Polarization difference and structure reconstruction-based mirror wafer two-dimensional code recognition method and system

Abstract

This invention discloses a method and system for recognizing QR codes on mirrored wafers based on polarization difference and structural reconstruction, belonging to the field of QR code recognition technology. The method includes the following steps: Step 1: Acquire parallel polarization and vertical polarization images of the wafer mirror surface with the same field of view, exposure, and frame rate using a single-channel grayscale. Step 2: Calculate a polarization difference image using the parallel and vertical polarization images, generate a reflective mask based on an adaptive threshold, and simultaneously calculate and preprocess the polarization fusion image to obtain the fused image. This invention achieves pixel-level accurate detection of highly reflective areas through parallel / vertical dual polarization image difference. Combined with an adaptive weighted polarization fusion strategy, it can fundamentally suppress strong mirror reflection interference from the wafer mirror surface, while significantly enhancing the texture contrast between QR code markings and the background. This solves the technical problem of traditional imaging methods where mirror reflection completely obscures QR code features, making recognition impossible.

Description

Technical Field

[0001] This invention relates to the field of QR code recognition technology, specifically to a method and system for QR code recognition on mirror wafers based on polarization difference and structural reconstruction. Background Technology

[0002] In the entire semiconductor manufacturing process, each wafer requires a unique identifier, typically a QR code directly laser-etched onto the wafer's surface. This code carries complete traceability information from crystal pulling, slicing, grinding to every step of the integrated circuit manufacturing process, serving as the cornerstone of intelligent manufacturing and data management. However, the wafer surface itself has extremely high reflectivity, much like a mirror, which presents unprecedented challenges to the reliable reading of QR codes. When traditional QR code recognition technology is applied in this scenario, its performance drops sharply due to the strong, dynamically changing reflections, resulting in low read rates and severely impacting the automation level of the production line and the integrity of data traceability.

[0003] Existing solutions mainly revolve around improving lighting, enhancing image processing algorithms, or combining deep learning. Among them, the most similar implementations include recognition methods based on multi-angle dynamic lighting, imaging methods based on polarization optics, and anti-interference recognition methods based on deep learning.

[0004] A recognition method based on multi-angle dynamic lighting is currently the most direct physical approach used in industry to address the high reflectivity problem. The core of this scheme is the use of a ring or dome-shaped light source composed of multiple independently controlled LED light sources. Through program control, these light sources are illuminated sequentially or in a specific pattern at extremely high frequencies. The principle is that for mirrored surfaces, reflections are highly directional. By changing the lighting angle, it is always possible to ensure that the camera sensor avoids the main specular reflection spots under a certain lighting configuration, thus capturing a clear QR code image. The system continuously acquires multiple images under different lighting conditions and then uses image selection algorithms (such as selecting the image with the highest overall contrast or the fewest reflective areas) or image fusion algorithms (such as stitching together the best-quality areas from multiple images into a "shadowless" image) to synthesize a final image for recognition. The advantage of this method lies in its clear physical principle and its ability to effectively avoid reflections to a certain extent. However, its disadvantages are significant: firstly, to cover a sufficient number of angles, a complex and expensive light source system is required, increasing hardware costs and system complexity. Secondly, dynamic lighting and multiple data acquisitions significantly reduce the system's throughput, making it difficult to meet the cycle time requirements of high-speed production lines. Finally, for wafer surfaces with varying curvature or microscopic irregularities, the reflection patterns are extremely complex, and simple multi-angle lighting may not be able to completely eliminate all interference, especially when the QR code area happens to overlap with a persistent reflective spot, in which case this method is ineffective.

[0005] This imaging method is based on polarization optics. It utilizes the polarization properties of light waves to suppress specular reflection. Its core hardware consists of a linear polarizer (analyzer) mounted in front of the camera lens and another polarizer (polarizer) mounted in front of the light source. By adjusting the relative angle (usually orthogonal) between the polarizer and the analyzer, specular reflection light from the wafer surface (whose polarization direction is roughly parallel to the incident light) can be suppressed to the greatest extent, while diffuse reflection light (whose polarization direction changes randomly) can partially pass through the analyzer. Theoretically, this can effectively "filter out" glaring reflections, highlighting the diffuse reflection signal generated by changes in the surface microstructure at the QR code laser marking location. The advantage of this method is that it directly attacks the reflection problem at a physical level, and its effect is significant under certain ideal conditions. However, its disadvantages are: firstly, the polarizer significantly attenuates the light intensity; to obtain a sufficient signal-to-noise ratio, stronger illumination or longer exposure times are required, which may lead to reduced energy efficiency or motion blur. Secondly, the reflection from the actual wafer surface is not ideal specular reflection; it may contain some diffuse reflection components. Furthermore, the diffuse reflection signal from the QR code markings may be very weak, leading to a decrease in overall image contrast and making the QR code features appear blurry. Finally, this method has limited effectiveness for mixed reflected light with complex polarization states, and system debugging (such as precise calibration of the polarization angle) is quite cumbersome.

[0006] A Deep Learning-Based Anti-Interference Recognition Method. With the development of artificial intelligence, end-to-end deep learning models have been introduced to solve the QR code recognition problem in complex environments. This approach typically employs a semantic segmentation network with an encoder-decoder structure, such as U-Net or its variants. The network takes a single image of a wafer, possibly containing reflections, captured under normal lighting conditions as input. Its training objective is to directly output a binary segmentation mask, where white pixels represent the QR code area and black pixels represent the background. By providing the network with a large number of labeled images containing various reflection conditions for training, the network can theoretically learn to "ignore" reflective areas and infer its complete shape based on the contextual structure and edge information of the QR code. The advantage of this method is that it avoids complex hardware configurations, and the inference process can be very fast once the model is trained. However, its fatal weaknesses are: First, the model performance is extremely dependent on the quality and diversity of the training data. Collecting and labeling a massive dataset covering all possible reflection patterns, different wafer materials, different etching processes, and different contamination conditions is extremely costly and almost impossible to complete. Second, the "black box" nature of the model raises questions about its reliability. When faced with unfamiliar reflective patterns, the model may produce unpredictable errors, such as misidentifying large reflective areas as QR codes or missing portions of the code covered by reflections. This is unacceptable in industrial settings where 100% reliability is required. Finally, deep learning models are typically computationally intensive, posing a challenge for deployment on edge devices with high real-time requirements.

[0007] Based on this, the present invention designs a mirror wafer QR code recognition method and system based on polarization difference and structure reconstruction to solve the above problems. Summary of the Invention

[0008] To address the aforementioned shortcomings of existing technologies, this invention provides a method and system for identifying QR codes on mirror wafers based on polarization difference and structural reconstruction.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] The mirror wafer QR code recognition method based on polarization difference and structure reconstruction includes the following steps:

[0011] Step 1: Acquire parallel polarization images of the wafer mirror surface with the same field of view, exposure, and frame rate for a single channel grayscale. With vertical polarization image ;

[0012] Step 2: Using parallel polarization images With vertical polarization image Calculate polarization difference image ,based on Generate reflective masks with adaptive thresholds Simultaneously calculate polarization fusion image And complete the preprocessing to obtain the fused image. ;

[0013] Step 3: Use merged images Based on this, candidate regions for QR codes are quickly identified through morphological processing, threshold segmentation, and connected component analysis; simultaneously, polarization difference images are incorporated. With the already generated reflective mask It accurately calibrates the reflective positions within the QR code candidate area and outputs the QR code candidate area and its corresponding accurate reflective submask. ;

[0014] Step 4: Using the QR code candidate region image and the corresponding precise reflective mask Using the inherent structure of the DataMatrix L-shaped Finder Pattern as input, the system accurately locates the QR code, calculates its four corner points, and performs perspective normalization. Simultaneously, it performs a matching transformation on the reflective mask, outputting a corrected reflective mask that is pixel-by-pixel aligned with the normalized QR code. Difference image after normalization ;

[0015] Step 5: Organize the QR code image Corrected reflective mask Taking an N×N standard module grid as input, the unknown modules that are obscured by reflection and normal reliable modules are first marked; the unknown modules are first repaired by weighted voting of eight neighborhoods, and then the remaining unknown modules are submitted to the Reed-Solomon decoder according to the erasure error. The DataMatrix error correction prior is used to complete the complete recovery and output a reliable module matrix that can be directly decoded.

[0016] Step 6: Complete Binary Module Matrix The data is fed into a standard data matrix code decoding library for decoding. The decoded information is then re-encoded into an ideal, noise-free standard QR code image, which is then compared with the original regular QR code image. The code is compared and the similarity is calculated. If the similarity is higher than the set threshold, the decoding is considered successful and reliable, and the valid data information obtained from the decoding is output. If the similarity is lower than the set threshold, the decoding result is determined to be unreliable, the current result is discarded, and the process returns to step 5 to perform module repair and error correction decoding 1 to 2 times. If the code still fails to meet the standard after repeated repairs, the QR code is marked as a decoding failure.

[0017] Furthermore, the specific steps for step 2 are as follows:

[0018] Step 21: Analyze the parallel polarization image With vertical polarization image Perform pixel-by-pixel interpolation and take the absolute value to obtain the pixel-level difference result. All coordinates corresponding The numerical values ​​are filled into the two-dimensional matrix according to their original spatial positions to form a complete polarization difference image. ;

[0019] Step 22: Calculate the threshold using an image statistical adaptive algorithm. ,by As the segmentation threshold, for Pixel-by-pixel binarization to generate a reflective mask ;

[0020] Step 23: Analyze the parallel polarization image With vertical polarization image A weighted linear fusion strategy is used to calculate polarization fusion images. ;

[0021] Step 24: Polarization fusion image Preprocessing is performed using grayscale normalization, contrast-limited adaptive histogram equalization, and noise reduction filtering.

[0022] Furthermore, step 3 is performed as follows:

[0023] Step 31: Process the fused image The QR code area is coarsely located to obtain the QR code candidate area;

[0024] Step 32: Use a reflective mask Based on this, the data is mapped to the candidate QR code area and verified a second time to generate a precise reflective mask corresponding to the candidate area. .

[0025] Furthermore, the specific steps in step 4 are as follows:

[0026] Step 41: Candidate Area Reflection Shielding and Edge Enhancement

[0027] Reflective pixel shielding: based on precise reflective masks The location marked as strongly reflective in the candidate region image is set as the local background mean of the candidate region;

[0028] Edge feature enhancement: Perform Canny edge extraction on the masked image;

[0029] Enhance the alternating black and white module outlines of the Finder Pattern to improve feature detection stability.

[0030] Step 42: The Data MatrixFinder Pattern is an L-shaped Finder Pattern formed by the left vertical positioning edge and the bottom horizontal positioning edge;

[0031] Step 43: Perform line fitting on the vertical and horizontal sides of the located L-shaped Finder Pattern to obtain the line equations of the two orthogonal positioning sides. Based on the square structure of the Data Matrix and the module size, calculate the line equations of the upper and right boundaries of the QR code. The lines of the two orthogonal positioning sides, the upper boundary line, and the right boundary line intersect each other. Solve to obtain the global coordinates of the four corner points of the QR code: upper left P1, upper right P2, lower right P3, and lower left P4.

[0032] Step 44: Perspective transformation and QR code normalization to obtain the pixel region of each module.

[0033] Target standard mesh definition: Map the four corner points to a standard square target mesh (size N×N modules, where N is the number of Data Matrix modules), and determine the four corner points Q1, Q2, Q3, and Q4 of the target;

[0034] Homography Matrix Solution: Based on the original four corner points {P1,P2,P3,P4} and the target four corner points {Q1,Q2,Q3,Q4}, calculate the perspective transformation homography matrix. ;

[0035] Inverse perspective correction: Perform inverse perspective transformation on the candidate region image, use nearest neighbor interpolation for boundary processing, and output a distortion-free, pixel-level aligned, regular QR code image. ;

[0036] Module mesh generation: Mesh according to standard module dimensions Divide the grid into regular N×N modules and determine the pixel region of each module;

[0037] Step 45: Apply a precise reflective mask Using the exact same homography matrix as the QR code image A perspective transformation is performed using interpolation to obtain a corrected reflective mask that is pixel-by-pixel aligned with the regular QR code. Synchronous polarization difference image Using the same homography matrix Perform perspective transformation to obtain the normalized difference image. .

[0038] Furthermore, the specific steps in step 5 are as follows:

[0039] Step 51: Module mesh partitioning, marking module states as unknown modules and normal trusted modules:

[0040] Module mesh mapping: Based on the N×N module partitioning determined in step 4, the... and Divided into a uniform rule grid, denoted as the first... Line number The column module is

[0041] Module state binary classification: To classify modules as unknown or normal trusted modules based on their status:

[0042] Trusted Module (Valid): There is no strong reflection in the corresponding mask area, that is... The module's grayscale is reliable;

[0043] Unknown module (to be fixed): The corresponding mask area is blocked by strong reflections, that is... The grayscale values ​​of the module are unreliable and need to be fixed.

[0044] Step 52, Initial value determination of trusted modules: For all trusted modules, the binary values ​​(0 / 1) of the module are calculated by sampling the central region. The core region, which is 1 / 2 to 2 / 3 the size of the module center, is selected, and the grayscale mean is calculated. Using module-local adaptive threshold Binarization yields the initial module values. ;

[0045] Initial module value The calculation is as follows:

[0046]

[0047] : Indicates that the data module is black; : Indicates that the data module is white;

[0048] Step 53: Based on the spatial continuity of the QR code, predict the module value of the unknown module using eight-neighbor weighted voting, and prioritize the recovery of unknown modules with clear structures:

[0049] Step 54: Treat modules that are still unknown after voting as erase errors with known locations, and use the prior erase error correction code of the Data MatrixECC200 standard Reed-Solomon error correction code to repair the erase error.

[0050] Furthermore, step 53 is performed as follows:

[0051] Step 531: Determine the structure of the unknown module;

[0052] Step 532: Repair modules with clear but unknown structures;

[0053] Neighborhood range:

[0054] Weighting rules: Modules that are directly adjacent vertically or horizontally have a weight of 2, and modules that are diagonally adjacent have a weight of 1. Only trusted modules participate in voting.

[0055] Step 533: For unknown modules that do not meet the criteria for clear structure, retain the "unknown" mark, do not perform voting repair, and generate a preliminary module matrix from the repaired unknown modules and trusted modules.

[0056] Furthermore, four rigid criteria are applied to all unknown modules. If all criteria are met, the module is a structurally clear unknown module and can be repaired first; if any criterion is not met, the module is a structurally ambiguous unknown module and its unknown status is retained.

[0057] Neighborhood sample sufficiency determination: Count the number of trustworthy modules within eight neighborhoods. ,and ≥5;

[0058] Neighborhood voting consensus determination: Calculate the weighted voting confidence Conf = maximum voting weight / total neighborhood weight, and Conf ≥ 0.6;

[0059] Local structural continuity determination: The binary changes of three consecutive modules in the row and column where the verification module is located conform to the grid space continuity characteristics of the DataMatrix module;

[0060] Reflective residual intensity determination: Extraction module in regional mean ,and .

[0061] A mirror wafer QR code recognition system based on polarization difference and structure reconstruction, comprising:

[0062] Image acquisition module: Used to acquire parallel polarization images of the wafer mirror surface. With vertical polarization image ;

[0063] Polarization preprocessing and reflection detection module: used to complete polarization difference calculation, reflection mask generation, adaptive polarization fusion and image preprocessing;

[0064] QR code coarse positioning and reflective calibration module: used for QR code candidate region extraction and precise calibration of candidate region reflective properties;

[0065] QR code precision positioning and perspective correction module: used for QR code precision positioning, distortion correction and mask alignment;

[0066] QR code module repair module: used for structural repair and error correction of reflective shielding modules;

[0067] Decoding and Verification Module: Used for QR code decoding and result reliability verification.

[0068] Compared with existing technologies, this invention has the following advantages and beneficial effects: This invention achieves pixel-level precise detection of highly reflective areas through parallel / vertical dual-polarization image difference. Combined with an adaptive weighted polarization fusion strategy, it can fundamentally suppress strong specular reflection interference from wafer mirrors, while significantly enhancing the texture contrast between QR code markings and the background. This solves the technical problem of traditional imaging methods where specular reflection completely obscures QR code features, making them unrecognizable. Simultaneously, the coarse positioning scheme employing top-hat transformation, Otsu's method segmentation, and multi-dimensional geometric feature constraints can quickly eliminate noise, impurities, and other interference, accurately locking onto QR code candidate areas. Furthermore, combining reflective mask clipping and polarization difference secondary verification enables fine-grained calibration of reflective positions within the candidate areas, providing a reliable basis for subsequent module repair, significantly improving positioning accuracy and anti-interference capabilities. In addition, relying on the structural prior of the Data Matrix L-shaped positioning pattern to complete fine positioning and perspective correction can effectively eliminate image distortion caused by shooting angle and wafer placement posture, outputting a pixel-level regular standard QR code grid, ensuring uniform and accurate module division, and laying a high-precision foundation for subsequent repair and decoding. Finally, a hierarchical repair strategy is adopted, which includes structural clarity determination, eight-neighbor weighted voting, and RS erasure error correction. Priority is given to the rapid repair of structurally clear occluded modules, while the error correction code is used to reliably restore structurally ambiguous modules. This approach balances processing efficiency and fully leverages the strong error correction capability of RS codes for erasing at known locations, significantly improving the complete recovery rate of QR code modules under reflective occlusion. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0070] Fig. 1 The flowchart shows a method for recognizing QR codes on mirror wafers based on polarization difference and structure reconstruction.

[0071] Fig. 2 This is a schematic diagram of precise QR code positioning and grid division based on Finder Pattern. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0073] Example 1: Please refer to the accompanying drawings in the instruction manual. Figs. 1-2 A method for recognizing QR codes on mirror wafers based on polarization difference and structure reconstruction includes the following steps:

[0074] Step 1: Acquire parallel polarization images of the wafer mirror surface with the same field of view, exposure, and frame rate for a single channel grayscale. With vertical polarization image ;

[0075] Step 2: Using parallel polarization images With vertical polarization image Calculate polarization difference image ,based on Generate reflective masks with adaptive thresholds Simultaneously calculate polarization fusion image And complete the preprocessing to obtain the fused image. ;

[0076] The specific steps for step 2 are as follows:

[0077] Step 21: Analyze the parallel polarization image With vertical polarization image Perform pixel-by-pixel interpolation and take the absolute value to obtain the pixel-level difference result. All coordinates corresponding The numerical values ​​are filled into the two-dimensional matrix according to their original spatial positions to form a complete polarization difference image. ;

[0078] Polarization difference image It can fully highlight highly reflective areas, creating an extreme grayscale distinction from QR code markings and wafer background;

[0079] Pure specular reflection area Medium brightness, The brightness is extremely low, and the pixel difference is large. The image shows high grayscale values; the diffuse reflection area (QR code markings) shows little difference in brightness and pixel difference between the two images. The image shows a low grayscale value.

[0080] Pixel-level difference results The calculation formula is as follows:

[0081]

[0082] Step 22: Calculate the threshold using an image statistical adaptive algorithm. ,by As the segmentation threshold, for Pixel-by-pixel binarization to generate a reflective mask ;

[0083] threshold The calculation is as follows: Statistics Full-image grayscale histogram;

[0084] The segmentation threshold with the largest inter-class variance can be calculated based on Otsu's method, or an adaptively calculated threshold can be obtained by using "mean + K × standard deviation" (K is 1.5~2.5) or a percentile threshold (the top 5%~10% of the grayscale values). ;

[0085] reflective mask The calculation formula is as follows:

[0086]

[0087] reflective mask The image is binary, with a pixel value of 1 indicating a strong reflective interference area and a pixel value of 0 indicating a non-reflective area (wafer background / QR code scratches); the positions of all reflective pixels are accurately marked to provide a basis for subsequent QR code module repair;

[0088] Step 23: Analyze the parallel polarization image With vertical polarization image A weighted linear fusion strategy is used to calculate polarization fusion images. ;

[0089] Polarization fusion image The calculation formula is as follows:

[0090]

[0091] Vertical polarization image weight, with a value ranging from 0.6 to 1.0, is used to preserve... Medium diffuse reflection signal, continuous suppression of specular reflection;

[0092] The polarization difference component weight, with a value of 0.3~0.7, is used to enhance the texture difference between the QR code scratches and the background.

[0093] The weights of the vertical polarization image and the polarization difference components can be adaptively adjusted: according to... Average brightness The reflectivity ratio is dynamically optimized to ensure that the fused image is neither too bright nor too dark.

[0094] The weights of the vertically polarized image and the polarization difference components can be adaptively adjusted as follows:

[0095] Step 231: Calculate the average brightness of the vertically polarized image Reflectance ratio of sum and difference images ;

[0096] 1. Average brightness of vertically polarized images and the average brightness of parallel polarized images The calculation is as follows:

[0097]

[0098]

[0099] W: Vertical polarization image Image width; H: Vertical polarization image Image height

[0100] 2. Reflectance ratio of differential image The calculation is as follows:

[0101]

[0102] ;

[0103] Step 232: Adaptive adjustment of α

[0104] Adjust the target:

[0105] Darker (smaller) → Increase α to improve overall brightness

[0106] Brighter (larger) → Decrease α to prevent overexposure.

[0107] Constraint interval: (Ensure the vertical polarization component dominates, and stably suppress reflection)

[0108] α adaptive adjustment calculation formula.

[0109]

[0110] The baseline weight is 0.8. Target brightness: 110; Brightness adjustment coefficient: 0.002;

[0111] Step 233: Adaptive β adjustment

[0112] Adjust the target:

[0113] Reflectivity ( (Large) → Increase the size of the differential components to highlight the QR code texture.

[0114] Weak reflectivity Small) → Turn it down to avoid amplifying the noise.

[0115] Constraint interval:

[0116] Adaptive calculation formula

[0117]

[0118] The baseline weight is 0.4. Reflectivity enhancement factor: 0.4;

[0119] Step 234: Final global brightness calibration (completely eliminate overbrightness / underbrightness)

[0120] After completing the preliminary calculations of α and β, the average brightness of the fused image is pre-calculated before the final constraints are applied.

[0121] Calculate the pre-fused average brightness The specific calculations are as follows:

[0122]

[0123] Brightness clamping calibration:

[0124]

[0125] We further constrain α∈[0.6,1.0] and β∈[0.3,0.7] to ensure the parameters are valid.

[0126] Step 24: Polarization fusion image Perform grayscale normalization, contrast-limited adaptive histogram equalization, and noise reduction filtering preprocessing;

[0127] 1. Gray-scale normalization

[0128] Polarization-fused images Pixel values ​​are linearly normalized to 0~255 (8-bit standard grayscale range) to avoid pixel value overflow or uneven distribution caused by weighted calculation.

[0129] 2. Contrast-limited adaptive histogram equalization (CLAHE)

[0130] The grayscale-normalized image is divided into several local sub-blocks of 8×8 / 16×16.

[0131] Perform histogram equalization independently on each sub-block and limit contrast gain (to avoid excessive noise enhancement).

[0132] Bilinear interpolation is used to splice sub-blocks to eliminate block artifacts;

[0133] Core function: Locally enhances the contrast of QR code scratches without amplifying background noise, preserving the structural details of the QR code.

[0134] 3. Noise reduction filtering (optional)

[0135] Based on the imaging noise intensity, a 3×3 median filter (to remove salt-and-pepper noise) or a small-radius Gaussian filter (to remove Gaussian noise) is used, with the filter kernel size not exceeding 3×3, to perform noise reduction filtering on the equalized image, resulting in the fused image. ;

[0136] Step 3: Use merged images Based on this, candidate regions for QR codes are quickly identified through morphological processing, threshold segmentation, and connected component analysis; simultaneously, polarization difference images are incorporated. With the already generated reflective mask It accurately calibrates the reflective positions within the QR code candidate area and outputs the QR code candidate area and its corresponding accurate reflective submask. ;

[0137] Step 3 is performed as follows:

[0138] Step 31: Process the fused image The QR code area is coarsely located to obtain the QR code candidate area;

[0139] Step 311: Process the fused image Perform a top-hat transform to highlight the QR code scratch area with dark texture in the image;

[0140] The top cap transformation uses 8×8~16×16 rectangular structural elements. It matches the size of the Data Matrix module;

[0141] Top hat transformation formula:

[0142]

[0143] The morphological opening operation (erosion followed by dilation) is used to eliminate small bright spots and smooth boundaries.

[0144] Transformation effect: Enhances the grayscale difference between the scratches and the background, and eliminates slight unevenness in the brightness of the mirror surface.

[0145] Step 312: Results of top-hat transformation The optimal segmentation threshold is calculated using the Otsu method. Binarization separation of foreground and background:

[0146] statistics The threshold value with the largest inter-class variance is selected from the grayscale histogram of the entire image. ;

[0147] Pixel-by-pixel binarization to generate a binary image :

[0148]

[0149] Pixel value 1 represents QR code scratches / noise foreground, and pixel value 0 represents wafer mirror background.

[0150] Step 313: Process the binary image Perform 8-neighbor connected component labeling, extract all independent foreground regions, and filter QR code candidate regions according to geometric features to obtain QR code candidate regions;

[0151] Connected component labeling: Records information such as the coordinates, area, circumscribed rectangle, and perimeter of each connected component;

[0152] Geometric features:

[0153] Area constraint: The area S of the connected region satisfies the following condition: total number of QR code modules × number of pixels per module ± 30% (excluding excessively small noise and excessively large impurities);

[0154] Aspect ratio constraint: The aspect ratio of the bounding rectangle is between 0.8 and 1.2 (the Data Matrix is ​​a square, close to 1:1).

[0155] Compactness Constraints: Connectivity Component Compactness ( The perimeter of the connected region is between 0.6 and 0.9 (excluding slender and broken parts).

[0156] Euler number constraint: The Euler number of the connected components conforms to the characteristics of the Data Matrix, with no excessive holes / breaks;

[0157] Output: After filtering, 1 to 3 QR code candidate regions are obtained. The coordinates of the bounding rectangle of each QR code candidate region are recorded. .

[0158] Step 32: Use a reflective mask Based on this, the data is mapped to the QR code candidate area and verified a second time to generate a precise reflective mask corresponding to the candidate area.

[0159] Region cropping mapping: Based on the coordinates of the bounding rectangle of the QR code candidate region, from the reflective mask... By cropping the corresponding sub-regions, the initial reflective submask is obtained.

[0160]

[0161] (x,y) are the local coordinates within the QR code candidate region. , () represents the global top-left corner coordinates of the QR code candidate area;

[0162] Secondary verification of reflection: combined with polarization difference image ,right The reflective pixels marked as 1 are subjected to secondary verification to remove mismarks caused by noise, retaining only pixels with strong specular reflection, thus obtaining the accurate reflective submask corresponding to each QR code candidate area. .

[0163] Step 4: Using the QR code candidate region image and the corresponding precise reflective mask Using the inherent structure of the DataMatrix L-shaped Finder Pattern as input, the system accurately locates the QR code, calculates its four corner points, and performs perspective normalization. Simultaneously, it performs a matching transformation on the reflective mask, outputting a corrected reflective mask that is pixel-by-pixel aligned with the normalized QR code. Difference image after normalization ;

[0164] Step 4:

[0165] Step 41: Candidate Area Reflection Shielding and Edge Enhancement

[0166] Reflective pixel shielding: based on precise reflective masks The location marked as a strong reflective element (pixel value = 1) in the candidate region image is set as the local background mean of the candidate region.

[0167] Edge feature enhancement: Perform Canny edge extraction on the masked image;

[0168] Enhance the alternating black and white module outlines of the Finder Pattern to improve feature detection stability.

[0169] Step 42: The Data MatrixFinder Pattern is an L-shaped Finder Pattern formed by the left vertical positioning edge and the bottom horizontal positioning edge;

[0170] The L-shaped Finder Pattern consists of alternating black and white modules of equal width, with the row / column grayscale projection forming a regular square wave.

[0171] Row / column grayscale projection calculation: Calculate grayscale projection column by column along the vertical direction. Calculate grayscale projection line by line along the horizontal direction. ;

[0172] Square wave feature matching: Search for square wave bands with equal periodicity, high contrast, and continuous alternation in the projection sequence to match the periodic features of the Finder Pattern module;

[0173] L-shaped Finder Pattern Positioning Edge Lock: Determines the precise pixel positions of the vertical positioning edge (left boundary) and the horizontal positioning edge (bottom boundary).

[0174] Step 43: Perform line fitting on the vertical and horizontal sides of the located L-shaped Finder Pattern to obtain the line equations of the two orthogonal positioning sides. Based on the square structure of the Data Matrix and the module size, calculate the line equations of the upper and right boundaries of the QR code. The lines of the two orthogonal positioning sides, the upper boundary line, and the right boundary line intersect each other. Solve to obtain the global coordinates of the four corner points of the QR code: upper left P1, upper right P2, lower right P3, and lower left P4.

[0175] Step 44: Perspective transformation and QR code normalization to obtain the pixel region of each module.

[0176] Target standard mesh definition: Map the four corner points to a standard square target mesh (size N×N modules, where N is the number of Data Matrix modules), and determine the four corner points Q1, Q2, Q3, and Q4 of the target;

[0177] Homography Matrix Solution: Based on the original four corner points {P1,P2,P3,P4} and the target four corner points {Q1,Q2,Q3,Q4}, calculate the perspective transformation homography matrix. ;

[0178] Inverse perspective correction: Perform inverse perspective transformation on the candidate region image, use nearest neighbor interpolation for boundary processing, and output a distortion-free, pixel-level aligned, regular QR code image. ;

[0179] Module mesh generation: Mesh according to standard module dimensions Divide the grid into regular N×N modules and determine the pixel region of each module.

[0180] Step 45: Apply a precise reflective mask Using the exact same homography matrix as the QR code image A perspective transformation is performed using interpolation to obtain a corrected reflective mask that is pixel-by-pixel aligned with the regular QR code. Synchronous polarization difference image Using the same homography matrix Perform perspective transformation to obtain the normalized difference image. .

[0181] Step 5: Organize the QR code image Corrected reflective mask Taking an N×N standard module grid as input, the unknown modules that are obscured by reflection and normal reliable modules are first marked; the unknown modules are first repaired by weighted voting of eight neighborhoods, and then the remaining unknown modules are submitted to the Reed-Solomon decoder according to the erasure error. The DataMatrix error correction prior is used to complete the complete recovery and output a reliable module matrix that can be directly decoded.

[0182] The specific steps for step 5 are as follows:

[0183] Step 51: Module mesh partitioning, marking module states as unknown modules and normal trusted modules:

[0184] Module mesh mapping: Based on the N×N module partitioning determined in step 4, the... and Divided into a uniform rule grid, denoted as the first... Line number The column module is

[0185] Module state binary classification: To classify modules as unknown or normal trusted modules based on their status:

[0186] Trusted Module (Valid): There is no strong reflection in the corresponding mask area, that is... The module's grayscale is reliable;

[0187] Unknown module (to be fixed): The corresponding mask area is blocked by strong reflections, that is... The grayscale values ​​of the module are unreliable and need to be fixed.

[0188] Step 52, Initial value determination of trusted modules: For all trusted modules, the binary values ​​(0 / 1) of the module are calculated by sampling the central region. The core region, which is 1 / 2 to 2 / 3 the size of the module center, is selected, and the grayscale mean is calculated. Using module-local adaptive threshold Binarization yields the initial module values. ;

[0189] Initial module value The calculation is as follows:

[0190]

[0191] : Indicates that the data module is black; : Indicates that the data module is white;

[0192] Step 53: Based on the spatial continuity of the QR code, predict the module value of the unknown module using eight-neighbor weighted voting, and prioritize the recovery of unknown modules with clear structures:

[0193] Step 531, Determining the structure of unknown modules:

[0194] Four rigid criteria are applied to all unknown modules. If all criteria are met, the module is a structurally clear unknown module and can be repaired first. If any criterion is not met, the module is a structurally ambiguous unknown module and its unknown label is retained.

[0195] Neighborhood sample sufficiency determination: Count the number of trustworthy modules within eight neighborhoods. ,and ≥5;

[0196] Neighborhood voting consensus determination: Calculate the weighted voting confidence Conf = maximum voting weight / total neighborhood weight, and Conf ≥ 0.6;

[0197] Local structural continuity determination: The binary changes of three consecutive modules in the row and column where the verification module is located conform to the grid space continuity characteristics of the DataMatrix module;

[0198] Reflective residual intensity determination: Extraction module in regional mean ,and ( (Adaptive weak reflectivity threshold).

[0199] Step 532: Repair modules with clear but unknown structures;

[0200] Neighborhood range:

[0201] Weighting rules: Modules that are directly adjacent vertically or horizontally have a weight of 2, and modules that are diagonally adjacent have a weight of 1. Only trusted modules participate in the voting.

[0202] The vote calculation method is as follows:

[0203]

[0204] δ(·) represents the weight of the neighborhood module and is the equality indicator function (1 for equality and 0 for inequality).

[0205] Step 533: For unknown modules that do not meet the criteria for clear structure, retain the "unknown" mark, do not perform voting repair, and generate a preliminary module matrix from the repaired unknown modules and trusted modules.

[0206] Step 54: Treat modules that remain unknown after voting as erase errors with known locations, and repair them using the prior erase error correction codes of the Data MatrixECC200 standard Reed-Solomon error correction code.

[0207] Traverse the modules that still retain the "unknown" mark after being repaired by weighted voting of eight neighbors, extract their row and column coordinate information, and generate a set of known location erasure errors. m is the total number of erase modules to be corrected. Corresponding to the The row and column coordinates of each erase module in an N×N module matrix.

[0208] Based on the Reed-Solomon (RS) error correction code characteristics built into the Data Matrix ECC200 standard (Data Matrix Code 200 Level Error Correction Code), its error correction capability for erasure errors at known locations has an upper limit of n−k (where n is the code length of the RS code and k is the information length of the RS code). This error correction capability is significantly better than the error correction performance for random errors.

[0209] Combine the preliminary module matrix output in step 53 (including the binary representation of trusted modules, the binary representation of voting repair modules, and unknown markers) with the erase location set. The input is a standard RS decoder. Based on prior information about the known erasure positions, the standard RS decoder performs optimal value recovery on all erasure modules through algebraic operations of the RS code, and finally outputs a complete binary module matrix with no unknown markers and no missing data. .

[0210] Step 6: Complete Binary Module Matrix The data is fed into a standard data matrix code decoding library for decoding. The decoded information is then re-encoded into an ideal, noise-free standard QR code image, which is then compared with the original regular QR code image. The code is compared and the similarity is calculated. If the similarity is higher than the set threshold, the decoding is considered successful and reliable, and the valid data information obtained from the decoding is output. If the similarity is lower than the set threshold, the decoding result is determined to be unreliable, the current result is discarded, and the process returns to step 5 to perform module repair and error correction decoding 1 to 2 times. If the code still fails to meet the standard after repeated repairs, the QR code is marked as a decoding failure.

[0211] The threshold is set at 95%;

[0212] This invention achieves pixel-level precise detection of highly reflective areas through parallel / vertical dual-polarization image difference. Combined with an adaptive weighted polarization fusion strategy, it can fundamentally suppress strong specular reflection interference from wafer mirrors, while significantly enhancing the texture contrast between QR code markings and the background. This solves the technical problem of traditional imaging methods where specular reflections completely obscure QR code features, making them unrecognizable. Simultaneously, a coarse positioning scheme employing top-hat transformation, Otsu's method segmentation, and multi-dimensional geometric feature constraints can quickly eliminate noise, impurities, and other interference, accurately locking onto QR code candidate areas. Further, by combining reflective mask clipping and polarization difference secondary verification, fine-tuning of reflective positions within candidate areas is achieved, providing a reliable basis for subsequent module repair, significantly improving positioning accuracy and anti-interference capabilities. Furthermore, relying on the structural prior of the Data Matrix L-shaped positioning pattern for fine positioning and perspective correction effectively eliminates image distortion caused by shooting angle and wafer placement posture, outputting a pixel-level regular standard QR code grid, ensuring uniform and accurate module division, and laying a high-precision foundation for subsequent repair and decoding. Finally, a hierarchical repair strategy is adopted, which includes structural clarity determination, eight-neighbor weighted voting, and RS erasure error correction. Priority is given to quickly repairing structurally clear obscured modules, while structurally blurred modules are reliably recovered using error-correcting codes prior to their removal. This approach balances processing efficiency with the strong error-correcting capability of RS codes for erasing at known locations, significantly improving the complete recovery rate of QR code modules under reflective obscuration. The entire process employs adaptive parameter adjustment and standardized algorithm design, requiring no manual intervention. It adapts to the automated recognition of QR codes on wafers of different reflective intensities and sizes, combining real-time performance and stability. It can be directly embedded into wafer production line inspection systems, demonstrating strong engineering practical value.

[0213] Example 2: A mirror wafer QR code recognition system based on polarization difference and structure reconstruction, utilizing a mirror wafer QR code recognition method based on polarization difference and structure reconstruction, including:

[0214] Image acquisition module: Used to acquire parallel polarization images of the wafer mirror surface. With vertical polarization image ;

[0215] 1. Image acquisition module: The ring-shaped LED white light source with integrated linear polarizer, high-resolution monochrome industrial camera with telecentric lens, and rotary linear analyzer driven by stepper motor are aligned with the QR code area to be tested on the wafer mirror. The telecentric lens ensures no dimensional distortion across the entire field of view.

[0216] The exposure time, gain, resolution, and frame rate of the high-resolution monochrome industrial camera are locked, as are the brightness, color temperature, and output power of the ring-shaped LED white light source. These settings remain unchanged throughout the process, ensuring that the brightness and signal-to-noise ratio of the two images are comparable.

[0217] Verify the stepper motor's rotational accuracy, the light source's light output stability, and the camera's image acquisition response speed to ensure there is no hardware delay or malfunction.

[0218] 2. Parallel polarization image Data acquisition (polarizer and analyzer polarization directions are parallel)

[0219] Polarization positioning: Drive the stepper motor to rotate the rotary linear analyzer in front of the telecentric lens to an angle that is completely parallel to the polarization direction of the linear polarizer integrated with the light output port of the ring LED white light source. The stepper motor is locked in position to ensure that there is no deviation in polarization direction.

[0220] Image Acquisition: A ring-shaped LED white light source is triggered to stably emit light. A high-resolution monochrome industrial camera acquires a frame of raw grayscale image with locked parameters. In this image, both specular and diffuse reflection light can pass through the polarization component in large quantities, and this image is denoted as a parallel polarization image. .

[0221] Data storage: The original grayscale format is cached to retain complete pixel brightness information without preprocessing or compression.

[0222] III. Vertically Polarized Images Acquisition (polarizer and analyzer polarization directions are perpendicular)

[0223] Rapid polarization state switching: The driving stepper motor rotates precisely 90° to ensure that the polarization directions of the linear analyzer and the linear polarizer are strictly orthogonal. The switching time is ≤ the production line cycle time threshold, avoiding wafer displacement.

[0224] Image acquisition: Maintaining all parameters of the light source and high-resolution monochrome industrial camera. The acquisition process was completely consistent, immediately triggering a high-resolution monochrome industrial camera to acquire a second frame of the original grayscale image. In this image, the specular reflection light from the wafer mirror was significantly suppressed, with only diffuse reflection light (QR code marking signal) passing through the polarization component, denoted as the vertically polarized image. .

[0225] Data synchronization: will and Pixel-aligned buffering ensures that the two-frame images are free from displacement, distortion, and brightness deviation.

[0226] Vertical polarization image and parallel polarization image The acquisition parameters were exactly the same: exposure, gain, light source power, and lens focal length were all the same, only the polarization angle was changed.

[0227] High-speed continuous acquisition: The total acquisition latency meets the high-speed cycle time of semiconductor production lines, avoiding image registration errors caused by wafer movement.

[0228] Raw grayscale output: High-resolution monochrome industrial cameras do not perform post-processing such as automatic white balance, automatic gain, or sharpening, outputting pure raw images to ensure the accuracy of polarization difference calculation.

[0229] Precise polarization angle: The parallel / perpendicular state is controlled by a stepper motor in a closed loop, with an angle error of ≤±0.5°, ensuring that the specular reflection light suppression effect meets the standard.

[0230] Polarization preprocessing and reflection detection module: used to perform polarization difference calculation, reflection mask generation, adaptive polarization fusion and image preprocessing in step 2;

[0231] QR code coarse positioning and reflective calibration module: used to perform QR code candidate region extraction and candidate region reflective calibration in step 3;

[0232] QR code fine positioning and perspective correction module: used to perform QR code fine positioning, distortion correction and mask alignment in step 4;

[0233] QR code module repair module: used to perform structural repair and error correction of the reflective shielding module in step 5;

[0234] Decoding and Verification Module: Used to perform QR code decoding and result reliability verification in step 6.

[0235] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for recognizing QR codes on mirror wafers based on polarization difference and structure reconstruction, characterized in that: Includes the following steps: Step 1: Acquire parallel polarization images of the wafer mirror surface with the same field of view, exposure, and frame rate for a single channel grayscale. With vertical polarization image ; Step 2: Using parallel polarization images With vertical polarization image Calculate polarization difference image ,based on Generate reflective masks with adaptive thresholds Simultaneously calculate polarization fusion image And complete the preprocessing to obtain the fused image. ; Step 3: Use merged images Based on this, candidate regions for QR codes are quickly identified through morphological processing, threshold segmentation, and connected component analysis; simultaneously, polarization difference images are incorporated. With the already generated reflective mask It accurately calibrates the reflective positions within the QR code candidate area and outputs the QR code candidate area and its corresponding accurate reflective submask. ; Step 4: Using the QR code candidate region image and the corresponding precise reflective mask Using the inherent structure of the DataMatrix L-shaped Finder Pattern as input, the system accurately locates the QR code, calculates its four corner points, and performs perspective normalization. Simultaneously, it performs a matching transformation on the reflective mask, outputting a corrected reflective mask that is pixel-by-pixel aligned with the normalized QR code. Difference image after normalization ; Step 5: Organize the QR code image Corrected reflective mask Taking an N×N standard module grid as input, the unknown modules that are obscured by reflection and normal reliable modules are first marked; the unknown modules are first repaired by weighted voting of eight neighborhoods, and then the remaining unknown modules are submitted to the Reed-Solomon decoder according to the erasure error. The Data Matrix error correction prior is used to complete the complete recovery and output a reliable module matrix that can be directly decoded. Step 6: Complete Binary Module Matrix The data is fed into a standard data matrix code decoding library for decoding. The decoded information is then re-encoded into an ideal, noise-free standard QR code image, which is then compared with the original regular QR code image. Compare and calculate the similarity. If the similarity is higher than the set threshold, the decoding is considered successful and reliable, and the valid data information obtained from the decoding is output. If the similarity is lower than the set threshold, the decoding result is determined to be unreliable, the current result is discarded, and the process returns to step 5 to repeat the module repair and error correction decoding 1 to 2 times. If the code still fails to meet the requirements after repeated repairs, then mark the QR code as having failed to decode.

2. The method for QR code recognition on mirror wafers based on polarization difference and structure reconstruction according to claim 1, characterized in that, The specific steps for step 2 are as follows: Step 21: Analyze the parallel polarization image With vertical polarization image Perform pixel-by-pixel interpolation and take the absolute value to obtain the pixel-level difference result. All coordinates corresponding The numerical values ​​are filled into the two-dimensional matrix according to their original spatial positions to form a complete polarization difference image. ; Step 22: Calculate the threshold using an image statistical adaptive algorithm. ,by As the segmentation threshold, for Pixel-by-pixel binarization to generate a reflective mask ; Step 23: Analyze the parallel polarization image With vertical polarization image A weighted linear fusion strategy is used to calculate polarization fusion images. ; Step 24: Polarization fusion image Preprocessing is performed using grayscale normalization, contrast-limited adaptive histogram equalization, and noise reduction filtering.

3. The method for QR code recognition on mirror wafers based on polarization difference and structure reconstruction according to claim 2, characterized in that, Step 3 is performed as follows: Step 31: Process the fused image The QR code area is coarsely located to obtain the QR code candidate area; Step 32: Use a reflective mask Based on this, the data is mapped to the candidate QR code area and verified a second time to generate a precise reflective mask corresponding to the candidate area. .

4. The method for QR code recognition on mirror wafers based on polarization difference and structure reconstruction according to claim 3, characterized in that, Step 4: Step 41: Candidate Area Reflection Shielding and Edge Enhancement Reflective pixel shielding: based on precise reflective masks The location marked as strongly reflective in the candidate region image is set as the local background mean of the candidate region; Edge feature enhancement: Perform Canny edge extraction on the masked image; Enhance the alternating black and white module outlines of Finder Pattern to improve feature detection stability; Step 42: The Data MatrixFinder Pattern is an L-shaped Finder Pattern formed by the left vertical positioning edge and the bottom horizontal positioning edge; Step 43: Perform line fitting on the vertical and horizontal sides of the located L-shaped Finder Pattern to obtain the line equations of the two orthogonal positioning sides. Based on the square structure of the Data Matrix and the module size, calculate the line equations of the upper and right boundaries of the QR code. The lines of the two orthogonal positioning sides, the upper boundary line, and the right boundary line intersect each other. Solve to obtain the global coordinates of the four corner points of the QR code: upper left P1, upper right P2, lower right P3, and lower left P4. Step 44: Perspective transformation and QR code normalization to obtain the pixel region of each module. Target standard mesh definition: Map the four corner points to a standard square target mesh (size N×N modules, where N is the number of Data Matrix modules), and determine the four corner points Q1, Q2, Q3, and Q4 of the target; Homography Matrix Solution: Based on the original four corner points {P1,P2,P3,P4} and the target four corner points {Q1,Q2,Q3,Q4}, calculate the perspective transformation homography matrix. ; Inverse perspective correction: Perform inverse perspective transformation on the candidate region image, use nearest neighbor interpolation for boundary processing, and output a distortion-free, pixel-level aligned, regular QR code image. ; Module mesh generation: Mesh according to standard module dimensions Divide the grid into regular N×N modules and determine the pixel region of each module; Step 45: Apply a precise reflective mask Using the exact same homography matrix as the QR code image A perspective transformation is performed using interpolation to obtain a corrected reflective mask that is pixel-by-pixel aligned with the regular QR code. Synchronous polarization difference image Using the same homography matrix Perform perspective transformation to obtain the normalized difference image. .

5. The mirror wafer QR code recognition method based on polarization difference and structure reconstruction according to claim 4, characterized in that, The specific steps for step 5 are as follows: Step 51: Module mesh partitioning, marking module states as unknown modules and normal trusted modules: Module mesh mapping: Based on the N×N module partitioning determined in step 4, the... and Divided into a uniform rule grid, denoted as the first... Line number The column module is ; Module state binary classification: To classify modules as unknown or normal trusted modules based on their status: Trusted Module (Valid): There is no strong reflection in the corresponding mask area, that is... The module's grayscale is reliable; Unknown module (to be fixed): The corresponding mask area is blocked by strong reflections, that is... The grayscale values ​​of the module are unreliable and need to be fixed. Step 52, Initial value determination of trusted modules: For all trusted modules, the binary values ​​(0 / 1) of the module are calculated by sampling the central region. The core region, which is 1 / 2 to 2 / 3 the size of the module center, is selected, and the grayscale mean is calculated. Using module-local adaptive threshold Binarization yields the initial module values. ; Initial module value The calculation is as follows: ； : Indicates that the data module is black; : Indicates that the data module is white; Step 53: Based on the spatial continuity of the QR code, predict the module value of the unknown module using eight-neighbor weighted voting, and prioritize the recovery of unknown modules with clear structures: Step 54: Treat modules that are still unknown after voting as erase errors with known locations, and use the prior erase error correction code of the Data MatrixECC200 standard Reed-Solomon error correction code to repair the erase error.

6. The mirror wafer QR code recognition method based on polarization difference and structure reconstruction according to claim 5, characterized in that, Step 53 is performed as follows: Step 531: Determine the structure of the unknown module; Step 532: Repair modules with clear but unknown structures; Neighborhood range: ; Weighting rules: Modules that are directly adjacent vertically or horizontally have a weight of 2, and modules that are diagonally adjacent have a weight of 1. Only trusted modules participate in voting. Step 533: For unknown modules that do not meet the criteria for clear structure, retain the "unknown" mark, do not perform voting repair, and generate a preliminary module matrix from the repaired unknown modules and trusted modules.

7. The method for QR code recognition on mirror wafers based on polarization difference and structure reconstruction according to claim 6, characterized in that, For all unknown modules, perform four rigid checks. If all four checks are met, the module is a structurally clear unknown module and can be repaired first. If any one check is not met, the module is a structurally ambiguous unknown module and its unknown label is retained. Neighborhood sample sufficiency determination: Count the number of trustworthy modules within eight neighborhoods. ,and ≥5; Neighborhood voting consensus determination: Calculate the weighted voting confidence Conf = maximum voting weight / total neighborhood weight, and Conf ≥ 0.6; Local structural continuity determination: The binary changes of three consecutive modules in the row and column where the verification module is located conform to the grid space continuity characteristics of the Data Matrix module; Reflective residual intensity determination: Extraction module in regional mean ,and .

8. A mirror wafer QR code recognition system based on polarization difference and structure reconstruction, utilizing the mirror wafer QR code recognition method based on polarization difference and structure reconstruction as described in any one of claims 1-7, characterized in that, include: Image acquisition module: Used to acquire parallel polarization images of the wafer mirror surface. With vertical polarization image ; Polarization preprocessing and reflection detection module: used to complete polarization difference calculation, reflection mask generation, adaptive polarization fusion and image preprocessing; QR code coarse positioning and reflective calibration module: used for QR code candidate region extraction and precise calibration of candidate region reflective properties; QR code precision positioning and perspective correction module: used for QR code precision positioning, distortion correction and mask alignment; QR code module repair module: used for structural repair and error correction of reflective shielding modules; Decoding and Verification Module: Used for QR code decoding and result reliability verification.

Images