is no less than Or the contribution value of propagation at the page level within the group. Reaching the propagation contribution threshold At that time, the graph propagation module writes the corresponding data block within the error correction code group into the error correction group propagation candidate data block list. The candidate record fields include data block identifier, trigger risk value, trigger page index, propagation contribution value, group association edge identifier, and error correction code group identifier. This candidate list is generated only based on the risk vector, propagation contribution, and group association edge, and does not read the subsequent judgment results of S5. , and In this specific implementation, the training set with recovery failure records is used for calibration. Take 0.65 and Take 2;

[0104] When constructing the page token, the order of the token fields and the normalization parameters are specified by the token configuration table. The table fields include the start and end dimensions of visual features, the scattering response dimension, the frequency domain outlier dimension, the readout reconstruction residual dimension, the initial outlier score dimension, the missing test mask dimension, and the configuration version number. The graph propagation module only accepts the version configured during training. When there is a discrepancy, the token is regenerated instead of being reasoned directly.

[0105] After spatial adjacency edges are generated, the graph construction module merges duplicate edges of the same grid pair. If two nodes simultaneously satisfy both spatial adjacency and optical influence conditions, both edge types are retained and written into the edge record respectively. If the center distance... If the distance between two pages is less than one page grid and the two pages belong to different data blocks, a cross-block flag is added to the spatial adjacent edge. This flag is used to determine whether the defect has spread across data blocks.

[0106] The generation of grouped edges involves reading the error correction group table. Table fields include data block identifier, error correction code group identifier, list of pages within the group, redundancy level, and the number of decoding failures in the past. When the number of historical failures within the same error correction code group exceeds the group risk quantile threshold, the graph construction module adds a risk prior to the grouped edge. This prior is used as an input to the attention calculation for edge features instead of directly overriding the detection results;

[0107] Point spread function radius of the optical influence edge Calculated from the point source exposure calibration page, the calibration module first measures the point diffusion intensity in the calibration map as it drops to its peak value. pixel radius at Then, based on the physical pixel size corresponding to the current imaging magnification. Converted to physical radius , and according to Determine the page grid coverage area and center distance. No more than The node pairs with each page grid spacing generate optical influence edges; calibration tables are established for different imaging magnifications and oblique incidence angles. When a table entry is not hit, linear interpolation is performed according to adjacent magnifications and angles, and the graph record is marked as interpolation calibration.

[0108] During graph attention parameter training, historically labeled defective regions and data block restoration failure records are used together as training validation information. The loss function includes page influence discrimination loss, data block influence probability loss, and edge type consistency regularization term. After training... Stored separately by spatial adjacency edges, group association edges, and optical influence edges, so that the same abnormal page cell has different propagation intensities to adjacent page cells, data blocks in the same group, and optical diffusion range;

[0109] After propagation is complete, the risk mapping header also outputs a page-level propagation contribution value. ,in The average of the multi-head attention coefficients. Used to identify pages that are not covered by local anomaly scores but are affected by the propagation of neighboring defects, this field is written to the updated page token table and participates in the candidate interpretation of affected data blocks in S6;

[0110] Data block risk vector The updated page token table, coding unit graph, and list of candidate data blocks for error correction group propagation are written to the graph propagation result cache. As input to the defect probability graph fusion in S5, the candidate list also retains the trigger rule description, risk source page and propagation edge type. After S5 generates the defect mask, S6 filters it into the range of affected data blocks according to the mask coverage relationship.

[0111] In this specific embodiment, S5 includes:

[0112] The conformal calibration module reads the normal holographic page calibration set and groups the calibration page grids according to the page partitions formed by S1. Each calibration page grid and the page grid to be inspected has an initial anomaly score. Data block risk vector corresponding page cell component scattering change and read out the reconstructed residual value All four inputs were cropped before entering S5. Scope, missing test items are from Set the value to zero and record the source of the missing measurement;

[0113] The same fusion rule is used to calculate the defect probability map values ​​for both the calibration set and the pages to be inspected. The calculation formula is as follows: ,in , , and The weights are non-negative and sum to 1. All four input terms are dimensionless normalized values. and In the same The numerical space indicates that the larger the value, the stronger the inconsistency of the defects;

[0114] For intra-page partitions The module reads the set of defect probabilities for that partition from the normal calibration set, obtained according to the same fusion rule. According to the target coverage rate Calculate the conformal calibration threshold In this specific implementation method Take 0.95, Represents the empirical quantile function, if partitioned The number of calibration samples is lower than the preset minimum number of samples. Then, the normal calibration probabilities of adjacent partitions are merged and the confidence flag of the threshold is written as pending review;

[0115] In Before the holographic page to be inspected, the threshold correction module calculates the partition. average grayscale value of the background and scattering baseline value ,Will and The threshold correction mapping table is used as the query key to obtain the correction coefficient. The threshold correction mapping table is fitted by normal holographic page calibration samples under different exposure backgrounds and scattering baseline levels. The table fields include background grayscale range, scattering baseline range, correction coefficient, number of valid samples and version number. When no table entry is hit, linear interpolation of adjacent intervals is used.

[0116] The corrected conformal calibration threshold is based on Confirmed, among which and All The dimensionless threshold boundary within the range is derived from a mixed validation set of normal pages and labeled defective pages. Replace the corresponding page partition Participation in the same defect probability map The threshold determination is used simultaneously for the calculation of the re-inspection threshold;

[0117] The conformal calibration module also generates a calibration record for each page partition. The record fields include the number of calibration pages, target coverage, empirical quantile, threshold before correction, threshold after correction, background grayscale range, scattering baseline range, source of correction coefficient, and validity period. The calibration record uses the same weight table version as the defect probability map and is saved with the detection task number.

[0118] The defect probability map generation module no longer performs a second set of fusion or additional logistic regression on inconsistent scores, but instead... As a defect probability map in the page The pixel values ​​at each location are written to the result cache. The cache field simultaneously stores four normalized input items, their corresponding weights, the blended values ​​before cropping, and the values ​​after cropping. The system includes partition number, threshold version, missing test mask, weight table version, number of calibration samples, threshold correction coefficient, retest ratio coefficient, judgment timestamp, source of calibration partition samples, adjacent partition merging flag, pruning reason and output page number, thereby ensuring that the calibration object of the conformal calibration threshold and the subsequent comparison object are completely consistent.

[0119] The classification module will The page grid is marked to identify the defect area. Pages marked as suspected re-examination areas are designated as normal areas, while the remaining pages are marked as normal areas. The re-examination ratio coefficient is... The recall rate is calibrated using a validation set for re-inspection, which is set to 0.8 in this specific implementation. After merging the defective and suspected re-inspection regions based on connectivity, a defect mask is generated. The mask recording simultaneously retains the category labels of each page. , , and confidence markers;

[0120] Before use, the normal holographic page calibration set undergoes defect elimination processing. The calibration management module removes pages that failed historical re-inspection, manually marked defective pages, and abnormal pages with readout reconstruction residuals, retaining only the normal pages that are read out stably. The calibration record includes page number, partition number, background grayscale mean, scattering baseline value, initial anomaly score, data block risk component, readout reconstruction residual, and model version number.

[0121] Inconsistent weights , , and The goal, determined through a validation set grid search, is to achieve a defect recall rate of at least [percentage missing]. To minimize the false positive rate of normal pages under the condition of [condition], the parameter table stores the weights, validation set number, recall rate, false positive rate and effective time. If the model version of the detection task is different from the weight table version, the same version of the weight table is used or the output is stopped to determine the defect label.

[0122] The threshold correction mapping table is constructed using two-dimensional binning based on the background gray-level mean range and the scattering baseline range, with each entry storing the correction coefficient. The sample size, background range, scattering range, and validity period are specified in this specific implementation method. Restricted to This is used to compensate for the effect of the difference between the in-page exposure background and the scattering baseline on the conformal calibration threshold, so as to avoid misjudging a uniform background that is too bright as a defect.

[0123] After the defect probability map is generated, the post-processing module performs a neighborhood consistency check on isolated single-page cells to determine defects. If the defect in that page cell is... Just over If all eight neighboring pages are below the re-inspection threshold, the page is marked as isolated and awaiting re-inspection and is not included in the determined defect mask. If two or more adjacent pages exceed the re-inspection threshold at the same time, they are retained according to the original label, and the neighbor consistency verification result is written into the mask attribute table.

[0124] When generating masks for identified defective and suspected re-inspection areas, the module retains two binary images. and ,in Only includes pages with identified defects. Includes suspected re-inspection pages, final defect mask Both sub-masks have page grid indexes and partition thresholds so that S6 can distinguish between confirmed defects and suspected sources of re-inspection when outputting coordinates and re-inspection suggestions;

[0125] If the confidence flag for a certain partition is a calibration flag awaiting verification, the classification module will still calculate... and However, instead of outputting a definitive defect conclusion that cannot be verified, it will output more than... The page is marked as a high-priority re-examination page, and the reason for insufficient calibration samples, the adjacent partitions used, and the threshold version are written into the judgment result cache so that subsequent re-examination suggestions can indicate the need to supplement calibration samples.

[0126] Defect probability map Defect mask The partition correction threshold table, the list of suspected re-examination pages, and the confidence flags are written to the judgment result cache. The cache records also include the inconsistency score of each page. Corrected threshold The following information is included: re-inspection threshold, isolated re-inspection flag, threshold correction version number, classification generation time and judgment basis summary, defect mask and location information, along with the determined defect source, suspected re-inspection source, calibration confidence source, neighborhood consistency verification result, threshold source description, mask merging batch, re-inspection priority and page status summary, which serve as inputs for S6 output defect coordinates, defect area, defect severity value, affected data block range and re-inspection recommendations.

[0127] In this specific embodiment, S6 includes:

[0128] The output generation module reads the defect mask written by S5. And the page position information written by S1, for The middle label represents the page cell with identified defects or suspected re-inspection. Eight-neighbor connectivity merging is performed to obtain the set of defective connected domains. Each connected component Save the set of page indexes contained therein Composition of region labels, mean defect probability, maximum defect probability, and intersecting data block identifiers;

[0129] For connected components Module based All page boundaries Find the circumscribed polygon in the unified page coordinate system, and retrieve the channel-saved data from the S1 page grid image record. , and Calculate or read its inverse matrix respectively. , and The coordinates of the three original imaging paths were then calculated back, and the calculation rule was as follows: ,in For reflection, scattering, or readout reconstruction channels, To unify page coordinates, For channel The original image coordinates are used as the primary coordinate system, and the output records are mapped to the three original image coordinates.

[0130] The defect type determination module uses connected components The three types of covered grid image patches and the updated grid tokens are used as measurement objects. The linear brightness and darkness changes in the surface reflection image are statistically analyzed for these measurement objects. Scattering variation in oblique incidence scattering images Read the mean value of the reconstruction residuals in the reconstructed page image. and the average value of the updated page tokens And according to each type Score The defect types are determined, and the type set includes local scattering, scratches, contamination, and microstructural damage. The weights are obtained by training on labeled defect pages. For connected components Corresponding to the Type score of class defect, , , and Connected components The linear brightness variation, scattering variation, mean of reconstructed residuals, and mean of updated page tokens. , and The first The class defects correspond to the weights of the three scalar features mentioned above. For the first Class defect corresponding to the average value of the updated page token The weight vector, For the first Bias terms for class defects;

[0131] when After obtaining the highest type score, if the difference between the highest and second-highest scores is lower than the type confidence threshold... If the defect type is positive, write it as a suspected composite defect and retain the labels of the first two types; otherwise, output the error message. The corresponding single defect type, By defining the obfuscated boundaries of the defect type validation set, the type result and the connected components Coordinate record binding;

[0132] The defect area is calculated from the physical dimensions of the page, and the calculation rule is as follows: ,in and The physical width and height of the page grid in the unified page coordinate system are expressed in square millimeters. If the boundary of the connected domain only covers a part of a page grid, the area of ​​the page grid is weighted according to the proportion of pixels covered by the mask. The area conversion factor comes from the imaging magnification calibration record.

[0133] Defect severity module by Calculate the defect severity value ,in The mean probability of defects within the connected domain. This represents the defect area after normalization based on page area. To read the mean of the reconstructed residuals within the connected domain, The number of affected data blocks is a value normalized to the total number of data blocks within the page. , , and The sum is 1, and in this specific embodiment, the values ​​are taken as 0.35, 0.25, 0.25, and 0.15 respectively. The larger the value, the more severe the defect.

[0134] The affected data block range was obtained by merging two parts, the first part being the area related to the defect mask. The second part is the list of candidate data blocks for S4 error correction packet propagation, which consists of spatially intersecting data blocks. And not with Spatially intersecting data blocks; for the second part, S6 was generated in S5. Based on this, check whether the candidate data block is consistent with The affected data block range is only added if the overriding data block belongs to the same error correction code group, or if it is connected to a identified defective page or a suspected re-inspection page via a group association edge. The data block risk threshold is obtained by calibration from the recovery failure record. In this specific implementation, it is set to 0.6. Candidate data blocks that do not meet the mask filtering conditions are only retained as propagation observation records.

[0135] Before outputting defect severity and data block range, the measurement summary module performs a step for each connected component. Generate defect measurement records, with record fields including the set of measurement object pages. The system includes three-way image patch references, page coordinate polygons, physical area, defect probability statistics, readout reconstruction residual statistics, type confidence, and a list of intersecting data blocks. Any field that comes from alternative calculations or low-confidence page cells will have a source flag written for the re-inspection personnel to trace.

[0136] The re-inspection suggestion module is based on the connected component labels, The system generates suggested fields for the suspected re-inspection area and the range of affected data blocks. This is especially important when the connected component contains a suspected re-inspection page or has a type confidence level lower than [a certain threshold]. Output suggestions for re-sampling and re-inspection at the same location. Or the number of affected data blocks exceeds It outputs data block redundancy readout and error correction review suggestions. When only low-severity definitive defects exist, it outputs periodic re-inspection suggestions. The final output record includes defect coordinates, defect type, defect area, defect severity value, affected data block range, re-inspection suggestions, model version number, and calibration threshold version number.

[0137] 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.

[0138] This invention applies multi-image registration, page fusion feature extraction, normal patch memory distance calculation, page token map propagation, and conformal calibration threshold determination to the holographic page defect detection process in a continuous manner. This enables visible surface anomalies, scattering enhancement anomalies, and readout reconstruction anomalies to corroborate each other under a unified page coordinate system. As a result, an executable technical processing chain is formed to address the problems of early defects in holographic pages being difficult to detect and the range of defect influence being difficult to determine.

[0139] The present invention further constructs each coding unit or data block into a page token containing visual patch features, scattering intensity, frequency domain outliers, and readout reconstruction residuals. A coding unit graph is established based on page spatial adjacency, data block or error correction code grouping relationship, and optical point diffusion influence range. The data block risk vector is obtained through graph attention propagation, and then classified in combination with the conformal calibration threshold of the intra-page partition. This makes the defect judgment not only dependent on the local outliers of a single page, but also able to reflect the propagation influence of defects on adjacent page and data blocks in the same group.

Claims

1. A holographic page defect detection method based on computer vision, characterized in that, include: S1. Acquire surface reflection, oblique incidence scattering and readout reconstruction images of the holographic page to be inspected, complete page coordinate registration based on the periodic characteristics of the synchronization mark or coding array, divide the page grid according to coding unit or data block, and obtain the page grid image set and position information. S2. Input each page grid into the trained self-supervised visual encoder, extract multi-scale patch features, and calculate local gray-level uniformity value, frequency domain scattering response value and readout reconstruction residual value to form page grid fusion features; S3. Based on the partition patch memory of the normal holographic page, calculate the distance between the page grid fusion feature and the normal patch feature of the corresponding region to obtain the initial anomaly score; S4. Construct a page token that includes visual patch features, scattering response, frequency domain outliers and readout reconstruction residuals for coding units or data blocks. Establish a coding unit graph based on spatial adjacency, error correction grouping and point diffusion influence range. Obtain the data block risk vector through graph attention propagation. S5. Determine the conformal calibration threshold based on the normal and abnormal score distribution of the page partition and the preset target coverage quantile. Combine the initial abnormal score, data block risk vector, scattering change and readout reconstruction residual value to generate a defect probability map and defect mask. Then classify the page into confirmed defects, suspected re-inspection and normal areas according to the conformal calibration threshold. S6. Output the defect coordinates, defect type, defect area, defect severity value, affected data block range, and re-inspection suggestions based on the defect mask and location information.

2. The holographic page defect detection method based on computer vision according to claim 1, characterized in that, S1 includes: The surface reflection image, the oblique incidence scattering image, and the readout reconstructed page image are acquired respectively under the same page positioning reference. Extract the marker center point from the synchronization marker, or extract the horizontal and vertical periodic vectors from the periodic features of the coding array; The page coordinate transformation matrix is ​​obtained based on the center point of the mark or the horizontal periodic vector and the vertical periodic vector; The three types of images are mapped to a unified page coordinate system using the page coordinate transformation matrix, and page grid index, page grid boundary and page grid position information are generated based on the coding unit spacing or data block boundary.

3. The holographic page defect detection method based on computer vision according to claim 1, characterized in that, S2 include: Each page area is cropped into a page patch that includes the page body and the preset neighboring border; The page patch is input into the trained self-supervised visual encoder, and the patch representations output by different coding layers are extracted and pooled to obtain visual patch features; The ratio of the standard deviation of gray level within a page grid area to the gray level mean plus a preset constant is used as the local gray level uniformity value. A frequency domain transformation is performed on the grid region in the oblique incidence scattering image, and the proportion of the energy of the preset frequency band to the total energy of the grid is used as the frequency domain scattering response value. The pixel difference statistics between the readout reconstructed page image and the normal reference reconstructed page image within the same page grid area are used as the readout reconstruction residual value; The visual patch features, the local grayscale uniformity value, the frequency domain scattering response value, and the readout reconstruction residual value are normalized and spliced ​​together to obtain the page fusion features.

4. The holographic page defect detection method based on computer vision according to claim 1, characterized in that, S3 includes: The normal holographic page is divided into multiple partitions according to its position within the page, and the normal page grid fusion features are saved for each partition to form the partition patch memory library; For the holographic page to be inspected, retrieve a preset number of neighboring normal page fusion features within the corresponding partition; An initial anomaly score is generated based on the mean distance, nearest neighbor distance, or weighted distance between the feature distances between the page fusion feature to be inspected and the neighboring normal page fusion features; The initial anomaly score is bound to the page index and then used to construct subsequent page tokens and determine defects.

5. The holographic page defect detection method based on computer vision according to claim 1, characterized in that, S4 includes: Using the page index as the node identifier, the visual patch feature, the scattering response, the frequency domain outlier, the readout reconstruction residual, and the initial outlier score are concatenated into a page token vector; Spatial adjacency edges are generated according to the contact relationship of the page grid boundaries, group association edges are generated according to the grouping of the same data block or the same error correction code, and optical influence edges are generated according to the page grid coverage determined by the point spread function radius. Assign trainable attention weights to each connected edge according to the edge type, and update the lattice token vector through graph attention propagation; The updated page token vector is mapped to the probability value of being affected by defects corresponding to each data block, and the data block risk vector is composed of the probability values ​​of being affected by defects.

6. The holographic page defect detection method based on computer vision according to claim 1, characterized in that, S5 include: Based on the normal holographic page calibration set, the inconsistency score is calculated in each page partition, consisting of the initial anomaly score, the page grid component corresponding to the data block risk vector, the scattering change, and the readout reconstruction residual value. According to the preset target coverage, the quantiles are obtained from the empirical distribution of the non-consistency scores, and the quantiles are used as the conformal calibration thresholds corresponding to the partitions within the page. The initial anomaly score, the page component corresponding to the data block risk vector, the scattering change, and the readout reconstruction residual value are normalized and fused according to a preset weight to obtain a defect probability map. Pages in the defect probability map that reach the conformal calibration threshold are marked as identified defect areas. Pages that do not reach the conformal calibration threshold but reach the re-inspection threshold are marked as suspected re-inspection areas. The remaining pages are marked as normal areas. The defect mask is generated from the identified defect areas and the suspected re-inspection areas. The re-inspection threshold is obtained by multiplying the conformal calibration threshold by a preset scaling factor.

7. The holographic page defect detection method based on computer vision according to claim 6, characterized in that, S6 include: The connected pages in the defect mask are merged, and the defect coordinates are calculated inversely based on the merged page boundaries and page coordinate transformation matrix. The defect type is determined based on the linear brightness and darkness variation in the surface reflection image, the scattering variation in the oblique incidence scattering image, the reconstruction residual distribution in the readout reconstruction page image, and the page grid token vector. The defect type includes local scattering, scratches, contamination, and microstructure damage. Convert the merged page area into the defect area; The defect severity value is calculated based on a weighted combination of the defect probability map mean, defect area, readout reconstruction residual mean, and number of affected data blocks. The data blocks that intersect with the defect mask and the data blocks whose probability of being affected by defects after graph attention propagation reach the data block risk threshold are identified as the affected data block range, and re-inspection suggestions are output for suspected re-inspection areas.

8. The holographic page defect detection method based on computer vision according to claim 3, characterized in that, The self-supervised visual encoder obtained through training is acquired using normal holographic pages and holographic pages containing labeled defective regions as training samples. During training, occlusion recovery, adjacent page grid consistency constraints, and cross-imaging mode feature alignment are performed on the page grid patches to map the patch representations of the same page grid in the surface reflection image, oblique incidence scattering image, and readout reconstructed page image to the same feature space. After training is complete, the encoder parameters are frozen, and the page fusion features of the normal holographic page are written into the partition patch memory.

9. The holographic page defect detection method based on computer vision according to claim 6, characterized in that, Before being applied to the holographic page to be inspected, the conformal calibration threshold is also corrected by partitioning based on the background exposure distribution within the page. The partition correction includes calculating the average background grayscale value and the scattering baseline value of each partition within a page, and inputting the average background grayscale value and the scattering baseline value into a threshold correction mapping table to obtain the threshold correction coefficient; The corrected conformal calibration threshold is determined based on the conformal calibration threshold and the threshold correction coefficient. The revised conformal calibration threshold is used to replace the conformal calibration threshold of the corresponding page partition in the defect probability map threshold determination and re-inspection threshold calculation.

10. The holographic page defect detection method based on computer vision according to claim 5, characterized in that, The graph attention propagation uses a set of attention parameters corresponding to the edge type to calculate attention coefficients for spatially adjacent edges, grouped associated edges, and optically influential edges, respectively. The attention coefficients are normalized within the range of incoming edges of the same node, and the normalized attention coefficients are used to perform weighted aggregation of adjacent page token vectors to obtain the updated page token vector. When the probability values ​​of multiple page token vectors within the same error correction code group that are affected by defects reach the intra-group propagation threshold, the data blocks within the error correction code group that are not covered by the defect mask but have a group association edge with the defective page are added to the candidate list of affected data blocks. The candidate list of affected data blocks serves as the input for generating the range of affected data blocks and re-inspection recommendations.

Images