Pixel-level instrument matrix calibration method for four-channel metasurface polarizing elements

Abstract

The application discloses a pixel-level instrument matrix calibration method of a four-channel metasurface polarization element, and particularly relates to the technical field of metasurface polarization detection and edge computing, and comprises the following steps: a polarization state generator is controlled to output a plurality of groups of known standard polarization states of Stokes vectors in a fixed order; four-channel subgraphs are collected; dark field subtraction, flat field correction and initial alignment are performed to obtain a standard response group; on the edge computing side, a local position group is unfolded in each channel subgraph for each target position, and is combined item by item according to the relationship that the four-channel responses can constitute the same observation vector to obtain a candidate position group set; and on the edge computing side, candidate position group construction and established pixel determination are first performed on the four-channel responses, and then pixel-level instrument matrix fitting, checking and abnormal recalculation are performed based on the established pixels.

Description

Technical Field

[0001] This invention relates to the field of metasurface polarization detection and edge calculation technology, and more specifically, to a pixel-level instrument matrix calibration method for a four-channel metasurface polarization element. Background Technology

[0002] In the instrument matrix calibration of a four-channel metasurface polarization element, the focus of existing technologies is mainly on establishing the correspondence between the four-channel observed intensity and the Stokes vector. In engineering implementation, it is generally necessary to first use a polarization state generator to output multiple sets of standard polarization states of known Stokes vectors, then sequentially collect the four-channel response intensity of each pixel, and then perform dark field correction, flat field correction and sub-image registration on the collected results. Subsequently, the instrument matrix is ​​solved by pixel, and the calibration results are checked and corrected in combination with the condition number, crosstalk degree or test state reconstruction error. Taking a four-channel metasurface polarization system for measuring live corneal reflectance as an example, the calibration process must be completed directly on the near-end processing side, without relying on frequent disassembly and readjustment or long-term offline recalculation. At the same time, under the constraints of local computing power and continuous acquisition time, the four spatially separated sub-images must be mapped to directly usable pixel-level calibration results. However, under these conditions, the following phenomenon tends to occur in the field: although sub-map registration has been completed and instrument matrix has been obtained at some locations, and the resulting matrix is ​​numerically inverseable and no obvious abnormalities have been found in the local verification results, subsequent measurements will still repeatedly show sudden changes in local Stokes reconstruction results, discontinuous connection of polarization parameters at adjacent locations, and difficulty in stably coinciding with repeated calibration results of the same device. The reason for this is that the existing processing method assumes that the registered locations with the same number can be directly used as the same calibration object to participate in matrix solving, without first confirming whether the four-channel response actually corresponds to the same physical location. The technical problem to be solved by this application is: how to first identify the established pixels that correspond to the four-channel responses in the pixel-level instrument matrix calibration process of a four-channel metasurface polarization element under edge computing conditions, and then perform instrument matrix fitting based on the established pixels, so as to avoid the calibration results formed by misaligned observations from entering the subsequent measurement link. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a pixel-level instrument matrix calibration method for a four-channel metasurface polarization element. This method first constructs candidate position groups and determines established pixels on the edge computing side of the four-channel response, and then performs pixel-level instrument matrix fitting, verification, and anomaly recalculation based on the established pixels, thereby solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a pixel-level instrument matrix calibration method for a four-channel metasurface polarization element, comprising: S1. Control the polarization state generator to output multiple sets of known Stokes vector standard polarization states in a fixed order, acquire four-channel sub-images, and perform dark field subtraction, flat field correction and initial alignment to obtain standard response sets; S2. On the edge computing side, expand the local position group for each target position in each channel sub-map, and combine them item by item according to the relationship that the four channel responses can form the same observation vector to obtain the candidate position group set; S3. Solve the instrument matrix for each candidate position group under all standard polarization states, and compare the position correspondence order, matrix row order and reconstruction direction between each standard polarization state. Retain the candidate position groups with unchanged position combination relationship and consistent reconstruction direction as the established pixel groups to obtain the established pixel set. S4. On the edge computing side, perform pixel-level instrument matrix fitting on the four-channel response intensity of each established pixel group under all standard polarization states, and calculate the condition number, crosstalk amount and test state reconstruction error to obtain the pixel-level instrument matrix set and the abnormal pixel set. S5. Write the pixel-level instrument matrix set into the calibration data table at the corresponding pixel position and generate a calibration identifier. Bind the calibration identifier to the subsequent measurement data. Repeat the pixel group determination and instrument matrix fitting for the corresponding positions of abnormal pixel sets, and output the pixel-level calibration results.

[0005] In a preferred embodiment, S1 includes: S11. Control the polarization state generator to output each standard polarization state in a predetermined order, and simultaneously acquire four-channel sub-images, dark field images and uniform illumination flat field images when each standard polarization state is output. Establish a correspondence relationship according to the acquisition time marker and the standard polarization state marker to obtain the original response group. S12. Perform dark field subtraction and flat field correction on each of the four-channel sub-images in the original response group according to the pixel ratio of the corresponding flat field image to obtain the channel correction group; S13. Based on the reflection center position and boundary expansion direction in the four-channel sub-graph under the same standard polarization state, perform inter-channel position mapping and initial alignment on the channel correction group to obtain the standard response group corresponding to each standard polarization state.

[0006] In a preferred embodiment, S2 includes: S21. On the edge computing side, perform fixed-length expansion around the same-numbered positions and adjacent positions of each target position in each channel subgraph to form local position groups corresponding to each channel, and complete the position numbering according to the fixed channel order, and output the local position group table. S22. Perform item-by-item pairing for different channel local locations in the local location group table, construct candidate location groups corresponding to each target location, and generate candidate cost groups based on channel location difference, local response difference and adjacent direction difference, and output candidate location group set.

[0007] In a preferred embodiment, S2 further includes: S23. Perform multi-view consistency verification within the same target location on the candidate position set, retain the candidate position set where the channel position difference is in a fixed order and does not reverse, the local response difference is transmitted in the same direction between each standard polarization state and the adjacent direction difference is continuously inherited, delete the candidate position set with cross conflict, and output the consistent candidate set. S24. Sort the consistent candidate set by candidate cost group, and perform forward verification and reverse verification on the first candidate position group of the sorted set and the first candidate position group of the adjacent target positions. Keep the candidate position groups that are consistent in the verification as the candidate position results of the corresponding target positions, and output the candidate position set.

[0008] In a preferred embodiment, S3 includes: S31. Perform joint solution on the four-channel response sequence and corresponding standard Stokes sequence of each candidate position group under all standard polarization states. Calculate the back substitution residual sequence, matrix row order sequence and position corresponding order sequence of each candidate position group respectively. Then, accumulate the back substitution residual sequence according to the standard polarization state order, count the number of shifts in the matrix row order sequence item by item, and count the number of jumps in the position corresponding order sequence item by item to obtain the initial matrix group, back substitution result group and order record group. S32. Based on the initial matrix group, the back-substitution result group, and the sequence record group, construct the receiving edge according to the adjacent order of the standard polarization state. When the corresponding position order of two adjacent standard polarization states is the same, the matrix row order is the same, and the positive and negative directions of the Stokes components obtained by back-substitution are consistent, retain the corresponding receiving edge. When the corresponding position order is different, the matrix row order is different, or the positive and negative directions of the Stokes components are opposite, delete the corresponding receiving edge to obtain the candidate receiving chain group.

[0009] In a preferred embodiment, S3 further includes: S33. Perform forward back substitution check and reverse reconstruction check on the candidate receiving chain group. Compare the matrix row correspondence results obtained by the forward check with the position correspondence results obtained by the reverse check according to the standard polarization state order. Delete receiving edges that have channel back pointing, row order reversal or reconstruction direction reversal. Reconstruct the candidate receiving chain group after each round of deletion until the receiving edge set obtained by two consecutive rounds of reconstruction is completely the same, and obtain a stable receiving chain group. S34. The candidate position groups in the stable receiving chain group that simultaneously satisfy the conditions of covering all standard polarization states, having the same position correspondence order between the first state and the last state, and having the same matrix row order between the first state and the last state are determined as the established pixel groups, and written into the established pixel table according to the target position to obtain the established pixel set.

[0010] In a preferred embodiment, S4 includes: S41. On the edge calculation side, extract the four-channel response intensity of each pixel group under all standard polarization states in a fixed order according to the standard polarization state, and arrange the four-channel response intensity into a response matrix according to the channel order. Arrange the corresponding standard Stokes vectors into a standard matrix in the same order to obtain the pixel solution group. S42. Based on the pixel solution group, perform generalized inverse solution and back-substitution recombination on each response matrix and the corresponding standard matrix to obtain the full-state instrument matrix, the full-state back-substitution result, and the instrument matrix group with one left after being eliminated in turn according to a single standard polarization state for each established pixel group.

[0011] In a preferred embodiment, S4 further includes: S43. For each full-state instrument matrix and leave-one instrument matrix group, calculate the singular value order and use the ratio of the first singular value to the last singular value as the condition number. Calculate the matrix inner product after row normalization and use the sum of the absolute values ​​of the inner products of non-corresponding rows as the crosstalk quantity. Calculate the sum of the absolute values ​​of the component differences between the test state reconstruction result and the corresponding standard Stokes vector as the test state reconstruction error to obtain the matrix index group. S44. Perform item-by-item verification on each matrix index group. When the last singular value of the global instrument matrix is ​​zero, the row order of the left-one instrument matrix group is different from that of the global instrument matrix, the sum of the absolute values ​​of the inner products of non-corresponding rows is reversed between the left-one instrument matrix group and the global instrument matrix, or the positive and negative directions of the components of the test state reconstruction result are different from the positive and negative directions of the components of the corresponding standard Stokes vector, write the corresponding established pixel group into the abnormal pixel set, and write the global instrument matrix corresponding to the remaining established pixel groups into the pixel-level instrument matrix set.

[0012] In a preferred embodiment, S5 includes: S51. Write the pixel-level instrument matrix set into the calibration data table in the order of pixel positions, and form a calibration load according to the pixel position, matrix row order, matrix element arrangement order, corresponding standard polarization state order and generation round. Perform fixed-length encoding and cyclic redundancy check calculation on the calibration load to obtain the calibration identifier group corresponding to each pixel position. S52. For each subsequent frame of measurement data, retrieve the corresponding calibration identifier from the calibration data table according to the pixel position, write the calibration identifier, measurement time identifier, channel order and pixel position together into the measurement binding table, and point the calibration identifier in the measurement binding table back to the calibration load in the calibration data table to obtain the measurement binding result.

[0013] In a preferred embodiment, S5 further includes: S53. For each abnormal location in the abnormal pixel set, the original candidate location group, the corresponding order of the adjacent established pixel group, and the matrix row order of the abnormal location in each channel sub-image are used as recalculation inputs. The established pixel group determination and instrument matrix fitting are repeatedly executed. The recalculated matrix row order, reconstruction direction, and calibration load are compared with the original binding record in the measurement binding result to obtain the abnormal recalculation result. S54. Based on the abnormal recalculation results, when the recalculated matrix row order is the same as the matrix row order of the adjacent pixel group, the reconstruction direction is the same as the reconstruction direction of the original bound record, and the cyclic redundancy check result is consistent, the abnormal record in the calibration data table is replaced with the recalculated calibration load and the corresponding calibration identifier is updated; when the matrix row order is different, the reconstruction direction is different, or the cyclic redundancy check result is inconsistent, the corresponding abnormal position is written into the failure position table, and the pixel-level calibration result is output.

[0014] The technical effects and advantages of this invention are as follows: 1. This scheme first constructs candidate location groups and determines the established pixels, and then performs instrument matrix fitting on the established pixels. This can exclude misaligned observations before solving the problem, and relatively reduce the abrupt changes in local reconstruction and spatial discontinuities caused by splicing from different physical locations. 2. By comparing the position correspondence order, matrix row order and reconstruction direction under the standard polarization state full sequence, and retaining the candidate position group with continuous consistency through the reconstructing of the successor chain, the stability of pixel-level calibration objects and cross-cycle consistency can be relatively improved. 3. Perform full-state solution and leave-one-out solution simultaneously on the established pixel group, and perform cross-checking by combining the condition number, crosstalk amount and test state reconstruction error. This can relatively suppress ill-conditioned matrices that are only valid in a single solution from entering the formal calibration results. 4. Perform dark field subtraction, flat field correction and initial alignment on the four-channel sub-image first, and then expand the local position group and combine the candidate position group under a unified reference. This can relatively reduce the interference of illumination non-uniformity, dark current bias and channel position offset on subsequent matrix solving. 5. By writing the pixel-level instrument matrix into the calibration data table, generating calibration identifiers and establishing retracement binding with subsequent measurement data, and performing local recalculation and replacement judgment on abnormal pixels, the traceability of calibration results and online maintenance capabilities can be relatively improved. Attached Figure Description

[0015] Figure 1 This is a flowchart outlining the method steps of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Refer to the instruction manual appendix Figure 1 The pixel-level instrument matrix calibration method for a four-channel metasurface polarization element of the present invention includes: S1. Control the polarization state generator to output multiple sets of known Stokes vector standard polarization states in a fixed order, acquire four-channel sub-images, and perform dark field subtraction, flat field correction and initial alignment to obtain standard response sets; This implementation method is used to uniformly collect, numerically correct, and spatially align the raw acquisition results of a four-channel metasurface polarization element in pixel-level instrument matrix calibration. This ensures that subsequent local position group expansion and candidate position group construction are based on the four-channel response under the same standard polarization state, the same acquisition cycle, and the same spatial reference. The principle is to first establish a one-to-one correspondence between the standard polarization state and the acquisition results, then eliminate numerical deviations caused by dark current, uneven illumination, and dead pixels, and finally map the four-channel sub-images to a unified position reference, forming a standard response group that can be directly read by subsequent steps. This implementation process includes the following steps: The polarization state generator outputs each standard polarization state sequentially according to the output sequence number in the standard polarization state table. The standard polarization state table includes at least the standard polarization state identifier, Stokes four-component values, output sequence number, and calibration round identifier. Each output of a set of standard polarization states triggers a synchronous exposure, obtaining four-channel sub-images for the same exposure. Before the acquisition of that set of standard polarization states begins, the dark field image of the corresponding channel is read. After the acquisition of that round of standard polarization states is completed, the uniform illumination flat field image is read. The four-channel sub-images, dark field image, and flat field image are written into the original response table, respectively. Subsequently, a joint correspondence is established between the acquisition time identifier, exposure round identifier, and standard polarization state identifier to form an original response record. Each original response record includes at least the standard polarization state identifier, acquisition time identifier, exposure round identifier, four-channel sub-image data, four-channel dark field image data, and flat field image data. When a complete four-channel sub-image is not obtained in the same output, or when the dark field image or flat field image is missing, the record corresponding to the standard polarization state identifier is written into the missing record table, and this record is prohibited from entering the subsequent correction process. Read each original response record in the original response table, and perform dark field subtraction on each of the four-channel sub-images under each standard polarization state in a fixed channel order. The subtraction method is to subtract the dark field gray value of the same position in the dark field image of the same channel from the original gray value of each pixel in the current channel sub-image to obtain the dark field subtracted image. Then, perform flat field correction on the dark field subtracted image. The correction method is to divide the dark field subtraction value at the position by the flat field correction value after the corresponding flat field image is normalized by the channel mean to obtain the corrected gray value. Write the four-channel corrected sub-image into the channel correction table. For positions with a flat field correction value of zero, mark the position as a zero value position and replace it with the median gray value of the adjacent valid position in the same channel. For consecutive bad pixel locations, the bidirectional average value is used to fill in the gaps when there are valid neighboring pixels in both the row and column directions. When there are only valid neighboring pixels in one direction, the average value in that direction is used to fill in the gaps. When there are no valid neighboring pixels in either direction, the bad pixel identifier is retained and written into the abnormal location table. After the correction is completed, all corrected grayscale values ​​are restricted to the effective quantization range of the imaging sensor. Values ​​exceeding the upper limit are written as the upper limit value, and values ​​below the lower limit are written as the lower limit value, so as to obtain the channel correction group. The channel correction group under the same standard polarization state is retrieved. The correction sub-image of the fixed reference channel is used as the position reference image. The reflection center position and boundary position are extracted from the position reference image. The reflection center position is taken as the gray-weighted center coordinate of the reflection kernel of the channel. The boundary position is taken as the set of positions where the gray level outside the reflection kernel continuously decreases radially and first enters the background stable region. The boundary unfolding direction is generated according to the radial line connecting the reflection center position to each boundary position and written into the position mapping table. Then, the other three channel correction sub-images are respectively mapped to the position reference image. The mapping method is to calculate the radial distance and azimuth angle from each position to be mapped to the reflection center position, and select the pixel position with the smallest radial distance difference and azimuth angle difference in the target channel as the corresponding mapping position to obtain the mapping position pair from each channel to the reference channel. After mapping is completed, the four-channel sub-graphs are uniformly written back to the reference channel coordinate system to form a standard response record corresponding to the current standard polarization state. The standard response records of all standard polarization states are summarized and written into the standard response group. When a channel does not have a mappable position in the boundary region, the position is recorded as a boundary overrun position and is not included in the writing of the standard response record of the current standard polarization state. At the same time, the overrun reason code is written into the position mapping table for removal during subsequent target position filtering. Through the above processing, a fixed correspondence chain is formed between the standard polarization state, the acquisition result, the correction result and the alignment result. When reading the standard response group in subsequent steps, the four-channel response can be directly extracted according to the standard polarization state order and the unified coordinate position, avoiding the distortion of the candidate position group caused by the chaotic acquisition attribution, the inconsistency between the dark field and flat field correction caliber or the inconsistency between the position references of the channels. In practical applications: When the polarization state generator outputs four sets of linear polarization states and one set of circular polarization states in sequence, the edge computing side can simultaneously obtain four-channel sub-images in the single exposure corresponding to each set of outputs, and attach the corresponding dark field image and flat field image to the exposure record respectively; then perform pixel-by-pixel dark field subtraction and flat field correction on the four-channel sub-images, and then use the reflection center position of the first channel sub-image as the reference center to map the other three channels to the same coordinate system, finally obtaining the standard response groups corresponding to the five standard polarization states, which can be directly called for subsequent local position group expansion and candidate position group construction.

[0018] S2. On the edge computing side, expand the local position group for each target position in each channel sub-map, and combine them item by item according to the relationship that the four channel responses can form the same observation vector to obtain the candidate position group set; This implementation method is used to form a set of candidate positions for each target position based on a standard response group, which can be used to determine the establishment of a pixel group. The processing logic is as follows: first, the local positions of the four channels are expanded under a unified coordinate reference, then the four-channel combination is constructed according to a fixed channel order, and then consistency verification and adjacent verification are performed to output the candidate position results. This implementation process includes the following steps: Local position groups are expanded around each target location to provide a unified input for the four-channel combination. The input includes a standard response group, a position mapping table, and a target position table. The target positions are taken from the valid positions of the standard response map of the reference channel. Boundary out-of-bounds positions and bad pixel positions are not included in the target position table. The processing actions are as follows: determine the same-signature position of the target position in the reference channel, and determine the corresponding same-signature position in the other three channels according to the position mapping table; perform adjacent expansion with the same-signature position of each channel as the center and according to the preset expansion length. The expansion length is given by the preset configuration, and the expansion order is fixed as center, top, right, bottom, left, upper left, upper right, lower right, lower left; the expanded positions are numbered according to the fixed channel order and local order; the output is a local position group table, which must at least write the target position identifier, channel identifier, local position number, local position coordinates, and relative displacement for subsequent item-by-item pairing and reading; the exception handling is as follows: when a local position goes out of bounds, write the out-of-bounds reason code and delete the position; when the number of valid local positions in any channel is less than the preset expansion number, write the corresponding target position into the local missing table; Candidate position groups are constructed based on the local position group table to compress the four-channel local positions into verifiable objects. The inputs are the local position group table, the standard response group, and the fixed channel order. The processing is as follows: for each target position, one local position is selected from each of the four channel local position groups in the fixed channel order to form a four-channel combination. All available combinations are traversed to obtain candidate position groups. For each candidate position group, a candidate cost group is calculated, where the channel position difference is the sum of the absolute values ​​of the coordinate differences between the local positions of each non-reference channel and the local position of the reference channel, the local response difference is the sum of the absolute values ​​of the differences between the corrected grayscale values ​​of each non-reference channel and the corrected grayscale value of the reference channel, and the adjacent direction difference is the sum of the absolute values ​​of the azimuth angle differences between the local positions of two adjacent channels under the fixed channel order. The output is a candidate position group set, which must contain at least the target position identifier, the four-channel local position number string, the channel position difference, the local response difference, and the adjacent direction difference. The exception handling is as follows: if any channel local position is missing, out of bounds, or has an empty corrected grayscale value, the corresponding candidate position group is deleted and a deletion reason code is written. The system performs a consistency check on the candidate position set to delete combinations that do not satisfy the spatial continuity and standard polarization state transfer relationships. The inputs are the candidate position set, the standard response set, and the standard polarization state table. The processing steps are as follows: For each candidate position set, generate a channel position difference sequence, a local response difference sign sequence, and an adjacent direction difference sequence; perform an order check on the channel position difference sequence, requiring the arrangement order to be consistent with the fixed channel order; perform a transfer check on the local response difference sign sequence, requiring no sign bounce (positive, negative, positive or negative, positive, negative) according to the standard polarization state order, with zero values ​​being assigned to the previous non-zero sign; perform a continuity check on the adjacent direction difference sequence, requiring continuous continuity according to the fixed channel order; the output is a consistent candidate set; exception handling is as follows: When two candidate position sets within the same target location occupy the same channel and the same local location and their combination relationships intersect, retain the candidate position set whose candidate cost group lexicographical order comes first, and write the rest into the conflict deletion table. The system performs sequential sorting and adjacent verification on the consistent candidate group set to determine the candidate position results for each target location. The inputs are the consistent candidate group set and the target position table. The processing steps are as follows: First, the target positions are sorted. The sorting comparison order is fixed as channel position difference, local response difference, and adjacent direction difference. If all three are the same, the four-channel local position number string is retained in lexicographical order. The first candidate position group is selected and subjected to forward verification with the first candidate position groups of the left and upper neighbor target positions, and reverse verification with the first candidate position groups of the right and lower neighbor target positions. The comparison content is the channel position increment order and the sign propagation direction of the local response difference. If the forward verification result matches the reverse verification result, the first candidate position group is retained as the candidate position result for the corresponding target position. The output is the candidate position group set, which is read for pixel group determination. The exception handling is as follows: If the verification is inconsistent, the next candidate position group is replaced sequentially according to the sorting order and the verification is repeated. If all verifications are inconsistent, the corresponding target position is written to an empty candidate result table. Through the above processing, candidate position results with clear sources, fixed comparison rules, clear conflict resolution, and spatial continuity can be output for each target position. These results can then be directly used to establish pixel group determination. In practical applications: the edge computing side expands a 3x3 local position group for each target position within the effective area of ​​the reference channel in four channels, constructs a four-channel candidate position group, calculates three types of cost values, removes symbol backtracking and cross-conflict combinations, sorts them according to a fixed comparison order, and performs bidirectional verification in combination with adjacent target positions to finally obtain the candidate position results for the corresponding target position.

[0019] S3. Solve the instrument matrix for each candidate position group under all standard polarization states, and compare the position correspondence order, matrix row order and reconstruction direction between each standard polarization state. Retain the candidate position groups with unchanged position combination relationship and consistent reconstruction direction as the established pixel groups to obtain the established pixel set. This implementation method is used to identify which candidate position groups can maintain stable positional correspondences and stable matrix representations across all standard polarization states, based on the formation of candidate position groups, and outputs a set of valid pixels for subsequent pixel-level instrument matrix fitting. The processing principle is as follows: First, a joint solution relationship is established between the four-channel response sequence of each candidate position group under all standard polarization states and the corresponding standard Stokes sequence to obtain comparable matrix results and order results. Then, results that can be continuously connected between adjacent standard polarization states are linked into a connection chain. Subsequently, erroneous connection edges that are only valid in unidirectional solutions are deleted through bidirectional verification. Finally, candidate position groups covering all standard polarization states and maintaining consistency from the stable connection chain groups are extracted and written into the valid pixel table as valid pixel groups. This implementation process includes the following steps: A joint solution is performed on each candidate position group to convert them into subsequently verifiable matrix results and ordering results. The inputs are a set of candidate position groups, a standard response group, and a standard polarization state table. Each candidate position group contains a string of four-channel local position numbers in a fixed channel order, and the standard polarization state table contains a standard polarization state identifier, Stokes four-component values, and a standard polarization state sequence number. The processing action is as follows: for each candidate position group, the four-channel response intensities of that candidate position group under all standard polarization states are extracted sequentially according to the standard polarization state order to form a four-channel response. The Stokes sequence is constructed by arranging the four Stokes components corresponding to the standard polarization state in the same order. Then, a four-row, four-column joint solution system is built using the four-channel response sequence as the observation and the standard Stokes sequence as the known. The solution and back-substitution are performed for each standard polarization state to obtain the back-substitution results corresponding to each standard polarization state. For each standard polarization state, the sum of the absolute values ​​of the component differences between the back-substitution result and the corresponding standard Stokes four components is calculated as the back-substitution residual value of that standard polarization state, and arranged in the order of the standard polarization states to form a back-substitution residual sequence. A matrix row order sequence is formed based on the correspondence between the matrix rows obtained from each solution and the fixed channel order. A position-corresponding order sequence is formed based on the local position number strings of the four channels within the candidate position group. The back-substitution residual sequence is accumulated item by item according to the standard polarization state order. The number of row position changes between adjacent states in the matrix row order sequence is counted item by item. The number of position number changes between adjacent states in the position-corresponding order sequence is counted item by item. This yields an initial matrix group, a back-substitution result group, and an order record group, which are written into the initial matrix table, back-substitution result table, and order record table, respectively, for use in the construction of the connecting edge. Anomaly handling is as follows: when any candidate position group lacks four-channel response intensity under a certain standard polarization state, or when an empty matrix row appears in the solution system, the candidate position group is written into the solution failure table and does not enter the connecting edge construction process. A connection edge is constructed based on the initial matrix group, back-substitution result group, and order record group to retain the expression relationship of continuously valid candidate position groups on the standard polarization state order. The input quantities are the initial matrix table, back-substitution result group, order record table, and standard polarization state table. The processing action is as follows: for each candidate position group, the Stokes component symbols in the current position corresponding order, matrix row order, and back-substitution result are read pairwise according to the adjacent order of the standard polarization state. The Stokes component symbols are recorded as positive for positive, negative for negative, and zero for inheriting the nearest non-zero symbol of the corresponding component of the previous standard polarization state. If there is no non-zero value before the current component, it is recorded as zero. A connection judgment is constructed by taking two adjacent standard polarization states as a pair. When the corresponding position order, matrix row order, and symbols of the four Stokes components of the pair of standard polarization states are the same, the connection edge between the pair of standard polarization states is retained, and the previous state identifier, the next state identifier, the candidate position group identifier, the corresponding position order, the matrix row order, and the direction consistency result are written into the connection edge table. When the position correspondence order is different, the matrix row order is different, or the sign of any Stokes component is inconsistent, the connecting edge between the standard polarization states is deleted and a deletion reason code is written. After all adjacent standard polarization state pairs are processed, the connecting edges retained in the same candidate position group are connected in the order from the previous state to the next state to obtain the candidate connecting chain group, and written into the candidate connecting chain list for bidirectional verification. The exception handling is as follows: when a candidate position group does not form a connecting edge between any adjacent standard polarization state pairs, the candidate position group is written into the chain break record table and is not entered into subsequent verification. For each candidate receiving chain group, a forward-backward substitution check and a reverse recombination check are performed to delete receiving edges that are valid only under unidirectional receiving but not under reverse checking. The inputs are the candidate receiving chain list, the backward substitution representative, the order record table, and the receiving edge table. The processing actions are as follows: For each candidate receiving chain group, a forward-backward substitution check is performed item by item from the first state to the last state according to the standard polarization state order. The matrix row order and the backward substitution result corresponding to the current receiving edge are read, and a sequence of forward matrix row corresponding results is generated. At the same time, a reverse recombination check is performed item by item from the last state to the first state according to the standard polarization state reverse order. The sequence corresponding to the current position and the four-channel local position number string within the candidate position group are read, and a sequence of reverse position corresponding results is generated. Next, the result sequence corresponding to the forward matrix rows is compared item by item with the result sequence corresponding to the reverse positions according to the standard polarization state order. When channel back-indexing, row order reversal, or reconstruction direction reversal occurs at the same standard polarization state position, the connecting edge associated with that position is deleted. Channel back-indexing is defined as the local position number of any channel in the later standard polarization state returning to the local position number that was deleted in the previous standard polarization state. Row order reversal is defined as the matrix row order of two adjacent standard polarization states being reversed from the fixed channel order. Reversal of reconstruction direction is defined as the sign of the same Stokes component changing from positive to negative or from negative to positive in the preceding and following standard polarization states. After each round of deletion, the candidate connecting chain groups are reconnected according to the currently retained connecting edges to obtain a new set of connecting edges. When the edge identifier, previous state identifier, and subsequent state identifier in the connecting edge set after two consecutive rounds of reconstruction are completely identical, the reconstruction is stopped and the current result is written into the stable connecting chain list to obtain a stable connecting chain group. The exception handling is as follows: when a candidate connecting chain group is deleted to an empty chain during the reconstruction process, the corresponding candidate position group is written into the empty chain list and the subsequent verification of the candidate position group is stopped. The established pixel group is extracted from the stable connection chain group and used as the unique input object for subsequent pixel-level instrument matrix fitting. The input quantities are the stable connection chain list, the candidate position group set, and the target position table. The processing action is as follows: read the stable connection chain group item by item according to the target position, and search the coverage of each candidate position group corresponding to the same target position in the stable connection chain list. When the stable connection chain group corresponding to a certain candidate position group covers all standard polarization states at the same time, and the position correspondence order of the first state and the last state is the same, and the matrix row order of the first state and the last state is the same, the candidate position group is determined as the established pixel group. If multiple candidate position groups satisfy the conditions exist for the same target location, they are uniquely retained in lexicographical order according to the candidate position group identifier. The target position identifier, four-channel local position number string, position correspondence order of the first and last states, matrix row order of the first and last states, and corresponding stable succession chain group identifier of the retained candidate position group are written into the established pixel table. After all target locations are processed, the established pixel set is obtained and read by the subsequent pixel-level instrument matrix fitting step. The exception handling is as follows: when there is no candidate position group that satisfies the conditions of covering all standard polarization states and having consistent first and last states for the same target location, the target location is written into the established missing table and removed from the input of the subsequent pixel-level instrument matrix fitting. Through the above processing, erroneous combinations in the candidate position group that are valid only in local standard polarization states or only in unidirectional solutions can be deleted, and only valid pixel groups that maintain consistent position correspondence, matrix row relationship and reconstruction direction in all standard polarization states can be retained, thereby providing stable input for subsequent pixel-level instrument matrix fitting; In practical applications: For multiple candidate position groups formed at a target location, the edge computing side first extracts the corresponding four-channel response sequences according to the five standard polarization states and solves them jointly to obtain the initial matrix group, back-substitution result group, and order record group for each candidate position group; then, it constructs connecting edges according to the adjacent order of the five standard polarization states, retaining connecting edges with consistent position correspondence order, matrix row order, and Stokes component signs; subsequently, it performs forward back-substitution verification and reverse reconstruction verification on the candidate connecting chain groups, deleting connecting edges with channel back-pointing and row order reversal, until the reconstruction results of two consecutive rounds are consistent; finally, it retains candidate position groups from the stable connecting chain groups that cover the five standard polarization states and have consistent first and last position correspondence order and matrix row order, and writes them into the established pixel table as the established pixel group for the target location.

[0020] S4. On the edge computing side, perform pixel-level instrument matrix fitting on the four-channel response intensity of each established pixel group under all standard polarization states, and calculate the condition number, crosstalk amount and test state reconstruction error to obtain the pixel-level instrument matrix set and the abnormal pixel set. This implementation method, based on a determined set of established pixels, performs pixel-level instrument matrix solving, index calculation, and anomaly splitting on each established pixel group. It outputs a pixel-level instrument matrix set that can be directly entered into the calibration data table, as well as an anomaly pixel set that requires recalculation. The processing logic is as follows: First, the four-channel response intensities of each established pixel group under all standard polarization states are organized with the corresponding standard Stokes vectors into a uniformly arranged solution object. Then, mutually verifiable matrix results are formed through full-state solving and leave-one-out solving. Next, the condition number, crosstalk, and test-state reconstruction error are calculated. Finally, the established pixel groups are split into valid matrices and anomaly pixels according to fixed verification rules. This implementation process includes the following steps: The four-channel response intensities are extracted and a pixel solution group is constructed to convert the established pixel group into a unified matrix solution input. The input consists of an established pixel table, a standard response group, and a standard polarization state table. The standard polarization state table includes at least a standard polarization state identifier, a standard polarization state sequence number, and the corresponding Stokes four-component values. The established pixel table also includes at least a target position identifier and a four-channel local position number string. The processing action is as follows: On the edge calculation side, the corrected response intensities of the four-channel local positions of each established pixel group under all standard polarization states are read sequentially according to the standard polarization state order. The four-channel response intensities under the same standard polarization state are then processed in a fixed channel order. The intensity values ​​are arranged in one row, and all standard polarization states are arranged in multiple rows to form a response matrix. The Stokes four-component values ​​corresponding to each standard polarization state are read synchronously and arranged in the same standard polarization state order to form a standard matrix. Each established pixel group is written into a pixel solution table, which includes at least the target position identifier, established pixel group identifier, response matrix, standard matrix, and standard polarization state order string. The exception handling is as follows: when any established pixel group is missing four-channel response intensity under a certain standard polarization state, or when the number of rows in the response matrix is ​​not equal to the number of rows in the standard matrix, the established pixel group is written into the solution missing table and is not entered into the subsequent generalized inverse solution. The generalized inverse problem and back-substitution recombination are performed to obtain the full-state instrument matrix and the left-one-out instrument matrix group. The input is the pixel solution table. The processing steps are as follows: Perform the generalized inverse problem on the response matrix and standard matrix of each established pixel group to obtain the full-state instrument matrix; then substitute the full-state instrument matrix back into the standard matrix to obtain the full-state back-substitution result; then, in turn, remove one row of response data and one row of standard Stokes data corresponding to a single standard polarization state according to the standard polarization state order, and re-perform the generalized inverse problem on the remaining rows to obtain the left-one-out instrument matrix group corresponding to each removal position, and perform back-substitution on each left-one-out instrument matrix to obtain the left-one-out back-substitution result group; write the full-state instrument matrix, the full-state back-substitution result, the left-one-out instrument matrix group, the left-one-out back-substitution result group, and the corresponding removal standard polarization state identifier into the matrix result table; the exception handling is as follows: when the response matrix has fewer than four rows after removal, or the generalized inverse problem result is an empty matrix, write the corresponding established pixel group into the solution exception table and stop the subsequent index calculation of the established pixel group. The calculation matrix index group is used to convert the matrix results into verifiable fixed indexes. The inputs are a matrix result table and a test state table. The test state table includes a test state identifier, a test state sequence number, and four Stokes component values ​​for the test state. The test states are derived from independent test acquisition records from the same round as the standard polarization state acquisition. The processing actions are as follows: Singular value decomposition is performed on each full-state instrument matrix and each left-one instrument matrix. The singular values ​​are arranged in descending order to form a singular value order, and the ratio of the first singular value to the last singular value is used as the condition number. Euclidean norm normalization is performed on each full-state instrument matrix and each left-one instrument matrix row by row. All non-corresponding row pairs are enumerated, the absolute value of the inner product of each non-corresponding row vector is calculated, and the absolute values ​​of the inner products of all non-corresponding rows are added together as the crosstalk quantity. Back-substitution reconstruction is performed on the Stokes four-component input of each test state in the test state table into the global instrument matrix to obtain the test state reconstruction result. The sum of the absolute values ​​of the component differences between the test state reconstruction result and the corresponding test state Stokes four-component is calculated as the test state reconstruction error. The global condition number, the number of leave-one condition, the global crosstalk, the leave-one crosstalk, the test state reconstruction error, and the test state reconstruction result are written into the matrix index table. Anomaly handling is as follows: when the last singular value of any matrix is ​​zero, the condition number field of that matrix is ​​directly written as a zero denominator anomaly flag. When a test state is missing, the corresponding set pixel group is written into the test state missing table, and the test state reconstruction error calculation is not performed. Perform item-by-item verification and output pixel-level instrument matrix sets and abnormal pixel sets to split established pixel groups according to fixed rules; the inputs are matrix index table, matrix result table, and established pixel table; the processing actions are: read the full-state instrument matrix, the one-left-one instrument matrix group, and the matrix index group item by item according to the established pixel group; when the last singular value of the full-state instrument matrix is ​​zero, write the established pixel group into the abnormal pixel table; when the row order of any one-left-one instrument matrix is ​​different from the row order of the full-state instrument matrix, or when the full-state crosstalk quantity and each one-left-one crosstalk quantity are arranged in a fixed order string from small to large, and the one-left-one order string is inconsistent with the full-state order string, write the established pixel group into the abnormal pixel table; When the positive and negative directions of any Stokes component in the test state reconstruction result are different from the positive and negative directions of the corresponding four Stokes components in the test state, and the component is not zero, the established pixel group is written into the abnormal pixel table; the full-state instrument matrix corresponding to the other established pixel groups is written into the pixel-level instrument matrix table. The pixel-level instrument matrix table includes at least the target position identifier, the established pixel group identifier, the full-state instrument matrix, the matrix row order, and the corresponding standard polarization state order string; the abnormal handling is as follows: when the same established pixel group meets multiple abnormal conditions at the same time, only one abnormal record is retained, and the abnormal reason code is written in a fixed order according to the last singular value of zero, the matrix row order is different, the crosstalk order string is inconsistent, and the reconstruction direction is different. Through the above processing, the pixel positions in the established pixel group that are only temporarily established under the whole state solution but are inconsistent in leave-one mutual verification, crosstalk calculation or test state back substitution can be separated out, and only the whole state instrument matrix that is consistent with the whole state result and the leave-one result and has the same test state reconstruction direction can be retained and entered into the pixel-level instrument matrix set. In practical applications: The edge computing side extracts the four-channel response intensity of a certain established pixel group according to five standard polarization states and constructs a five-row, four-column response matrix, and simultaneously constructs a five-row, four-column standard matrix; generalized inverse is performed on the response matrix and the standard matrix to obtain the global instrument matrix, and then any one of the five standard polarization states is removed in turn to obtain five one-left-one instrument matrices; then the condition number and crosstalk are calculated for the global instrument matrix and the five one-left-one instrument matrices respectively, and back substitution is performed using independent test states to obtain the test state reconstruction error; if the last singular value of the global instrument matrix of the established pixel group is non-zero, the row order of the one-left-one matrix is ​​consistent with the row order of the global matrix, the crosstalk order is consistent, and the test state reconstruction direction is consistent, then its global instrument matrix is ​​written into the pixel-level instrument matrix table; otherwise, it is written into the abnormal pixel table.

[0021] Furthermore, in this embodiment, the standard polarization state is output by the polarization state generator according to the standard polarization state table. The Stokes four-component values ​​in the standard polarization state table are used to construct the standard matrix. The number of standard polarization states is not less than four, and there is no linear repetition relationship between the standard polarization states. The standard polarization state table is also configured with at least one set of independent test states. The independent test states do not participate in the full-state instrument matrix solution and the leave-one instrument matrix solution, but are only used for post-calibration error calculation. For each established pixel group, the edge computing side arranges its four-channel response intensity into a response matrix, arranges the Stokes four-component values ​​in the standard polarization state table into a standard matrix, and records the linear mapping result between the response matrix and the standard matrix as the instrument matrix A of the established pixel group. The row order of the instrument matrix A is bound to the fixed channel order, and the column order of the instrument matrix A is bound to the Stokes four-component order. Pixel-level calibration is performed with pixel groups as the smallest processing object. Each record in the pixel-level instrument matrix table only affects the four-channel response intensity at the corresponding target location. Channel-level calibration is performed with channel identifiers as the processing object. It is used to record dark field subtraction, flattening correction, inter-channel position mapping, and fixed channel order within the same channel. It does not replace the solution results of pixel-level instrument matrix A. When subsequent measurement data enters polarization reconstruction, dark field subtraction, flattening correction, and channel order reading are performed first according to the channel identifier. Then, the corresponding instrument matrix A and calibration identifier are retrieved from the calibration data table according to the pixel position. The four-channel measurement response at the same pixel position is input into instrument matrix A for Stokes reconstruction. Pixel positions that are not written to the calibration data table, written to the failure position table, or whose calibration identifier fails to point back are not included in the subsequent Stokes reconstruction result output. After each calibration cycle is completed, the edge computing side continuously reads the temperature record, assembly record, and test state reconstruction error record. When the difference between the current temperature in the temperature record and the calibration temperature of this cycle exceeds the preset temperature difference limit, or when any assembly mark appears in the assembly record indicating disassembly or assembly of metasurface polarization element, re-fixing of optical path, reinstallation of sensor, or re-alignment of polarization state generator, or when the independent test state reconstruction error is higher than the post-calibration error limit of this cycle, a recalibration trigger record is generated. After the recalibration trigger record is generated, the edge computing side re-executes the standard polarization state acquisition, candidate position group construction, pixel group determination, instrument matrix A fitting, abnormal pixel recalculation, and calibration flag update, and writes the test state reconstruction error, condition number, and crosstalk before and after recalibration into the post-calibration error table. The test state reconstruction error in the post-calibration error table does not participate in the solution of instrument matrix A, but is only used to determine whether the calibration result of this cycle is included in the calibration data table.

[0022] S5. Write the pixel-level instrument matrix set into the calibration data table at the corresponding pixel position and generate a calibration identifier. Bind the calibration identifier to the subsequent measurement data. Repeat the pixel group determination and instrument matrix fitting for the corresponding positions of abnormal pixel sets, and output the pixel-level calibration results. This implementation method is used to solidify pixel-level instrument matrix results into traceable calibration records, and after separating abnormal pixels from the normal release link, performs local recalculation and replacement judgment, finally outputting pixel-level calibration results that can be directly used for subsequent measurements. Its processing logic is as follows: first, the pixel-level instrument matrix set is written into the calibration data table and a retrievable calibration identifier is generated; then, the calibration identifier is bound to subsequent measurement data frame by frame; subsequently, the abnormal pixel positions are recalculated and checked against the original bound records item by item; finally, replacement updates or failure registration are performed according to the verification results. This implementation process includes the following steps: The pixel-level instrument matrix set is written into the calibration data table and a calibration identifier group is generated, which serves as the sole calibration basis for subsequent measurement calls. The input quantities are the pixel-level instrument matrix table and the current calibration round identifier. The pixel-level instrument matrix table includes at least the pixel position, the full-state instrument matrix, the matrix row order, and the corresponding standard polarization state order string. The processing actions are as follows: the pixel-level instrument matrix table is read item by item in the row and column order of the pixel position, and the corresponding full-state instrument matrix is ​​written into the calibration data table. Then, the calibration load is composed in a fixed order of pixel position, matrix row order, matrix element arrangement order, corresponding standard polarization state order, and generation round. The matrix element arrangement order adopts the fixed order of matrix expansion by row, and the generation round is taken from the current calibration round identifier. Subsequently, fixed-length encoding is performed on the calibration load. The bit width of the fixed-length encoding is given by a preset configuration, and the floating-point matrix elements are represented by fixed segments of sign bit, integer bit, and decimal bit. After encoding, cyclic redundancy check (CRC) calculation is performed on the encoding result to obtain the CRC value. The calibration load encoding result and the CRC value are combined as calibration identifiers and written into the calibration data table to form a calibration identifier group. The output consists of the calibration data table and the calibration identifier group, which are used for subsequent measurement binding and reading. Anomaly handling is as follows: when the calibration load at a certain pixel position is empty, the number of matrix elements is insufficient, or the length after fixed-length encoding does not meet the preset encoding length, the pixel position is written into the empty load record table, and no calibration identifier is generated. The calibration identifier is bound to the measurement data of each subsequent frame to establish a one-to-one correspondence between the measurement results and the calibration basis. The inputs are the calibration data table, the calibration identifier group, and the measurement data of each subsequent frame. The measurement data of each subsequent frame includes at least the measurement time identifier, channel order, and pixel position. The processing action is as follows: for each frame of measurement data, the corresponding calibration identifier in the calibration data table is retrieved item by item according to the pixel position, and the calibration identifier, measurement time identifier, channel order, and pixel position are written into the measurement binding table in a fixed field order. Then, using the calibration identifier and pixel position as a joint back pointer, the calibration identifier in the measurement binding table is back pointered to the corresponding calibration load in the calibration data table to form the measurement binding result. After all pixel positions of the measurement data in each frame are processed, the measurement binding result of that frame is written into the binding result table for abnormal recalculation and verification. The output is the measurement binding table and the measurement binding result. The abnormal handling is as follows: when a calibration identifier is not found at a certain pixel position, or when the same calibration identifier corresponds to multiple calibration loads at the same pixel position, the pixel position is written into the binding abnormal table, and the binding writing of that pixel position is stopped. The system recalculates the abnormal locations within the abnormal pixel set and generates abnormal recalculation results to determine whether abnormal records can be repaired through local recalculation. Inputs include an abnormal pixel table, candidate location sets, established pixel table, measurement binding results, standard response sets, and standard polarization state tables. The processing steps are as follows: For each abnormal location, the system reads the original candidate location sets in each channel sub-image and retrieves the adjacent established pixel sets from the established pixel table. Adjacent established pixel sets are selected according to a four-adjacency rule, in the order of left neighbor, right neighbor, top neighbor, and bottom neighbor. When multiple established pixel sets exist within a four-adjacency range, they are retained in a fixed order of left, top, right, and bottom. The positions and row order of the retained adjacent established pixel groups are read again as the recalculation constraint input for the abnormal position. The established pixel group judgment and instrument matrix fitting are repeated to obtain the recalculated matrix row order, reconstruction direction and calibration load. The reconstruction direction is taken as the positive and negative direction string of each Stokes component of the test state reconstruction result. The calibration load is still generated in a fixed order of pixel position, matrix row order, matrix element arrangement order, corresponding standard polarization state order and generation round. Then, the original binding record of the abnormal position is retrieved from the measurement binding result. The matrix row order, reconstruction direction and calibration load are compared item by item in a fixed comparison order to obtain the abnormal recalculation result and write it into the abnormal recalculation table. The abnormal handling is as follows: when there are no adjacent established pixel groups at the abnormal position, or the original candidate position group is empty, the abnormal position is directly written into the failure position table and does not enter the replacement judgment. Based on the abnormal recalculation results, perform replacement updates or failure registration to output the final pixel-level calibration results; the inputs are the abnormal recalculation table, calibration data table, and failure location table; the processing action is as follows: for each abnormal location in the abnormal recalculation table, read the recalculated matrix row order, reconstruction direction, and calibration load, and compare them item by item with the cyclic redundancy check results of the corresponding adjacent pixel group, the reconstruction direction of the original bound record, and the original calibration load. When the recalculated matrix row order is the same as the matrix row order of adjacent pixel groups, the reconstruction direction is the same as the reconstruction direction of the original binding record, and the cyclic redundancy check value recalculated by the recalculated calibration load is consistent with the cyclic redundancy check value of the original calibration load record, the abnormal record in the calibration data table is replaced with the recalculated calibration load, and the corresponding calibration identifier and the retracement relationship in the measurement binding table are updated synchronously; when the matrix row order is different, the reconstruction direction is different, or the cyclic redundancy check value is inconsistent, the abnormal position is written into the failure position table, which includes at least the pixel position, failure cause code, original calibration identifier, and calibration round; after all abnormal positions are processed, the updated calibration data table and the failure position table are written together into the pixel-level calibration result table; the abnormal handling is as follows: when the same abnormal position meets multiple inconsistencies at the same time, only one failure record is retained, and the failure cause code is written in a fixed order of different matrix row order, different reconstruction direction, and inconsistent cyclic redundancy check value; Through the above processing, the instrument matrix of normal pixel positions can be written into the calibration data table in a fixed encoding form and maintain a stable retracement relationship with subsequent measurement data. At the same time, abnormal pixel positions are recalculated locally before deciding whether to replace or fail, avoiding writing unverified recalculation results directly back to the calibration link. In practical applications: The edge computing side first writes the full-state instrument matrix into the calibration data table according to the pixel position order, and generates calibration loads and cyclic redundancy check values ​​to form calibration identifiers according to the fixed field order; for each subsequent frame of measurement data, the corresponding calibration identifier is retrieved according to the pixel position and written into the measurement binding table; when an abnormal position is detected, its original candidate position group and the established pixel groups of its left and upper neighbor positions are read as recalculation constraints, the established pixel group judgment and instrument matrix fitting are re-executed, and the row order, reconstruction direction and calibration load of the recalculated matrix are compared with the original binding record item by item; if the three comparison results are consistent, the abnormal record in the calibration data table is replaced and the calibration identifier is updated; if any item is inconsistent, the abnormal position is written into the failure position table, and finally the complete pixel-level calibration result is output.

[0023] The working principle of this scheme is as follows: First, the polarization state generator outputs multiple sets of known Stokes vector standard polarization states in a fixed order, simultaneously acquiring four-channel sub-images and completing dark field subtraction, flat field correction, and initial alignment to bring the raw data to a unified and comparable state. Then, around each target position, local position groups are expanded in the four channels to generate multiple candidate position groups, no longer directly assuming that positions with the same number necessarily correspond to the same physical position. Next, the candidate position groups are jointly solved under all standard polarization states. By comparing the position correspondence order, matrix row order, and reconstruction direction, positions in the preceding and following directions are selected. Candidate position groups that maintain consistency in the standard polarization state are identified as valid pixel groups. Based on this, pixel-level instrument matrix fitting is performed on these valid pixel groups, and cross-checks are conducted using full-state solutions, leave-one-out solutions, condition numbers, crosstalk, and test-state reconstruction errors. Reliable results are written into the pixel-level instrument matrix set, while unreliable results are classified into the abnormal pixel set. Finally, the reliable matrix is ​​written into the calibration data table according to pixel position, and calibration identifiers are generated and bound to subsequent measurement data frame by frame. Abnormal pixel positions are then locally recalculated and replaced to obtain pixel-level calibration results that can be directly used for subsequent polarization measurements. For example, in a live corneal reflectance measurement scenario, a four-channel metasurface polarization element divides the same reflected light into four spatially separated sub-images. Superficially, positions with the same number in the four sub-images seem to be directly pieced together into a four-channel observation. However, in reality, due to the influence of assembly and adjustment deviations, local distortion, thermal drift, and pixel inhomogeneity, the four positions with the same number may not actually correspond to the same physical point. The purpose of this solution is to first correct and align the four sub-images on the edge calculation side, and then find all the possible corresponding positions around each position. It compares each group to see which group can maintain the same positional and matrix relationships under multiple standard polarization states. Only such combinations are considered as real usable pixels for calibration. After doing this, the result is not a result that appears to be able to calculate the matrix, but a matrix result calculated after confirming that the correspondence is correct. During subsequent measurements, each pixel can call the pre-bound calibration mark, and abnormal positions can be recalculated or judged as invalid. Therefore, it is more suitable for rapid recall and stable measurement when the actual equipment is working continuously.

[0024] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A pixel-level instrument matrix calibration method for a four-channel metasurface polarization element, characterized in that, include: S1. Control the polarization state generator to output multiple sets of known Stokes vector standard polarization states in a fixed order, acquire four-channel sub-images, and perform dark field subtraction, flat field correction and initial alignment to obtain standard response sets; S2. On the edge computing side, expand the local position group for each target position in each channel sub-map, and combine them item by item according to the relationship that the four channel responses can form the same observation vector to obtain the candidate position group set; S3. Solve the instrument matrix for each candidate position group under all standard polarization states, and compare the position correspondence order, matrix row order and reconstruction direction between each standard polarization state. Retain the candidate position groups with unchanged position combination relationship and consistent reconstruction direction as the established pixel groups to obtain the established pixel set. S4. On the edge computing side, perform pixel-level instrument matrix fitting on the four-channel response intensity of each established pixel group under all standard polarization states, and calculate the condition number, crosstalk amount and test state reconstruction error to obtain the pixel-level instrument matrix set and the abnormal pixel set. S5. Write the pixel-level instrument matrix set into the calibration data table at the corresponding pixel position and generate a calibration identifier, then bind the calibration identifier to the subsequent measurement data. Repeat the pixel group determination and instrument matrix fitting for the corresponding positions of the abnormal pixel set, and output the pixel-level calibration results.

2. The pixel-level instrument matrix calibration method for a four-channel metasurface polarization element according to claim 1, characterized in that: S1 includes: S11. Control the polarization state generator to output each standard polarization state in a predetermined order, and simultaneously acquire four-channel sub-images, dark field images and uniform illumination flat field images when each standard polarization state is output. Establish a correspondence relationship according to the acquisition time marker and the standard polarization state marker to obtain the original response group. S12. Perform dark field subtraction and flat field correction on each of the four-channel sub-images in the original response group according to the pixel ratio of the corresponding flat field image to obtain the channel correction group; S13. Based on the reflection center position and boundary expansion direction in the four-channel sub-graph under the same standard polarization state, perform inter-channel position mapping and initial alignment on the channel correction group to obtain the standard response group corresponding to each standard polarization state.

3. The pixel-level instrument matrix calibration method for a four-channel metasurface polarization element according to claim 2, characterized in that: S2 includes: S21. On the edge computing side, perform fixed-length expansion around the same-numbered positions and adjacent positions of each target position in each channel subgraph to form local position groups corresponding to each channel, and complete the position numbering according to the fixed channel order, and output the local position group table. S22. Perform item-by-item pairing for different channel local locations in the local location group table, construct candidate location groups corresponding to each target location, and generate candidate cost groups based on channel location difference, local response difference and adjacent direction difference, and output candidate location group set.

4. The pixel-level instrument matrix calibration method for a four-channel metasurface polarization element according to claim 3, characterized in that: S2 also includes: S23. Perform multi-view consistency verification within the same target location on the candidate position set, retain the candidate position set where the channel position difference is in a fixed order and does not reverse, the local response difference is transmitted in the same direction between each standard polarization state and the adjacent direction difference is continuously inherited, delete the candidate position set with cross conflict, and output the consistent candidate set. S24. Sort the consistent candidate set by candidate cost group, and perform forward verification and reverse verification on the first candidate position group of the sorted set and the first candidate position group of the adjacent target positions. Keep the candidate position groups that are consistent in the verification as the candidate position results of the corresponding target positions, and output the candidate position set.

5. The pixel-level instrument matrix calibration method for a four-channel metasurface polarization element according to claim 4, characterized in that: S3 includes: S31. Perform joint solution on the four-channel response sequence and corresponding standard Stokes sequence of each candidate position group under all standard polarization states. Calculate the back substitution residual sequence, matrix row order sequence and position corresponding order sequence of each candidate position group respectively. Then, accumulate the back substitution residual sequence according to the standard polarization state order, count the number of shifts in the matrix row order sequence item by item, and count the number of jumps in the position corresponding order sequence item by item to obtain the initial matrix group, back substitution result group and order record group. S32. Based on the initial matrix group, the back-substitution result group, and the sequence record group, construct the receiving edge according to the adjacent order of the standard polarization state. When the corresponding position order of two adjacent standard polarization states is the same, the matrix row order is the same, and the positive and negative directions of the Stokes components obtained by back-substitution are consistent, retain the corresponding receiving edge. When the corresponding position order is different, the matrix row order is different, or the positive and negative directions of the Stokes components are opposite, delete the corresponding receiving edge to obtain the candidate receiving chain group.

6. The pixel-level instrument matrix calibration method for a four-channel metasurface polarization element according to claim 5, characterized in that: S3 also includes: S33. Perform forward back substitution check and reverse reconstruction check on the candidate receiving chain group. Compare the matrix row correspondence results obtained by the forward check with the position correspondence results obtained by the reverse check according to the standard polarization state order. Delete receiving edges that have channel back pointing, row order reversal or reconstruction direction reversal. Reconstruct the candidate receiving chain group after each round of deletion until the receiving edge set obtained by two consecutive rounds of reconstruction is completely the same, and obtain a stable receiving chain group. S34. The candidate position groups in the stable receiving chain group that simultaneously satisfy the conditions of covering all standard polarization states, having the same position correspondence order between the first state and the last state, and having the same matrix row order between the first state and the last state are determined as the established pixel groups, and written into the established pixel table according to the target position to obtain the established pixel set.

7. The pixel-level instrument matrix calibration method for a four-channel metasurface polarization element according to claim 6, characterized in that: S4 includes: S41. On the edge calculation side, extract the four-channel response intensity of each pixel group under all standard polarization states in a fixed order according to the standard polarization state, and arrange the four-channel response intensity into a response matrix according to the channel order. Arrange the corresponding standard Stokes vectors into a standard matrix in the same order to obtain the pixel solution group. S42. Based on the pixel solution group, perform generalized inverse solution and back-substitution recombination on each response matrix and the corresponding standard matrix to obtain the full-state instrument matrix, the full-state back-substitution result, and the instrument matrix group with one left after being eliminated in turn according to a single standard polarization state for each established pixel group.

8. The pixel-level instrument matrix calibration method for a four-channel metasurface polarization element according to claim 7, characterized in that: S4 also includes: S43. For each full-state instrument matrix and leave-one instrument matrix group, calculate the singular value order and use the ratio of the first singular value to the last singular value as the condition number. Calculate the matrix inner product after row normalization and use the sum of the absolute values ​​of the inner products of non-corresponding rows as the crosstalk quantity. Calculate the sum of the absolute values ​​of the component differences between the test state reconstruction result and the corresponding standard Stokes vector as the test state reconstruction error to obtain the matrix index group. S44. Perform item-by-item verification on each matrix index group. When the last singular value of the global instrument matrix is ​​zero, the row order of the left-one instrument matrix group is different from that of the global instrument matrix, the sum of the absolute values ​​of the inner products of non-corresponding rows is reversed between the left-one instrument matrix group and the global instrument matrix, or the positive and negative directions of the components of the test state reconstruction result are different from the positive and negative directions of the components of the corresponding standard Stokes vector, write the corresponding established pixel group into the abnormal pixel set, and write the global instrument matrix corresponding to the remaining established pixel groups into the pixel-level instrument matrix set.

9. The pixel-level instrument matrix calibration method for a four-channel metasurface polarization element according to claim 8, characterized in that: S5 includes: S51. Write the pixel-level instrument matrix set into the calibration data table in the order of pixel positions, and form a calibration load according to the pixel position, matrix row order, matrix element arrangement order, corresponding standard polarization state order and generation round. Perform fixed-length encoding and cyclic redundancy check calculation on the calibration load to obtain the calibration identifier group corresponding to each pixel position. S52. For each subsequent frame of measurement data, retrieve the corresponding calibration identifier from the calibration data table according to the pixel position, write the calibration identifier, measurement time identifier, channel order and pixel position together into the measurement binding table, and point the calibration identifier in the measurement binding table back to the calibration load in the calibration data table to obtain the measurement binding result.

10. The pixel-level instrument matrix calibration method for a four-channel metasurface polarization element according to claim 9, characterized in that: S5 also includes: S53. For each abnormal location in the abnormal pixel set, the original candidate location group, the corresponding order of the adjacent established pixel group, and the matrix row order of the abnormal location in each channel sub-image are used as recalculation inputs. The established pixel group determination and instrument matrix fitting are repeatedly executed. The recalculated matrix row order, reconstruction direction, and calibration load are compared with the original binding record in the measurement binding result to obtain the abnormal recalculation result. S54. Based on the abnormal recalculation results, when the recalculated matrix row order is the same as the matrix row order of the adjacent pixel group, the reconstruction direction is the same as the reconstruction direction of the original bound record, and the cyclic redundancy check result is consistent, the abnormal record in the calibration data table is replaced with the recalculated calibration load and the corresponding calibration identifier is updated; when the matrix row order is different, the reconstruction direction is different, or the cyclic redundancy check result is inconsistent, the corresponding abnormal position is written into the failure position table, and the pixel-level calibration result is output.

Images