Holographic data recovery method based on image reconstruction

Abstract

This invention discloses a holographic data restoration method based on image reconstruction, belonging to the field of holographic storage data restoration. To solve the problem that it is difficult to restore the original file content under fragment reading, partial occlusion or complex disturbance of incomplete diffraction patterns and low signal-to-noise ratio page images, this invention improves the reliability and decodability of incomplete holographic page restoration by using registration and positioning, cross-domain feature fusion, holographic forward propagation constraints and diffusion-based completion to restore page data.

Description

Technical Field

[0001] This invention relates to the field of holographic data restoration, and more particularly to a method for holographic data restoration based on image reconstruction. Background Technology

[0002] Holographic storage technology achieves high-density data storage by recording page images or their diffraction-related information in a medium. The reading end typically needs to reconstruct the symbol array based on the diffraction pattern, page image, synchronization markers, and page layout parameters, and combine this with error correction and verification to obtain the original archive data. In applications such as disaster recovery, long-term archiving, and damage-resistant preservation, holographic media may only output partially observable data due to damage, obstruction, contamination, reading posture deviation, or environmental disturbances.

[0003] Existing processing methods mostly rely on relatively complete page images, fixed threshold segmentation, or single image enhancement processes, which do not make sufficient use of the physical correlation between missing regions, low signal-to-noise ratio regions, and diffraction domain information. At the same time, if there is a lack of forward propagation consistency constraints and code decodability feedback during the page image completion process, the problem of continuous image appearance but unstable code decision can easily occur, leading to an increase in the bit error rate of missing page recovery.

[0004] Furthermore, when the same holographic page comes from multiple media fragments, existing methods struggle to integrate reliable observations, physical residuals, and supplementary priors from different fragments within the page coordinate system. The overlapping and missing information between fragments cannot be fully coordinated, affecting the recovery value of the original archival content in fragmented media.

[0005] Therefore, a holographic data restoration method is needed to address the shortcomings of the existing technologies. Summary of the Invention

[0006] One objective of this invention is to propose a holographic data restoration method based on image reconstruction. Addressing the problem that existing technologies struggle to reliably restore original file content using incomplete diffraction patterns and low signal-to-noise ratio page images under conditions of fragment reading, partial occlusion, or complex environmental disturbances, this invention proposes a technical solution involving registration and localization of the data to be restored, construction of a defect mask and pixel confidence map, fusion of spatial reconstruction features and frequency domain reconstruction features, introduction of holographic forward propagation physical residuals, and combination of diffusion-based completion and decoding feedback to restore page data. This invention achieves the technical effects of improving the continuity of fragmented page restoration, the consistency of holographic propagation, and the decodability of code elements.

[0007] This invention provides a holographic data restoration method based on image reconstruction, comprising: S1, acquiring the data to be restored output from the reader, registering, normalizing, and locating the page region of the data to be restored to obtain a page observation image, a defect mask, and a pixel confidence map. The data to be restored includes a measured diffraction pattern, a page image, synchronization marker information, and holographic page layout parameters; S2, extracting spatial reconstruction features from the page observation image, the defect mask, and the pixel confidence map, and extracting frequency domain reconstruction features from the frequency domain representation of the measured diffraction pattern and the page observation image; S3, generating cross-domain attention weights based on the defect mask and the pixel confidence map, fusing the spatial reconstruction features and the frequency domain reconstruction features to obtain a coarsely restored page image; S4, using a holographic forward propagation model to... The coarse restored page image is converted into a simulated diffraction pattern. The physical residual between the simulated diffraction pattern and the measured diffraction pattern is calculated within the reliable region. Frequency domain gating weights and diffusion prior weights are generated from the physical residuals. The reliable region is determined on the page plane by the defect mask and pixel confidence threshold, and mapped to the diffraction plane reliable weights through the holographic forward propagation model. S5, The coarse restored page image, defect mask, pixel confidence map, frequency domain gating weights and diffusion prior weights are input into the diffusion-type completion module to complete the defect region and the region where the pixel confidence is less than the pixel confidence threshold. During the completion process, a decoding feedback map is generated based on the symbol probability matrix and reliability matrix. After the completion is completed, the original archive data is restored based on the page format constraints and error correction verification.

[0008] Optionally, S1 includes:

[0009] The page coordinate system is determined using the synchronization marker information and holographic page layout parameters. The page image is mapped to the page coordinate system, and the measured diffraction pattern is aligned with the geometric scale and / or frequency domain of the page coordinate system to generate a diffraction alignment record.

[0010] The mapped page image is normalized to grayscale and the background is subtracted to obtain the page observation image, wherein 1 in the defect mask represents a defect and 0 represents no defect;

[0011] The defect mask is generated based on the page boundary, the occlusion position of the synchronization mark, and the pixel position with no effective light intensity response;

[0012] The confidence level of each pixel is calculated according to a preset weight based on the normalized light intensity deviation, local noise variance, synchronization mark registration residual, and defect mask value, and the calculation results are normalized to a numerical range of zero to one to form the pixel confidence map.

[0013] Optionally, S2 includes:

[0014] Spatial reconstruction features are obtained by sampling the symbol grid of the page observation image and determining the symbol candidate boundary based on the holographic page layout parameters.

[0015] Geometric features of synchronization markers, symbol edge features, and local texture features are extracted within the observation area defined by the defective mask.

[0016] Using the pixel confidence map as sampling weight, the symbol candidate boundary, synchronization marker geometric features, symbol edge features and local texture features are weighted and aggregated to obtain the spatial reconstruction features aligned with the page coordinate system.

[0017] Furthermore, the frequency domain reconstruction features are obtained as follows: Fourier transforms are performed on the measured diffraction pattern and the observed page image to obtain the diffraction spectrum and the page spectrum, respectively;

[0018] Diffraction fringe features, edge component features with frequencies greater than the edge frequency band threshold, and noise distribution features are extracted according to the frequency band division rules determined by the holographic page layout parameters.

[0019] The effective observation mask obtained from the defective mask is converted to the frequency domain to obtain the effective frequency domain coefficients. The effective frequency domain coefficients are then used to weight the diffraction fringe features, edge component features, and noise distribution features to obtain the frequency domain reconstructed features.

[0020] Optionally, S3 includes:

[0021] The spatial retention weight is calculated based on the defect mask and the pixel confidence map. The spatial retention weight takes a first preset weight at the position where the defect mask is represented as undefected and the pixel confidence is not less than the pixel confidence threshold, and takes a second preset weight at other positions. The first preset weight is greater than the second preset weight.

[0022] The frequency domain compensation weight is calculated based on the proportion of pixels that are defective as represented by the defective mask and the effective frequency domain coefficients obtained by converting the effective observation mask obtained from the defective mask.

[0023] The spatial preservation weight, frequency domain compensation weight, spatial reconstruction feature, and frequency domain reconstruction feature are input into the cross-domain attention module to obtain the cross-domain attention weight.

[0024] The spatial reconstruction features and frequency domain reconstruction features are spliced, channel-weighted, and page coordinate resampled using the cross-domain attention weights to obtain fused features. The fused features are then decoded into page images to obtain the coarse restored page image.

[0025] Optionally, S4 includes:

[0026] The holographic forward propagation model is established based on the recording wavelength, propagation distance, pixel spacing, and holographic page layout parameters;

[0027] The coarse restored page image is used as the amplitude constraint of the complex amplitude of the object plane, and the simulated diffraction pattern is obtained through phase modulation and diffraction propagation calculations.

[0028] The defective mask is represented as an undefective page plane region with a pixel confidence level not less than the pixel confidence level threshold. Page plane confidence weights are generated and the diffraction plane confidence weights are obtained by mapping the holographic forward propagation model. Light intensity residual, frequency band energy residual, and synchronization mark position residual are calculated within the weighted region defined by the diffraction plane confidence weights.

[0029] The light intensity residual, frequency band energy residual, and synchronization mark position residual are normalized and weighted by preset coefficients to obtain the physical residual;

[0030] The frequency domain gating weights and diffusion prior weights are generated by looking up the gating mapping table based on the physical residuals.

[0031] Optionally, S5 includes:

[0032] The diffuse completion module includes a conditional encoder, a noise prediction network, and a page image decoder;

[0033] The conditional encoder encodes the coarse restored page image, the defect mask, the pixel confidence map, the frequency domain gating weights, and the diffusion prior weights to obtain the completion conditional features;

[0034] The noise prediction network predicts noise components based on the current noise page image, completion condition features, and decoding feedback map in each completion iteration.

[0035] The page image decoder updates the current noisy page image based on the predicted noise components and retains the update result in the completion area. At the position where the missing mask is represented as not missing and the pixel confidence is not less than the pixel confidence threshold, the page observation image is retained according to the spatial retention weight.

[0036] Furthermore, the decoding feedback image is generated in the following way: the current restored page image after each preset number of completion iterations is divided into code element units according to the holographic page layout parameters;

[0037] The probability of each candidate symbol value is calculated for each symbol unit to form the symbol probability matrix;

[0038] The candidate symbol values ​​of the same symbol unit are sorted from high to low probability. The difference between the probability of the first candidate and the probability of the second candidate is used as the corresponding reliability. The reliability matrix is ​​generated by combining the error correction and verification syndrome.

[0039] Symbols with reliability less than the symbol reliability threshold and symbols that failed the error correction check are identified as the set of symbols to be fed back.

[0040] The set of code elements to be fed back is rasterized according to the position of the code element unit in the page coordinate system to obtain the decoding feedback map;

[0041] Furthermore, the spatial reconstruction branch, frequency domain reconstruction branch, cross-domain attention module, and diffusion completion module are obtained through sample training;

[0042] The training samples include complete page images, sample diffraction patterns generated from complete page images through a holographic forward propagation model, sample page observation images obtained by applying occlusion and noise to complete page images, sample defect masks, and sample pixel confidence maps;

[0043] The training loss includes page image reconstruction loss, diffraction pattern consistency loss, symbol cross-entropy loss, and decoding feedback region weighted loss. The decoding feedback region weighted loss applies a greater loss weight to the page coordinate region corresponding to the set of symbols to be fed back than to the non-feedback region.

[0044] Furthermore, when the data to be restored comes from multiple media fragments of the same holographic page, steps S1 to S4 are performed on each media fragment to obtain the coarse restored page image, physical residual, frequency domain gating weight, and diffusion prior weight corresponding to each media fragment.

[0045] Based on the consistency of the synchronous mark position residual and the overlapping area of ​​the page coordinate system, the page coordinates of each media fragment are spliced ​​together.

[0046] Within the stitched page coordinate system, if there are multiple media fragment observation results for the same pixel position or the same symbol neighborhood, the pixel or symbol neighborhood corresponding to the media fragment with the smallest local physical residual value is selected as the observation retention source based on the local physical residual map. A stitched defect mask and pixel confidence map are generated, and the area not covered by the observed retention pixels and the area where the pixel confidence after stitching is less than the pixel confidence threshold are used as the completion area and input into the diffusion completion module.

[0047] The beneficial effects of this invention are:

[0048] 1. By using synchronized labeling information and holographic page layout parameters, registration, normalization, and page region localization are completed, and defect masks and pixel confidence maps are generated, enabling subsequent reconstruction to distinguish between reliable observation areas, defective areas, and low-confidence areas, thereby improving the targeting of incomplete page data processing.

[0049] 2. By extracting symbol boundaries, synchronization markers and local texture features through the spatial reconstruction branch, and extracting diffraction fringes, high-frequency edges and noise distribution features through the frequency domain reconstruction branch, the two types of features are then fused using a mask-guided cross-domain attention module. This allows for compensation of missing content using frequency domain global constraints when local image information is insufficient.

[0050] 3. The physical residual between the simulated diffraction pattern and the measured diffraction pattern is calculated by using a holographic forward propagation model. The completion process is guided by frequency domain gating weights, diffusion prior weights, and decoding feedback maps. This ensures that the restored page image is simultaneously constrained by image structure, propagation consistency, and code decodability, thereby reducing the risk of unstable code decision in the restoration of incomplete pages. Attached Figure Description

[0051] 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:

[0052] Figure 1 This is a flowchart of a holographic data restoration method based on image reconstruction.

[0053] Figure 2 This is a flowchart of step S5, diffusion-based completion and decoding feedback, of the present invention. Detailed Implementation

[0054] 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.

[0055] refer to Figures 1-2 The holographic data restoration method based on image reconstruction includes: S1, acquiring the data to be restored output from the reader, registering, normalizing, and locating the page region of the data to be restored to obtain the page observation image, the defect mask, and the pixel confidence map. The data to be restored includes the measured diffraction pattern, the page image, the synchronization marker information, and the holographic page layout parameters; S2, extracting spatial reconstruction features from the page observation image, the defect mask, and the pixel confidence map respectively, and extracting frequency domain reconstruction features from the frequency domain representation of the measured diffraction pattern and the page observation image; S3, generating cross-domain attention weights based on the defect mask and the pixel confidence map, fusing the spatial reconstruction features and the frequency domain reconstruction features to obtain the coarsely restored page image; S4, using a holographic forward propagation model to restore the coarsely restored page image. The page image is converted into a simulated diffraction pattern. The physical residual between the simulated diffraction pattern and the measured diffraction pattern is calculated within the reliable region. Frequency domain gating weights and diffusion prior weights are generated from the physical residuals. The reliable region is determined on the page plane by the defect mask and pixel confidence threshold, and mapped to the diffraction plane reliable weights through the holographic forward propagation model. S5. The coarse restored page image, defect mask, pixel confidence map, frequency domain gating weights and diffusion prior weights are input into the diffusion-type completion module to complete the defect region and the region where the pixel confidence is less than the pixel confidence threshold. During the completion process, a decoding feedback map is generated based on the symbol probability matrix and reliability matrix. After the completion is completed, the original archive data is restored based on the page format constraints and error correction verification.

[0056] In this specific embodiment, S1 includes:

[0057] The data restoration system receives data packets to be restored in a single reading task at the reading end as the processing unit. The data packets to be restored include the measured diffraction pattern, the page image initially output by the reading end, synchronization mark information, and holographic page layout parameters. The system establishes a page-level record for the reading task. The page-level record includes the page identifier, medium fragment identifier, reading time, synchronization mark candidate coordinates, recording wavelength, number of page symbol rows and columns, symbol spacing, and page boundary size. In this specific embodiment, a defect mask value of 1 indicates a defect or unusable observation, and a defect mask value of 0 indicates the existence of a usable observation. The pixel confidence map ranges from zero to one, and the larger the value, the more suitable the pixel is as the observation basis for spatial reconstruction and physical residual calculation.

[0058] Synchronous labeling and filtering units in solving First, construct a candidate set of synchronization tags. Each candidate record includes a marker number, reading end coordinates, theoretical page coordinates, detection confidence, and neighborhood occlusion ratio. The system filters candidate synchronization markers according to the following rules: the detection confidence is not less than the marker threshold, the neighborhood occlusion ratio does not exceed the occlusion threshold, and the relative distance error with the adjacent synchronization marker does not exceed the page distance tolerance. When there are multiple candidates with the same number, the candidate with the highest detection confidence and the smallest relative distance error is selected. If the two are tied, the candidate with the smallest local noise variance is selected and a tie selection flag is written in the page-level record. This filtering result is used as the input for the registration module to solve the page coordinate system.

[0059] The registration module reads the marker number and reading end coordinates from the synchronization marker information, and constructs a page coordinate system based on the marker theoretical coordinates in the holographic page layout parameters. It then uses homography transformation to map the reading end image coordinates to the page coordinates. This homography transformation is denoted as... ,in Indicates a media fragment identifier or a single-page read identifier. By minimizing get, For the first The coordinates of each synchronization marker in the reading end coordinate system The theoretical coordinates of the synchronization marker in the page coordinate system are given. If fewer than four synchronization markers are involved in the solution or the registration residual exceeds a preset registration threshold, the system marks the fragment page-level record as pending verification and stops using it as a reliable observation input for subsequent stitching. After successful registration, the page image is resampled to the page coordinate grid. For the measured diffraction pattern, the system determines the diffraction alignment record based on the Fourier optical sampling theorem, and the geometric scale factor is determined by the pixel spacing of the page coordinates. Pixel pitch of the sensor at the reading end Transmission distance and recording wavelength The page coordinate frequency was determined jointly. With the sensor coordinates at the reading end according to Establish scale correspondence; frequency domain sampling mapping is determined by the number of sensor sampling points. and sensor pixel pitch Determined as , The system then generates the diffraction resampling coordinate table used by S4. The page image is resampled using bilinear interpolation, and the diffraction spectrum is resampled using frequency domain interpolation with a Hanning window. The geometric scale factor, frequency domain sampling mapping, resampling coordinate table, interpolation method, window function, and reading end coordinate range are all written into the diffraction alignment record.

[0060] The page region positioning unit generates the effective region within the page based on the page boundary size and the outer frame of the symbol grid. The system writes three types of cause codes—boundary defect, marker occlusion defect, and no light intensity response defect—to the resampled pixels outside the page boundary, the synchronous marker occlusion extension area, and the area with no effective light intensity response, respectively. These cause codes are associated with the defect mask. The shared storage enables S2 to distinguish between unsampled areas outside the boundary, synchronously marked occluded areas, and areas where real page content is missing when extracting spatial reconstruction features;

[0061] The normalization module performs dark background subtraction and grayscale normalization on the mapped page image to obtain the page observation image. In this specific implementation method, according to calculate, The grayscale value of the mapped page image. The background gray level is estimated from the dark field region outside the page boundary and the neighborhood outside the synchronization marker. and These are the 95th and 5th percentile values ​​of the grayscale of the effective sampled pixels within the page, respectively. To prevent zero grayscale measurement and to use a quantization interval of grayscale resolution for the same reading batch, the page observation image is written to the page observation cache and kept pixel-by-pixel aligned with the page coordinate system;

[0062] The mask generation module generates a defect mask based on the page boundary, the occlusion position of the synchronous marker, and the pixel position with no effective light intensity response. ,when When the image falls outside the page boundary, falls into the occlusion extension area of ​​the synchronization marker, corresponds to pixel saturation at the original read end, is below the dark noise upper limit, or lacks a resampling source, Set to 1 otherwise set to 0. The occlusion extension area of ​​the synchronization mark is obtained by expanding the theoretical outer frame of the synchronization mark by one symbol width. The threshold for determining no effective light intensity response is obtained by adding 3 times the standard deviation to the mean of the dark noise samples of the blank medium at the reading end. The system also saves the proportion of missing pixels for each symbol unit for S2 and S3 to determine the sampling weight.

[0063] The confidence calculation module calculates the pixel confidence corresponding to the normalized light intensity deviation, local noise variance, synchronization mark registration residual, and defective mask value in the page coordinate system, first using... Indicates the normalized light intensity deviation. Indicates Centered Neighborhood local noise variance This represents the residual term obtained by normalizing the registration residual of the most recent synchronization marker according to the page diagonal length, and then according to... Pixel confidence map ,in In this specific implementation, the weights are determined using complete page samples and occluded page samples before the reader is deployed and stored in the confidence weight table. This serves as the local noise variance benchmark for complete page samples within the same batch;

[0064] When the data to be restored comes from multiple media fragments of the same holographic page, the system identifies them by the fragments. Perform the registration, normalization, page region localization, defect mask generation, and pixel confidence calculation steps described above, respectively, so that each fragment forms an independent page observation image. Defective mask and pixel confidence map And record in the fragment index table Mean of registration residuals, effective observation area, and visible status of synchronization markers;

[0065] In scenarios with multiple fragments, the system also establishes a page coordinate covering index in stage S1. This is used to record the observation results of media fragments at each page coordinate pixel position. The coverage index field includes fragment identifier, pixel page coordinate, observation gray level, defect mask value, pixel confidence and registration residual summary. When there are multiple media fragment observation results at the same pixel position, S1 does not immediately merge the pixels, but writes all fragment observation results into the coverage index, so that S4 can select the observation and retain pixels according to the rule of the smallest local physical residual value in the local physical residual map.

[0066] After S1 is completed, the data restoration system will view the page image. Defective mask Pixel confidence map Page coordinate system, covering index The fragment index table is written into the restoration task cache and used as the input object for S2 to extract spatial reconstruction features and frequency domain reconstruction features;

[0067] To ensure that S2 can stably read the output of S1, the restoration task cache uses the same page coordinate size and the same pixel index order for the page observation image, defect mask, and pixel confidence map. The origin, row and column span, symbol grid version, and effective observation area are written in the cache header. If the size of any output object is inconsistent with the page layout parameters, the system switches the reading task status to the registration abnormal state and prohibits entry into the feature extraction branch until a consistent page coordinate cache is regenerated.

[0068] The aforementioned page-level records also store candidate synchronization marker filtering results, cause code statistics, dark field background parameters, confidence weight table version, and coverage index version. When subsequent steps discover a physical residual or decoding feedback anomaly in a fragment, the system can read back these fields to locate the source of the anomaly and re-execute the registration and confidence calculation of the corresponding fragment without changing the original read data.

[0069] In this specific embodiment, S2 includes:

[0070] The feature extraction module reads the page observation image written by S1. Defective mask Pixel confidence map Based on the holographic page layout parameters, a symbol grid is established according to the number of symbol rows, number of symbol columns, symbol spacing, and theoretical position of the synchronization marker in the page layout parameters. Each symbol unit in the grid is denoted as... , The symbol cell number is used. The candidate boundary of the symbol is obtained by translating the theoretical boundary in the page coordinate system through the local residual correction. The candidate boundary coordinates, cell center, cell effective observation ratio and adjacent cell number are written into the symbol sampling table.

[0071] The local residual correction amount is jointly determined by the synchronization mark monography transformation residual field and the edge response of the page observation image. The system first interpolates the registration residuals of each synchronization mark to the symbol grid according to the inverse distance ratio to form the residual field. Then, within a search window of one symbol width on both sides of the symbol theoretical boundary, the edge response peak offset is searched. ,according to Obtain symbol unit Boundary correction amount, and Based on the complete page sample, S2 only uses the boundary correction amount to perform feature extraction once, and subsequent S3 and S5 do not rewrite the code symbol candidate boundary in reverse;

[0072] The spatial reconstruction branch extracts synchronization marker geometric features, symbol edge features, and local texture features within the observation area defined by the missing mask. Synchronization marker geometric features include synchronization marker center offset, side length scaling ratio, and rotation angle residual. Symbol edge features are formed by Sobel gradient magnitude and boundary orientation histogram. Local texture features are derived from... Neighborhood gray-level co-occurrence statistics and local binary pattern statistics are formed, all Pixels are not included in the statistics. Below the pixel confidence threshold Pixels are only included in the aggregation as low-weight samples;

[0073] For each symbol unit The spatial reconstruction branch uses the pixel confidence map as the sampling weight to perform weighted aggregation, according to... Calculate the spatial aggregation vector. To observe the image in pixels from the page The generated grayscale, gradient, edge direction, and texture joint features, To prevent zero weight and ensure that the minimum confidence weight of a pixel has the same dimension, if the denominator is less than the preset effective sampling threshold, the spatial aggregation vector of the symbol unit is filled by the median vector of the adjacent undamaged symbol units and a low confidence flag is written into the symbol sampling table.

[0074] The spatial reconstruction branch concatenates the symbol candidate boundaries, synchronization marker geometric features, symbol edge features, and local texture features into spatial reconstruction features aligned with the page coordinate system. ,in Each record includes a symbol unit number. Page coordinate boundaries, synchronization marker residual summary, edge strength, texture response, effective observation ratio and low confidence flag. The system performs boundary consistency verification on adjacent code units. When adjacent candidate boundaries overlap by more than half a code width, the boundary with the largest confidence weighted edge strength value is retained and the deleted boundary is written into the pruning record.

[0075] The frequency domain reconstruction branch is used to reconstruct the measured diffraction pattern. Page observation images Perform a two-dimensional Fourier transform to obtain the diffraction spectrum. and page spectrum The frequency band division rules are determined by the symbol spacing, page size, and recording wavelength in the holographic page layout parameters. The system divides the frequency domain plane into a synchronization marker low-frequency band, a symbol main frequency band, an edge high-frequency band, and a noise estimation frequency band, and assigns frequencies greater than the edge frequency band threshold. The components are labeled as edge component candidates; where, and Frequency domain coordinates;

[0076] The frequency domain reconstruction branch generates diffraction fringe features, edge component features, and noise distribution features according to the frequency band. The diffraction fringe features include the principal direction of the fringes, peak frequency, and peak width. The edge component features include the proportion of high-frequency energy and directional energy difference. The noise distribution features include the mean energy and variance within the noise estimation frequency band. The edge frequency band threshold... The Nyquist frequency corresponding to the symbol spacing in the page layout parameters is multiplied by a preset ratio. In this specific embodiment, the ratio is calibrated by the frequency band energy distribution of the complete page sample and stored in the frequency band parameter table.

[0077] The mask frequency domain conversion module first uses a defective mask. Generate an effective observation mask ,in A value of 1 indicates that the pixel at that coordinate on the page has an available observation, while a value of 0 indicates a missing or unavailable observation. The effective observation mask is then transformed to the frequency domain to obtain the effective frequency domain coefficients. Specifically, for... Perform a two-dimensional Fourier transform and obtain the energy spectrum Then calculate the normalized mean energy according to the frequency band, and follow... Get the first The effective coefficients in the frequency domain of each frequency band, Indicates the first A set of frequency points in a frequency band. This indicates the energy spectrum reference value of the effective observation mask for a complete page of the same format in this frequency band. To prevent zero energy, this specific implementation uses the energy spectrum instead of the real part, in order to avoid phase sign cancellation and improve stability under occlusion boundary perturbations. As a relative weighting and gating effective information support, it increases overall with the increase of available observation coverage and connectivity, but it does not require strict physical inversion of the effective observation mask from the frequency domain effective coefficients;

[0078] The frequency domain reconstruction branch uses effective frequency domain coefficients to weight diffraction fringe features, edge component features, and noise distribution features, according to... Generate the first Frequency domain aggregation vector of each frequency band The effective signal characteristics consist of diffraction fringes and edge components. As for the noise distribution characteristics, if a certain frequency band If the frequency band is below the effective threshold in the frequency domain, it is retained but participates in the cross-domain attention calculation as a frequency band to be compensated in S3.

[0079] The frequency domain parameter table is created from a complete page sample of the same format before deployment. The table fields include format identifier, symbol spacing, recording wavelength, frequency band number, upper and lower limits of the frequency band, and edge frequency band threshold. The frequency domain effective threshold and update time are used to calculate the upper and lower limits of the frequency band in real time according to the symbol interval when the format identifier of the current reading task cannot be matched with the table. The frequency domain reconstruction branch takes the frequency domain effective threshold as the lower limit of the full page sample statistics and then writes the task into the supplementary calibration record.

[0080] After S2 is completed, the feature extraction module will reconstruct the spatial features. Frequency domain reconstruction features Symbol sampling table, effective frequency band coefficient The pruned records are written to the feature cache and used as input for S3 to compute cross-domain attention weights and generate coarse restored page images.

[0081] In this specific embodiment, S3 includes:

[0082] The cross-domain fusion module reads the spatial reconstruction features output by S2. Frequency domain reconstruction features Effective frequency band coefficient and the defect mask output by S1 and pixel confidence map First, calculate the spatially preserved weights in the page coordinate system. ,when and hour, Take the first preset weight Otherwise, take the second preset weight. In this specific implementation method Furthermore, both were calibrated using a full-page observation fidelity experiment and then written into a spatially reserved weight table;

[0083] The cross-domain fusion module calculates the frequency domain compensation weights based on the proportion of missing pixels represented by the defective mask and the effective frequency domain coefficients. The proportion of missing pixels is denoted as... , The total number of pixels in the page coordinate system, the first... The frequency domain compensation weights for each frequency band are based on get, , and The parameters are derived from the frequency domain compensation parameter table and satisfy the non-negativity constraint. The larger the value, the more compensation from the frequency domain reconstruction branch is needed for that frequency band during fusion;

[0084] The cross-domain attention module converts spatial and frequency domain reconstruction features into page coordinate tokens of the same channel dimension. The spatial token consists of symbol unit number, boundary coordinates, edge intensity, texture response, and spatial preservation weights. The frequency domain token consists of frequency band number, fringe direction, peak frequency, high-frequency energy percentage, noise energy, and frequency domain compensation weights. The system obtains the query vector through linear projection. Key vector Sum value vector and in accordance with Computation symbol unit For frequency band Cross-domain attention weights The projection dimension is given by the training configuration table;

[0085] To make the missing region more reliant on frequency domain compensation, the cross-domain attention module writes spatial preservation weights and frequency domain compensation weights into the attention bias, according to... Correct attention score, For code unit The inner space retains the weighted mean. The mask bias coefficients obtained from the occlusion samples are corrected and then normalized to obtain the cross-domain attention weights. If the spatial effective observation ratio of a certain symbol unit is lower than the preset unit threshold, the attention of that unit is normalized only in the frequency band where the frequency domain effective coefficient is not lower than the frequency domain effective threshold.

[0086] As an optional improvement and stabilization process, the cross-domain attention module performs computational... A frequency band candidate set is established for each symbol unit. The candidate set consists of frequency bands with effective coefficients in the frequency domain not lower than the effective threshold in the frequency domain, high-frequency edge bands that match the edge direction of the symbol, and low-frequency bands for synchronization markers. If the value is empty, the main frequency band of the symbol is retained as a backoff candidate and an empty candidate flag is written. If the difference in the corrected attention score of multiple frequency bands is less than the tie threshold, the frequency band with the largest effective coefficient value in the frequency domain is selected first. If they are still tied, the frequency band with the smallest noise energy value is selected to participate in the dominant weighting.

[0087] The cross-domain fusion module utilizes cross-domain attention weights to stitch together spatial and frequency domain reconstructed features, perform channel weighting, and resample page coordinates, according to... Obtain symbol-level fusion features. This is a channel-weighted network obtained from training with samples. Square brackets indicate channel concatenation. The statistical vector of weights is preserved in the space within the symbol unit, and the fused features are resampled back to the page coordinate grid to form a fused feature map. ;

[0088] As an optional improvement in fusion mode selection, channel-weighted network Before output, the fusion module sets candidate fusion modes for the spatial and frequency domain branches. These candidate fusion modes include spatial preservation mode, frequency compensation mode, and equalization fusion mode. The mode evaluation value is determined according to... calculate, The normalized noise energy of the candidate frequency band. , and From the fusion mode weight table, the system selects... The largest candidate fusion mode, when the evaluation values ​​are tied, the candidate that is consistent with the previous adjacent symbol unit mode is selected first, so as to reduce inter-block mutation after page coordinate resampling;

[0089] Page image decoder reads fused feature map Output coarse restored page image In this specific embodiment, the page image decoder adopts a convolutional decoding structure that includes residual convolutional blocks and upsampling blocks. The output channel is the grayscale channel of the page image. The output value is cropped and kept in the range of zero to one. The decoder also outputs the coarse decision confidence at the symbol unit level, which is used to perform consistency verification with the symbol probability matrix when S5 generates the decoding feedback map.

[0090] After the coarse restored page image is generated, the cross-domain fusion module performs... Perform page format constraint checks, comparing the synchronization marker region, symbol boundary region, and missing region with the holographic page format parameters. If the coarse decision confidence of a symbol unit is lower than the coarse decision threshold, then the unit number is added to the coarse restoration low-confidence set. The set does not change the coarse restored page image output by S3, but serves as a priori hint field when S5 generates the decoding feedback graph;

[0091] When the same holographic page has multiple media fragments, S3 performs S3 on each fragment Generate coarse restored page images respectively and fusion feature map The missing percentage of each fragment is recorded in the fragment fusion cache. The average spatial retention weight and frequency domain compensation weight summary are used by S4 to calculate the physical residual and determine the observation reliability of each fragment when performing page coordinate stitching.

[0092] After S3 is completed, the cross-domain fusion module will add cross-domain attention weights. fusion feature map Coarse reproduction page image Space retention weight Frequency domain compensation weights and coarsely restored low-confidence sets Write it to the restoration task cache as input for S4 to build the simulated diffraction pattern and calculate the physical residual;

[0093] The aforementioned cross-domain fusion process stores the candidate frequency band set, candidate fusion mode, parallel processing flag, and backoff candidate flag in the attention record. Subsequently, S4 and S5 can distinguish frequency domain compensation caused by defects, low-confidence decision caused by noise, and local resampling changes caused by fusion mode switching based on this record.

[0094] In this specific embodiment, S4 includes:

[0095] The physical consistency module reads the coarse restored page image output by S3. The S1 outputs the page coordinate system and holographic page layout parameters, and is based on the recording wavelength. Transmission distance Pixel pitch A holographic forward propagation model is established based on the page size. This model adopts angular spectrum propagation, and the complex amplitude of the object plane on the page coordinate grid is denoted as... , The phase template is obtained by the phase calibration before deployment at the reading end. If the phase template is missing, the phase estimation network trained with the same template sample outputs the phase and writes the phase estimation flag into the page-level record.

[0096] The physical consistency module performs phase modulation and diffraction propagation calculations on the complex amplitude of the object surface, according to... The simulated diffraction pattern was obtained. For the distance of transmission The corresponding angular spectral transfer function is calculated by setting frequencies exceeding the propagation passband to zero, and the diffraction pattern is simulated. Resample the measured diffraction pattern according to the diffraction alignment record of S1. The diffraction plane of the readout end or its frequency domain coordinate system, both of which maintain the same frequency domain sampling interval at the readout end;

[0097] The trusted region generation module first determines the page plane trusted weights based on the defect mask and pixel confidence thresholds. Then As a propagation operator with non-negative real amplitude input, consistent with the holographic forward propagation model. ,according to Obtain the non-negative weighted energy of the diffraction plane, and according to... Normalization yields the confidence weights of the diffraction plane, where and The quantile statistical value of the current fragment weighted energy is below the weighted noise threshold. The position is set to zero; this mapping ensures non-negativity of weights and a defined range by pruning the sum of squared amplitudes, and transforms the support range of the page plane reliable observations for the diffraction plane residual comparison into a weighted masking basis, if If the effective weight area is lower than the preset ratio, the physical consistency module will not output a high-confidence gating result, but will instead limit the frequency domain gating weight to the lower limit of the training sample statistics and mark the fragment as a low-physical-confidence fragment.

[0098] Within the weighted region defined by the confidence weights on the diffraction plane, the physical consistency module calculates the intensity residual, frequency band energy residual, and synchronization marker position residual between the simulated and measured diffraction patterns. The intensity residual is calculated according to... get, The average light intensity of the measured diffraction patterns in the same batch within the weighted region is given by the frequency band energy residual according to... get, Indicates the first Within each frequency band Weighted summation of energy, the zero-prevention quantity and They are dimensionless, respectively, and have the same dimensions as light intensity and frequency band energy;

[0099] The synchronization marker position residual is determined by the difference in synchronization marker positions obtained from the simulated diffraction pattern and the measured diffraction pattern under the same backpropagation positioning process. The system uses the angular spectrum backpropagation operator. Each and Numerical backpropagation to page plane intensity map and The simulated synchronization mark position is obtained by performing normalized cross-correlation localization with the synchronization mark template within a preset search window around the theoretical location of the same synchronization mark. Synchronous marking position with actual measurement ; Synchronization mark position residual according to calculate, To simulate and measure the number of synchronization markers that were successfully located on both sides, The diagonal length of the page; if the measured diffraction pattern fails to be backpropagated for positioning but the synchronization mark in the S1 page observation image is successfully positioned, then the page observation synchronization mark position saved in S1 is used as the reference. If the relevant peak value on the simulated or measured side is lower than the positioning threshold and there is no S1 back-off position, the mark will not participate in the mean calculation and will be written into the positioning failure flag. If the number of valid marks is insufficient, the fragment will be given a high penalty value for the synchronous mark position residual and will be stopped as a splicing reference.

[0100] The physical consistency module normalizes the light intensity residual, frequency band energy residual, and synchronization marker position residual, and weights them according to preset coefficients to obtain the physical residual. ,in Each coefficient is obtained by calibrating the propagation consistency between complete page and occluded page samples. A smaller value indicates a greater consistency between the coarsely restored page image and the measured diffraction pattern within the reliable region, and the system will... Simultaneously write to page-level records and fragment index tables;

[0101] The gating generation module looks up the gating mapping table based on the physical residual. The fields of the gating mapping table include residual interval, frequency band number, frequency domain gating coefficient, diffusion prior coefficient, applicable version, version number, and update time. Each record in the table is the residual interval. Furthermore, when the frequency band number is the main frequency band of the symbol, the frequency domain gating coefficient is taken as 0.70 and the diffusion prior coefficient is taken as 0.45. This record is obtained from the binning statistics of the reconstruction error of the occlusion samples before deployment. If the result does not fall within any closed interval, the nearest residual interval is used as the record and written into the nearest neighbor lookup table flag;

[0102] Frequency domain gating weights and diffusion prior weights are generated according to the lookup table results. The frequency domain gating weights are denoted as... It is used to update the amplitude of the S5-constrained noise prediction network in different frequency bands, and the diffusion prior weights are denoted as... And by Defective mask and pixel confidence map Expanding together to a page coordinate grid, the system improves performance in defective and low-confidence regions. Reduce in trusted regions This allows diffusion completion to prioritize the modification of missing or low-confidence pixels;

[0103] As an optional improvement to the gating protection process, after gating mapping is completed, the physical consistency module performs a gating feasibility check. The check includes whether the frequency domain gating weights are within the range of zero to one, whether the diffusion prior weights are greater in the defective region than in the reliable region, and whether the gating difference between adjacent fragments in the same frequency band exceeds the gating mutation threshold. If any check is infeasible or violates the above constraints, the system freezes the gating update for that frequency band. The regression is to the median gating value of the training samples, and... Limit the diffusion prior to a preset upper limit to avoid amplifying the noise prediction update magnitude in S5 due to infeasible gating results;

[0104] When the data to be restored comes from multiple media fragments of the same holographic page, the physical consistency module performs S1 to S4 on each fragment and then uses the position residual of the synchronization mark. To ensure consistency with the overlapping areas of the page coordinate system, page coordinate splicing is performed on the two fragments. and overlap consistency according to calculate, The overlapping region is defined as the area where both pixels are intact and have reached the pixel confidence threshold. If the splicing consistency threshold is exceeded, the coordinates of the fragment with the smallest residual value at the synchronization mark position are retained as the splicing reference, and the other fragment is marked as pending review.

[0105] Within the stitched page coordinate system, when multiple media fragment observations exist at the same pixel location, the system calculates a local physical residual map within the overlapping region. The local physical residual map has the same resolution as the page coordinate grid, and is a pixel-level residual map; its values ​​are calculated by back-projecting the diffraction plane weighted intensity residual onto the page plane, within the symbol neighborhood containing the current pixel. The mean value is obtained by internal smoothing, i.e. This reduces noise through symbol neighborhood statistics but still outputs residual values ​​based on pixel positions, and then... The pixel corresponding to the medium fragment with the smallest local physical residual value is selected as the observation retention pixel on a pixel-by-pixel basis; when the system is configured for symbol-level stitching, pixels within the same symbol neighborhood are selected as the observation retention pixels. The mean value is filled into all pixels in the neighborhood, and the best value is selected based on the symbol neighborhood. This represents the set of fragments that cover the coordinate pixels of this page or their corresponding symbol neighborhood and meet the registration pass condition;

[0106] After the number of pixels to be retained for observation is determined, the system generates a stitched image of the observed page. Defective mask and pixel confidence map ,in , and All inherited from Corresponding source observation records for fragments; if If empty, then Set to 1 and Set to 0, and uniformly complete the region according to The calculation, meaning the union of the unobserved preserved pixel coverage area and the low-confidence area, is the coarsely reconstructed page image after stitching. Frequency domain gating weights and diffusion prior weights As the unified input for S5;

[0107] For fragments whose physical residuals exceed the residual alarm threshold twice consecutively, the physical consistency module does not delete the page observation results of the fragment, but writes its physical constraint state as a low confidence constraint state. When splicing multiple fragments, it only allows them to supplement the page coordinate area not covered by other fragments. If all fragments are in a low confidence constraint state, the system retains the fragment with the smallest physical residual value as the reference fragment, and switches the diffusion prior weight to the conservative completion level and outputs a low confidence completion flag to S5.

[0108] After S4 is completed, the physical consistency module will simulate the diffraction pattern. Physical residuals Frequency domain gating weights Diffusion Prior Weights In addition, the splicing results generated in multi-fragment scenarios are written to the physical constraint cache as input for S5 diffusion completion and decoding feedback updates.

[0109] In this specific embodiment, S5 includes:

[0110] The completion module reads the coarse restored page image, defect mask, pixel confidence map, frequency domain gate weights, and diffusion prior weights written by S4; if S4 outputs a multi-fragment stitching result, it reads the stitched coarse restored page image. The missing mask after splicing Pixel confidence map after stitching Unified supplementation of regions Frequency domain gating weights and diffusion prior weights And the diffusion completion area is directly determined as ,in This is equivalent to the region obtained by combining the stitched defect mask and pixel confidence map according to the union rule of the defect area and the low confidence area; if S4 does not output the multi-fragment stitching result, then the region corresponding to a single fragment is read. , , , and and will The defective area and The low-confidence regions were jointly identified as diffusion completion regions. ;

[0111] The diffusion-based completion module includes a conditional encoder, a noise prediction network, and a page image decoder. The conditional encoder aligns the coarsely restored page image, the defect mask, the pixel confidence map, the frequency domain gating weights, and the diffusion prior weights by channel before inputting them into the convolutional encoder. The frequency domain gating weights are then mapped to the page coordinate grid through a frequency band embedding layer to obtain the completion conditional features. , The fields include page coordinate position encoding, coarse restoration grayscale, mask state, pixel confidence, frequency band gating response, and diffusion prior strength. These fields serve as conditional inputs to the noise prediction network in each completion iteration. Simultaneously, the frequency domain gating weights also participate in subsequent FFT update constraints as explicit frequency band filtering coefficients.

[0112] The completion process uses preset settings. The current noisy page image is denoted as the result of the next backdiffusion iteration. The noise prediction network in the first Read in the next iteration Complete the conditional features and decoding feedback diagram ,according to Predict noise components, This is a U-shaped convolutional noise prediction network trained from samples. The network outputs a noise estimation map of the same size as the page coordinate grid. If the decoding feedback map is empty, then... Setting the matrix to zero does not trigger symbol feedback bias;

[0113] The page image decoder updates the current noisy page image based on the predicted noise components, according to... Generate candidate update page images. , and For the spread noise scheduling table, the first The coefficients corresponding to the next iteration. For standard Gaussian noise terms, the noise scheduling table is calibrated from the complete page of training samples and stored in the model configuration file. The system then retains this in the completion region. The updated results;

[0114] To ensure that the frequency domain gating weights fall within the update amplitude constraint, the completion module will use candidate update page images. With the current noisy page image The difference is defined as the candidate update quantity. and to Performing a two-dimensional Fourier transform yields ; then according to For each frequency point, the corresponding frequency band Apply gating, and then obtain the gated page field update amount through inverse Fourier transform. Ultimately Superimpose the current noisy page image to obtain It then enters the spatially preserved weight constraint, which allows the frequency domain gating weights to explicitly suppress or amplify the spread update amplitude of the corresponding frequency band;

[0115] At locations where the defective mask represents an undefective image and the pixel confidence is not less than the pixel confidence threshold, the page image decoder retains the page observation image according to spatial preservation weights, with the update rule being... ,in To preserve pixels in the page observation image or the observation after stitching together multiple fragments obtained from S1, The spatial retention weights obtained from S3 are determined by this rule, which ensures that the reliable observation regions mainly inherit the original observations, while the missing and low-confidence regions mainly adopt the diffusion update results.

[0116] Each preset number of times After the completion iteration, the decoding feedback module will restore the current page image. The holographic page layout is divided into symbol units according to the parameters, and the probability of candidate symbol values ​​is calculated for each symbol unit. The symbol classifier reads the grayscale block, the edge difference of adjacent units, and the page format position code within the unit, and outputs the symbol probability matrix. ,in The code element unit number, For candidate symbol values, Candidate symbol values The classification score and probability matrix are written to the decoding cache;

[0117] The decoding feedback module sorts the candidate symbol values ​​of the same symbol unit from highest to lowest probability, uses the difference between the probability of the first and second ranked candidate symbols as the corresponding reliability, and generates a reliability matrix by combining the error correction and verification syndrome. Obtain candidate differences, according to To obtain the overall reliability, and Let represent the probabilities of being the first and second ranked candidates, respectively. Representation of symbol unit Whether it participates in the verification equation that failed the error correction check. This is the penalty coefficient for the syndrome. The smaller the value, the more feedback and correction are needed for that symbol;

[0118] When sorting code symbol candidates, the decoding feedback module uses the candidate probability from high to low as the main sorting rule, the legality of the candidate code symbol value in the page format code table as the first parallel rule, and the reduction of the syndrome after the candidate code symbol value participates in the error correction verification as the second parallel rule. When the probability difference between the first and second ranked candidates is less than the probability parallel threshold, the system does not force the output of a single stable code symbol, but writes the code symbol unit into the code symbol set to be fed back and retains two candidate values ​​for the next round of completion iteration to enhance diffusion correction in the corresponding page coordinate area.

[0119] when Less than the symbol reliability threshold or At that time, the decoding feedback module will output the symbol unit. Write the code set to be fed back, and then rasterize the code set according to the position of the code unit in the page coordinate system to obtain the decoding feedback map. The decoding feedback map takes 1 in the page coordinate region corresponding to the code to be fed back, and 0 in other regions, and is used as a conditional input to increase the diffusion prior weight of the corresponding region in the next round of noise prediction.

[0120] Before the decoding feedback graph is input into the next round of the noise prediction network, the completion module will... Multiply by the diffusion prior weights to form a feedback prior graph and in accordance with Increase the padded strength in low-reliability symbol regions. The feedback gain is calibrated by the decoding error rate of the validation set. If the same symbol still fails the error correction check after two consecutive feedbacks, the feedback gain will remain unchanged and a difficult-to-recover symbol flag will be written to avoid excessive smoothing in local areas.

[0121] The spatial reconstruction branch, frequency domain reconstruction branch, cross-domain attention module, and diffusion-type completion module are obtained through sample training. The training samples include complete page images, sample diffraction patterns generated from complete page images through a holographic forward propagation model, sample page observation images obtained by applying occlusion and noise to complete page images, sample defect masks, and sample pixel confidence maps. Occlusion samples are generated according to four types of rules: synchronous marker occlusion, random block defects, stripe perturbation, and multi-fragment cropping. During training, complete page images are used as supervision targets and sample page observation images are used as input observations.

[0122] The training loss includes page image reconstruction loss, diffraction pattern consistency loss, symbol cross-entropy loss, and decoding feedback region weighted loss, with the total loss calculated according to... calculate, Constrain the consistency between the restored page image and the complete page image. The diffraction pattern of the constrained restored page after holographic forward propagation is consistent with the diffraction pattern of the sample. Constrain symbol candidate probabilities, Apply a loss weight greater than that to the non-feedback region to the page coordinate region corresponding to the set of feedback symbols. , and Obtained from the bit error rate calibration of the validation set decoding;

[0123] In this specific implementation, the confidence weight table, frequency band parameter table, fusion mode weight table, gating mapping table, noise scheduling table, and various thresholds are all generated by the parameter calibration module before deployment. The parameter calibration module takes complete page samples, occluded page samples, multi-fragment cropped samples, and dark noise samples at the read end as inputs, performs normalization constraints on the weights, uses the quantile search result with the lowest bit error rate on the validation set for the thresholds, and uses a segmented monotonic mapping of no less than 8 physical residual bins for the gating mapping table. Among them, the pixel confidence threshold is searched in the range of 0.45 to 0.80, the effective frequency domain threshold is searched in the range of 0.20 to 0.70, the gating coefficient and diffusion prior coefficient are both cropped to the range of zero to one, and the table version number, sample batch, and update time are written into the model configuration file.

[0124] After the completion iteration is completed, the page format constraint module divides the final restored page image into symbol units according to the holographic page layout parameters, reads the symbol probability matrix and reliability matrix to perform symbol decision, interleaving restoration and error correction verification. If the error correction verification passes, the decoded bit stream is reassembled into the original file data according to the file fragment number and length field in the page format header. If the error correction verification fails, the positions of symbols with reliability below the threshold and the final physical residual are retained as verification information and a low-confidence restoration flag is output.

[0125] After the last completion iteration, the result verification unit reads the page header, data area and verification area according to the page format constraints. First, it selects the highest probability legal candidate for each symbol unit according to the symbol probability matrix. Then, it judges whether the page-level verification relationship is satisfied through the error correction verification syndrome. When multiple candidate combinations pass the error correction verification, the combination with the highest average comprehensive reliability is selected. If the average comprehensive reliability is still tied, the candidate combination with the smallest physical residual value is selected and the selection rule is written into the result record.

[0126] When restoring the original archive data, the system reassembles the bitstream according to the archive identifier, fragment number, fragment length and check field in the page format header, and performs hash check or length check on the reassembled archive data. If the check passes, the system outputs the original archive data and the restored page image. If the check fails, the system outputs the low confidence restoration flag, the set of code elements to be fed back, the difficult-to-recover code element flag and the final physical residual, so that it can be used for subsequent reading tasks or manual review.

[0127] After S5 is completed, the data restoration system writes the final restored page image, symbol probability matrix, reliability matrix, decoding feedback diagram, error correction verification results, and the restored original archive data into the result record as the output of this image reconstruction-based holographic data restoration method.

[0128] 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.

[0129] This invention is based on incomplete diffraction patterns, low signal-to-noise ratio page images, synchronization marker information, and holographic page layout parameters obtained at the reading end. It connects page coordinate registration, spatial domain reconstruction, frequency domain reconstruction, holographic forward propagation constraints, and diffusion-type completion into a continuous processing chain, so that the incomplete page data forms a transitive constraint relationship between the observation confidence region, the physical consistency region, and the region to be completed, thereby solving the problem that fragmented media are difficult to recover the original archive content.

[0130] In the improved technical solution, frequency domain gating driven by physical residuals and diffusion prior weights are used to adjust the reconstruction emphasis of different regions. The symbol decodability feedback mechanism maps the set of low-reliability symbols into a decoding feedback map and inputs it into the next round of completion. This makes the completion process not only pursue the continuity of the page image structure, but also take into account the consistency of holographic propagation and the requirements of error correction decoding, making it more suitable for holographic data restoration in fragment reading, partial occlusion and complex disturbance scenarios.

Claims

1. A holographic data retrieval method based on image reconstruction, characterized in that, include: S1. Acquire the data to be restored output from the reading end, perform registration, normalization, and page region localization on the data to be restored, and obtain the page observation image, defect mask, and pixel confidence map. The data to be restored includes the measured diffraction pattern, page image, synchronization marker information, and holographic page layout parameters; S2. Extract spatial reconstruction features from the page observation image, defect mask, and pixel confidence map respectively, and extract frequency domain reconstruction features from the frequency domain representation of the measured diffraction pattern and page observation image; S3. Based on the defect mask and pixel location... The confidence map generates cross-domain attention weights, which are then fused with spatial reconstruction features and frequency domain reconstruction features to obtain a coarse restored page image. S4: The coarse restored page image is converted into a simulated diffraction pattern using a holographic forward propagation model. The physical residual between the simulated diffraction pattern and the measured diffraction pattern is calculated within the confidence region. Frequency domain gating weights and diffusion prior weights are generated from the physical residuals. The confidence region is determined in the page plane by the defect mask and pixel confidence threshold, and is mapped to the diffraction plane confidence weights by the holographic forward propagation model. S5. Input the coarse restored page image, defect mask, pixel confidence map, frequency domain gating weight and diffusion prior weight into the diffusion-type completion module to complete the defective area and the area where the pixel confidence is less than the pixel confidence threshold. During the completion process, a decoding feedback map is generated based on the symbol probability matrix and reliability matrix. After the completion is completed, the original file data is restored based on the page format constraints and error correction verification.

2. The holographic data restoration method based on image reconstruction according to claim 1, characterized in that, S1 includes: The page coordinate system is determined using the synchronization marker information and holographic page layout parameters. The page image is mapped to the page coordinate system, and the measured diffraction pattern is aligned with the geometric scale and / or frequency domain of the page coordinate system to generate a diffraction alignment record. The mapped page image is normalized to grayscale and the background is subtracted to obtain the page observation image, wherein 1 in the defect mask represents a defect and 0 represents no defect; The defect mask is generated based on the page boundary, the occlusion position of the synchronization mark, and the pixel position with no effective light intensity response; The confidence level of each pixel is calculated according to a preset weight based on the normalized light intensity deviation, local noise variance, synchronization mark registration residual, and defect mask value, and the calculation results are normalized to a numerical range of zero to one to form the pixel confidence map.

3. The holographic data restoration method based on image reconstruction according to claim 1, characterized in that, The spatial reconstruction features in step S2 are obtained as follows: the page observation image is sampled using a symbol grid, and the symbol candidate boundaries are determined based on the holographic page layout parameters; the synchronization marker geometric features, symbol edge features, and local texture features are extracted within the observation area defined by the defective mask. Using the pixel confidence map as sampling weights, the symbol candidate boundary, synchronization marker geometric features, symbol edge features, and local texture features are weighted and aggregated to obtain the spatial reconstruction features aligned with the page coordinate system.

4. The image reconstruction based holographic data retrieval method of claim 1, wherein, The frequency domain reconstruction features in step S2 are obtained as follows: Fourier transforms are performed on the measured diffraction pattern and the page observation image to obtain the diffraction spectrum and the page spectrum; diffraction fringe features, edge component features with frequencies greater than the edge frequency band threshold, and noise distribution features are extracted according to the frequency band division rules determined by the holographic page layout parameters. The effective observation mask obtained from the defective mask is converted to the frequency domain to obtain the effective frequency domain coefficients. The effective frequency domain coefficients are then used to weight the diffraction fringe features, edge component features, and noise distribution features to obtain the frequency domain reconstructed features.

5. The image reconstruction based holographic data retrieval method of claim 1, wherein, S3 includes: calculating spatial preservation weights based on the defect mask and pixel confidence map, wherein the spatial preservation weights take a first preset weight at positions where the defect mask represents an undefected pixel and the pixel confidence is not less than the pixel confidence threshold, and take a second preset weight at other positions, with the first preset weight being greater than the second preset weight; calculating frequency domain compensation weights based on the proportion of pixels represented as defective by the defect mask and the frequency domain effective coefficients obtained by converting the effective observation mask obtained from the defect mask; inputting the spatial preservation weights, frequency domain compensation weights, spatial reconstruction features, and frequency domain reconstruction features into a cross-domain attention module to obtain the cross-domain attention weights; using the cross-domain attention weights to stitch together, channel-weighted, and page coordinate resampling of the spatial reconstruction features and frequency domain reconstruction features to obtain fused features, and decoding the page image of the fused features to obtain the coarse restored page image.

6. The image reconstruction based holographic data retrieval method of claim 1, wherein, S4 includes: establishing the holographic forward propagation model based on the recording wavelength, propagation distance, pixel spacing, and holographic page layout parameters; using the coarse restored page image as the amplitude constraint of the object plane complex amplitude, and calculating the simulated diffraction pattern through phase modulation and diffraction propagation; representing the defective mask as an undefective page plane region with a pixel confidence level not less than the pixel confidence level threshold to generate page plane confidence weights, and mapping the diffraction plane confidence weights to the holographic forward propagation model; calculating the light intensity residual, frequency band energy residual, and synchronization marker position residual within the weighted region defined by the diffraction plane confidence weights; normalizing the light intensity residual, frequency band energy residual, and synchronization marker position residual and weighting them according to preset coefficients to obtain the physical residual; and looking up the gating mapping table based on the physical residual to generate the frequency domain gating weight and diffusion prior weight.

7. The image reconstruction based holographic data retrieval method of claim 5, wherein, The diffusion-based completion module in step S5 includes a conditional encoder, a noise prediction network, and a page image decoder; the conditional encoder encodes the coarse restored page image, the defect mask, the pixel confidence map, the frequency domain gating weights, and the diffusion prior weights to obtain completion conditional features; In each completion iteration, the noise prediction network predicts noise components based on the current noisy page image, completion condition features, and decoding feedback map. The page image decoder updates the current noisy page image based on the predicted noise components and retains the update results in the completion region. At positions where the missing mask represents an unmissing image and the pixel confidence is not less than the pixel confidence threshold, the page observation image is retained according to the spatial retention weight.

8. The holographic data restoration method based on image reconstruction according to claim 7, characterized in that, The decoding feedback graph is generated as follows: the current restored page image after each preset number of completion iterations is divided into symbol units according to the holographic page layout parameters; the probability of each candidate symbol value is calculated for each symbol unit to form the symbol probability matrix; the candidate symbol values ​​of the same symbol unit are sorted from high to low probability, and the difference between the first and second candidate probabilities is used as the corresponding reliability, and the reliability matrix is ​​generated by combining the error correction verification syndrome; symbols with reliability less than the symbol reliability threshold and symbols that failed the error correction verification are identified as the set of symbols to be fed back; the set of symbols to be fed back is rasterized according to the position of the symbol units in the page coordinate system to obtain the decoding feedback graph.

9. The image reconstruction based holographic data retrieval method of claim 8, wherein, The spatial reconstruction branch, frequency domain reconstruction branch, cross-domain attention module, and diffusion completion module are obtained through sample training. The training samples include complete page images, sample diffraction patterns generated from complete page images through a holographic forward propagation model, sample page observation images obtained by applying occlusion and noise to complete page images, sample defect masks, and sample pixel confidence maps; the training loss includes page image reconstruction loss, diffraction pattern consistency loss, symbol cross-entropy loss, and decoding feedback region weighted loss, wherein the decoding feedback region weighted loss applies a greater loss weight to the page coordinate region corresponding to the set of symbols to be fed back than to the non-feedback region.

10. The holographic data restoration method based on image reconstruction according to claim 8, characterized in that, When the data to be restored comes from multiple media fragments of the same holographic page, steps S1 to S4 are performed on each media fragment to obtain the coarse restored page image, physical residual, frequency domain gating weight and diffusion prior weight corresponding to each media fragment; and page coordinates are stitched together for each media fragment based on the consistency of the synchronization mark position residual and the overlapping area of ​​the page coordinate system. Within the stitched page coordinate system, if there are multiple media fragment observation results for the same pixel position or the same symbol neighborhood, the pixel or symbol neighborhood corresponding to the media fragment with the smallest local physical residual value is selected as the observation retention source based on the local physical residual map. A stitched defect mask and pixel confidence map are generated, and the area not covered by the observed retention pixels and the area where the pixel confidence after stitching is less than the pixel confidence threshold are used as the completion area and input into the diffusion completion module.

Images