A western blot image high-throughput identification method and system

Abstract

The application provides a Western Blot image high-throughput identification method and system, relates to the technical field of image processing, and comprises the following steps: obtaining a Western Blot image set to be checked for duplication, and preprocessing into a standard query image set; extracting the features of each standard query image in the standard query image set under different modalities, and packing and converting into a high-dimensional feature tensor; performing two-by-two asymmetric intelligent routing on the images in the standard query image set, and determining each pair of compared query set and index set; calculating all Euclidean distances at one time, and screening out the best candidate geometric transformation parameters from the query set to the index set in each pair of comparison; performing vector flow consistency cleaning and homography matrix health degree checking on the candidate geometric transformation parameters, and determining whether there is a repeated area between the query set and the index set in each pair of comparison; and the application can ensure high recall rate and low false positive rate, and significantly improve the processing efficiency and evidence intuitiveness of large-scale image duplication checking.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a high-throughput identification method and system for Western blotting images. Background Technology

[0002] In recent years, the number of research papers in the biomedical field has surged, and academic misconduct, especially image manipulation and forgery, has become increasingly prominent. Western blotting images, in particular, have become a major target for image fraud due to their inclusion of key experimental conclusions and their visual characteristics such as "high background noise, simple texture structure, and blurred edges." Traditional peer review relies on manual visual checks, which is inefficient and struggles to identify elaborate manipulation techniques such as cropping, splicing, color inversion, and blurring.

[0003] Existing automated image deduplication technologies are mainly divided into two categories: hash fingerprint-based retrieval and feature point-based matching (such as SIFT and SURF). Hash algorithms are extremely sensitive to image rotation, stretching, and nonlinear deformation; even slight adjustments can lead to detection failure. General feature point matching algorithms struggle to distinguish between valid biological bands and random background noise, resulting in a high false positive rate. Furthermore, existing methods lack the ability to extract multimodal features (such as texture, inverted colors, and high-frequency noise residuals), and require the brute-force generation of numerous variants when dealing with geometric attacks, leading to high computational load and slow detection speed.

[0004] Therefore, existing technologies suffer from high false negative rates, low computational efficiency, and a lack of in-depth forensic capabilities, making it difficult to meet the needs of large-scale scientific research image screening. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a high-throughput identification method and system for Western Blot images, which significantly improves the processing efficiency and intuitiveness of large-scale image deduplication while ensuring high recall and low false alarm rate.

[0006] According to some embodiments, the present invention adopts the following technical solution: A high-throughput identification method for Western Blot images, comprising: Obtain the set of Western Blot images to be checked for deduplication, and perform adaptive normalization preprocessing on each image in the set to obtain the standard query image set; Construct a multimodal adversarial feature space and extract features of each standard query image in the standard query image set under different modalities; Each modal feature of each standard query image is packaged into a geometric variant tensor and then converted into a high-dimensional feature tensor. Based on the high-dimensional feature tensor, pairwise asymmetric intelligent routing is performed on the images in the standard query image set to determine the query set and index set in each pair. Through GPU-accelerated elastic matrix scanning, all Euclidean distances are calculated in a high-throughput manner to filter out the best candidate geometric transformation parameters from the query set to the index set in each pair. Vector flow consistency cleaning and homography matrix health checks are performed on the candidate geometric transformation parameters to determine whether there are duplicate regions between the query set and index set in each pair. If duplicates are found, a reverse evidence map based on visual normalization is generated to show pixel-level aligned tampering traces.

[0007] According to some embodiments, the present invention adopts the following technical solution: A high-throughput identification system for Western Blot images, comprising: The acquisition module is configured to: acquire the Western Blot image set to be deduplicated, and perform adaptive normalization preprocessing on each image in the image set to obtain a standard query image set; The extraction module is configured to: construct a multimodal adversarial feature space and extract features of each standard query image in the standard query image set under different modalities; The packaging module is configured to package each modal feature of each standard query image into a geometric variant tensor and convert it into a high-dimensional feature tensor. The duplicate detection module is configured to: perform pairwise asymmetric intelligent routing on the images in the standard query image set based on the high-dimensional feature tensor to determine the query set and index set in each pair; calculate all Euclidean distances in a high-throughput manner using GPU-accelerated elastic matrix scanning to select the best candidate geometric transformation parameters from the query set to the index set in each pair; perform vector flow consistency cleaning and homography matrix health verification on the candidate geometric transformation parameters to determine whether there are duplicate regions between the query set and the index set in each pair; if duplicates are determined to exist, generate a reverse evidence map based on visual normalization to show pixel-level aligned tampering traces.

[0008] According to some embodiments, the present invention adopts the following technical solution: A computer program product includes a computer program that, when executed by a processor, implements the aforementioned high-throughput identification method for Western Blot images.

[0009] According to some embodiments, the present invention adopts the following technical solution: A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned high-throughput identification method for Western Blot images.

[0010] According to some embodiments, the present invention adopts the following technical solution: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform a Western Blot image high-throughput identification method.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a multimodal adversarial feature space and extracts features from different modalities, effectively addressing various image tampering techniques such as color inversion and blurring, significantly improving detection robustness. By packaging features into geometric variant tensors and converting them into high-dimensional feature tensors, combined with asymmetric intelligent routing to determine the query set and index set, redundant computation is avoided. Furthermore, GPU-accelerated elastic matrix scanning is used to calculate all Euclidean distances in a high-throughput manner, quickly selecting the best candidate geometric transformation parameters and significantly improving the efficiency of large-scale image comparison. By performing vector flow consistency cleaning and homography matrix health checks on candidate transformation parameters, false positive matches are effectively eliminated, ensuring the accuracy of the judgment results. When duplicate judgments are found, a visually normalized inverse evidence map is generated, which can intuitively display tampering traces in a pixel-level aligned manner, providing irrefutable evidence support for academic misconduct review. Attached Figure Description

[0012] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0013] Figure 1 This is a flowchart of the method in Example 1.

[0014] Figure 2 This is a flowchart of feature extraction and packaging for Example 1.

[0015] Figure 3 This is a flowchart of the repeat detection process in Example 1. Detailed Implementation

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0019] Example 1 One embodiment of the present invention provides a high-throughput identification method for Western blotting images, which significantly improves the processing efficiency and intuitiveness of large-scale image deduplication while ensuring high recall and low false alarm rate. Figure 1 As shown, the specific steps include: Step S1: Obtain the Western Blot image set to be checked for deduplication, and perform adaptive normalization preprocessing on each image in the image set to obtain the standard query image set.

[0020] Each image in the Western Blot image set to be checked for plagiarism is a grayscale bar image generated by a Western Blot experiment. Due to differences in scanning equipment between different laboratories, the original images often contain invalid black background borders or noise in non-experimental areas. This step aims to clean the data of each image in the set to ensure the purity of subsequent feature extraction. If the input is a color image, it needs to be converted into a single-channel grayscale image first, specifically including: S11: For each image in the image set, use an adaptive threshold segmentation and contour search algorithm to remove invalid black borders around the image.

[0021] First, read the original data of the current image, set the binarization threshold (e.g., 25), convert the image into a binary mask, use the contour search algorithm to retrieve all connected components in the mask, and calculate the bounding rectangle of each connected component.

[0022] To distinguish between valid experimental regions and background noise, the system calculates the ratio of the area of ​​the bounding rectangle of each connected region to the total area of ​​the original image. If the area ratio of a bounding rectangle is less than a preset threshold (e.g., 2%), or if the aspect ratio of the rectangle is extremely skewed (e.g., greater than 20:1), the connected region is determined to be scanning noise, and its corresponding bounding rectangle is filtered out as an invalid candidate box (skipped) when selecting valid regions. Finally, among the remaining valid candidate boxes, the bounding rectangle with the largest area is selected as the valid experimental region of the current image, and the current image is cropped according to the coordinates of the rectangle, while expanding the buffer margin by a preset number of pixels (e.g., 2 pixels).

[0023] Perform the black border removal operation described above independently on each image in the image set to obtain the cropped image for each image.

[0024] S12: Based on the resolution of each cropped image, adaptive resampling is performed on the image using the Lanczos-4 window Sink interpolation algorithm.

[0025] To balance feature extraction richness with computational efficiency, a dynamic scaling rule is set independently for each image. Specifically, the resolution (H, W) of the current cropped image is obtained: if the longest side of the image is less than 500 pixels, a 1.5x upsampling is performed to enhance the details of weak bands; if the longest side of the image is greater than 2500 pixels, a 0.5x downsampling is performed to suppress high-frequency noise.

[0026] During scaling operations, the Lanczos-4 Windowed Sinc Interpolation algorithm is forced to be used instead of traditional bilinear or bicubic interpolation. The Lanczos-4 algorithm uses 8×8 neighboring pixels to perform weighted convolution through the Sinc function, which can preserve the edge sharpness and high-frequency texture features of weak stripes in the Western Blot image to the greatest extent, preventing the loss of key fingerprint information during image stretching or scaling.

[0027] The adaptive resampling operation described above is performed independently on each cropped image in the image set to obtain the standard query image set.

[0028] Step S2: Construct a multimodal adversarial feature space and extract features of each standard query image in the standard query image set under different modalities.

[0029] Furthermore, the multimodal adversarial feature space includes texture mode, inverse color mode, ghost layer mode, morphological blob mode, and dynamic Gamma penetration mode; wherein, the ghost layer mode extracts high-frequency residual skeletons through Gaussian blur and difference from the original image, adaptive threshold binarization, and morphological opening and closing operations, which are used to detect edge traces covered by blur processing.

[0030] Specifically, to counter common anti-forgery techniques in biological research images, such as "color inversion," "Gaussian blur," and "overexposure," five orthogonal feature modal channels were constructed. For each standard query image in the standard query image set, features under each modality were extracted in parallel, ensuring that regardless of how forgers adjust the image's photometric properties, at least one modality can capture key features, such as... Figure 2 As shown, the specific steps are as follows: S21: Extract Texture Modal Features For the current image in the standard query image set, the SIFT operator is used to extract the basic key points and 128-dimensional descriptors of the image as texture modality features.

[0031] In this embodiment, the contrast threshold of the SIFT operator is set to 0.003 and the edge threshold is set to 20 to adapt to the visual characteristics of low contrast and blurred edges of WB images.

[0032] Perform the above operations independently on each image in the image set.

[0033] S22: Extract Inverted Modal Features A bitwise NOT operation is performed on the current image in the standard query image set to generate an inverted copy. Key points and feature descriptors are then extracted from this copy using the SIFT operator. Subsequently, the feature descriptors are normalized using RootSIFT to serve as the feature representation of this modality.

[0034] This embodiment introduces an inverse color modality specifically designed to identify interchangeable alterations between "black background with white bands" and "white background with black bands." In biological experiments, this represents a change in development mode, but in image plagiarism detection, both essentially contain the same geometric structural information.

[0035] Perform the above operations independently on each image in the image set.

[0036] S23: Extract Ghost Layer Modal Features The current image in the standard query image set is Gaussian blurred and then differs from the original image. The high-frequency residual skeleton is then extracted using morphological opening and closing operations.

[0037] To counter fuzzy attacks, a "ghost layer" was constructed, the specific process of which is as follows: The original image is Gaussian blurred (Kernel Size = 5×5). The difference between the original and blurred images is calculated. High-frequency signals are extracted using adaptive threshold binarization. Then, morphological closing and opening operations are performed sequentially to fill in the gaps within the high-frequency signals and filter out isolated noise points in the background, resulting in a structurally continuous ghost layer binary image. This step strips away the smooth background of the image and is specifically used to capture residual edge traces hidden by the Gaussian blur. The AKAZE operator is then used to extract binary features from this ghost layer binary image.

[0038] Perform the above operations independently on each image in the image set.

[0039] S24: Extract morphological blob modality (Blob_Bright / Dark) features For the current image in the standard query image set, apply the morphological top hat transformation (TopHat) and bottom hat transformation (BlackHat) respectively.

[0040] In this embodiment, the top-hat transform is used to separate bright bands when the background is dark, and the bottom-hat transform is used to separate dark bands when the background is bright. Then, the two transformed images are used as inputs, and the SIFT operator is used to extract key points and feature descriptors of the blob region. The extracted feature descriptors are then normalized by RootSIFT to serve as the final feature representation of this morphological blob modality.

[0041] Perform the above operations independently on each image in the image set.

[0042] S25: Extract dynamic Gamma penetration mode features For the current image in the standard query image set, after applying nonlinear Gamma correction (γ=2.0), the AKAZE operator is used to extract key points and binary feature descriptors on the corrected image, which are then used as the final feature representation of the modality.

[0043] This embodiment uses Gamma nonlinear stretching to reveal texture details that were originally "overexposed" or "underexposed" in extremely dark or extremely bright areas, thereby achieving "penetration" detection of exposure attacks.

[0044] Perform the above operations independently on each image in the image set.

[0045] By performing steps S21 to S25 on each image in the standard query image set, a complete feature set of each image in the multimodal adversarial feature space is obtained, which is used for subsequent feature packaging and pairwise comparison.

[0046] Step S3: Pack each modal feature of each standard query image into a geometric variant tensor and convert it into a high-dimensional feature tensor.

[0047] To combat attacks involving arbitrary rotations and non-uniform stretching, the traditional one-to-one comparison method is abandoned. Instead, the massively parallel computing power of GPUs is utilized to transform the geometric transformation problem into a high-dimensional matrix operation problem. For each standard query image in the standard query image set, the following packaging operation is performed independently: Figure 2 As shown, the specific steps include: S31: For each modal feature extracted from the current image, generate multiple sets of corresponding affine transformation feature variants.

[0048] The preset set of geometric transformation parameters in this embodiment includes: scaling factors covering the range of 0.25x to 4.0x (approximately 16 levels, covering independent stretching of the X and Y axes); flip factors including horizontal flip (H-Flip) and vertical flip (V-Flip); and rotation factors including perturbations of ±5 degrees.

[0049] The above geometric transformation parameters are applied to each modal feature of the current image to generate multiple sets of affine transformation feature variants for each modality.

[0050] S32: Convert the features of all variants of the current image into FP16 half-precision floating-point numbers and stack them in GPU memory as a continuous high-dimensional feature tensor.

[0051] This embodiment utilizes the tensor concatenation operation (such as torch.cat) of a deep learning framework to concatenate the descriptors of all geometric variants belonging to the same image along the 0th dimension in GPU memory. For example, if the current image has 1000 feature points and 20 variants are generated, the dimension of the packed tensor will be approximately 20000×128. This step leverages the features of GPU TensorCores, which not only reduces memory usage but also increases the speed of subsequent distance calculations by orders of magnitude.

[0052] For each image in the standard query image set, the above steps S31 to S32 are performed independently to obtain the high-dimensional feature tensor corresponding to each image, which is used for index set construction and query matching when comparing pairs of images in the subsequent image set.

[0053] Step S4: Based on the high-dimensional feature tensor, perform pairwise asymmetric intelligent routing on the images in the standard query image set to determine the query set and index set in each pair; use GPU-accelerated elastic matrix scanning to calculate all Euclidean distances in a high-throughput manner to filter out the best candidate geometric transformation parameters from the query set to the index set in each pair; perform vector flow consistency cleaning and homography matrix health verification on the candidate geometric transformation parameters to determine whether there are duplicate regions between the query set and the index set in each pair; if duplicates are found, generate a reverse evidence map based on visual normalization to show pixel-level alignment tampering traces.

[0054] Furthermore, the asymmetric intelligent routing includes: using GPU to quickly calculate the bidirectional feature pre-detection scores of the two parties to be compared, and forcing the party with lower feature coverage to be used as the query set and the party with higher feature coverage to be used as the index set, so as to optimize the cache hit rate of matrix multiplication.

[0055] Furthermore, the GPU-accelerated elastic matrix scan includes: stacking the features of all geometric variants belonging to the same image into a continuous high-dimensional feature tensor along the 0th dimension in the GPU memory, calculating the Euclidean distance between all variants and the query set in parallel through a single matrix multiplication (torch.cdist), and then filtering valid matching points through a ratio test.

[0056] Furthermore, the vector flow consistency cleaning includes: calculating the displacement vectors of all matching point pairs, constructing an angle histogram and locking the main peak direction, removing outliers whose direction deviates from the main peak by more than a preset threshold or whose displacement length is abnormal, and ensuring that the remaining matching points follow the same overall motion trend.

[0057] Furthermore, the homography matrix health check includes: projecting the four boundary corner points of the query image onto the target plane through the homography matrix to form a quadrilateral, and detecting whether the quadrilateral satisfies the convex hull constraint, side length ratio constraint, and interior angle constraint; if it does not satisfy these constraints, it is determined to be an invalid distortion; if it does satisfy these constraints, the target image is further mapped back to the source image coordinate system using the inverse matrix, and the inverse color or Gamma parameter is automatically synchronized according to the matching modality to generate a pixel-level aligned red-green complementary color superposition evidence image.

[0058] This step is the core elastic scanning and geometric verification algorithm flow in this embodiment, including an asymmetric matching strategy based on intelligent routing, a GPU matrix-accelerated filtering mechanism, and physical constraint verification logic based on vector flow and polygon health. It is applied to pairwise comparisons of all image pairs in the standard query image set, such as... Figure 3 As shown, specifically: S41: Perform pairwise permutations of all images in the standard query image set, and perform asymmetric intelligent routing based on feature density for each image to be compared, to determine the query set and index set in the comparison pair.

[0059] To reduce redundant computations during pairwise comparisons within the image set, an asymmetric routing strategy was designed to optimize the efficiency of matrix multiplication, including: S411: When performing batch deduplication, the images in the standard query image set are arranged in pairs to generate multiple image pairs to be compared; for any image pair currently being processed (let's say image A and image B), the GPU is used to quickly calculate its bidirectional feature pre-detection score.

[0060] This embodiment calculates the feature coverage rates for A→B and B→A. Specifically, the basic features of image A and the high-dimensional feature tensor of image B are uniformly converted into FP16 half-precision floating-point numbers, and the Euclidean distance matrix between them is calculated in parallel using a GPU. Then, the nearest neighbor and second nearest neighbor distances of each feature point in the high-dimensional feature tensor are obtained and a ratio test is performed. The number of valid matching points that meet the ratio threshold condition is counted, and the maximum number of matching points under each feature mode is taken as the pre-detection score for A→B. Similarly, the reverse score is calculated.

[0061] S412: Compare the bidirectional pre-detection scores, force the locking of the side with lower feature coverage, and use its extracted feature matrix as the query set; lock the side with higher feature coverage, and use its packaged high-dimensional feature tensor (i.e., the feature set containing multiple geometric variations) as the index set.

[0062] This strategy ensures that in each pair, a small, feature-sparse matrix (Query) is always used to collide with a large, feature-dense matrix (Index). By leveraging the broadcast mechanism of matrix multiplication, this asymmetric design significantly improves the hit rate of the GPU cache, avoids unnecessary computational overhead, and maximizes system throughput.

[0063] For each image to be compared in the standard query image set, steps S411 to S412 above are performed independently to determine the respective query set and index set.

[0064] S42: For each pair of images, perform a GPU-accelerated elastic matrix scan to select the best candidate geometric transformation in that pair.

[0065] S421: Calculate the Euclidean distance matrix between the current query set tensor and the index set tensor using matrix multiplication (e.g., torch.cdist).

[0066] Specifically, the high-dimensional feature tensor is essentially formed by concatenating (stacking) the features corresponding to each feature point row by row. In the specific implementation, the feature descriptor tensors of the query set and the index set are loaded into the GPU memory, and the Euclidean distance between each feature in the query set and all features in the index set is solved in parallel through a single matrix operation (a tensor distance calculation instruction based on the L2 norm), thereby generating a fully connected matrix containing distance information of millions of matching pairs in an instant.

[0067] The underlying code for this single matrix operation does not use ordinary matrix dot product, but instead utilizes the tensor distance calculation function (torch.cdist, with p=2) of deep learning frameworks (such as PyTorch). After loading the two feature descriptor tensors, Query and Index, onto the GPU, the system calculates the L2 norm from each row (single descriptor) in Query to all rows (all descriptors) in Index in parallel using a single instruction.

[0068] S422: Perform a ratio test on the distance matrix to count the number of valid matching points under each geometric variant.

[0069] This embodiment performs the filtering in parallel on the GPU: it finds the minimum and second minimum values ​​in each row, and if the ratio of the two is less than a preset threshold (e.g., 0.87), it is determined to be a valid match. It counts the number of valid matches within each geometric variant block, sorts them in descending order, and selects the top three variant parameters as candidates for the current image pair.

[0070] For each image to be compared in the standard query image set, steps S421 to S422 are performed independently to select the best candidate geometric transformation parameters for each image.

[0071] S43: Perform geometric verification with multiple physical constraints on each candidate transformation in the image pair to determine whether there is a duplication between the image pairs.

[0072] To eliminate false positives, strict physical and geometric constraints were introduced to ensure that the matching results conform to real-world physical laws, including: S431: Perform Vector Flow Cleaning on the matching point pairs in the current image pair.

[0073] This embodiment calculates the displacement vectors of all matching point pairs, constructs an angle histogram of 36 intervals, and locks the dominant angle. Outliers whose directions deviate from the dominant angle by more than a preset threshold (e.g., 0.6 radians) and points with abnormal displacement lengths (e.g., greater than 3 times or less than 0.2 times the median length) are removed. This step ensures that all matching points follow the same overall motion trend.

[0074] S432: Calculate the homography matrix using the RANSAC algorithm and perform polygon health detection.

[0075] Using the valid feature matching point pairs retained after vector flow consistency cleaning, the final homography matrix is ​​calculated. Then, the four physical boundary corners of the original query image are extracted and projected onto the target plane using this final homography matrix, forming a quadrilateral. The quadrilateral is checked to ensure it meets the following constraints: 1) Convex hull constraint (must be a convex polygon); 2) Side length ratio constraint (the ratio of the shortest side to the longest side must be greater than 0.05); 3) Angle constraint (all interior angles must be between 30 and 150 degrees). If any of these conditions are not met, it is judged as an "hourglass" or "linear" distortion, indicating that the current feature alignment is a false positive caused by background noise, and is considered an invalid match. Based on this, it is determined that there is no overlapping region between the current image pairs. Conversely, if the homography matrix health check is passed, the match is confirmed to be valid, and it is ultimately determined that there is a real overlapping region between the current image pairs (i.e., image reuse or tampering).

[0076] For each image to be compared in the standard query image set, steps S431 to S432 are performed independently to determine whether the image pair has a duplicate region.

[0077] S44: If it is determined that there is a duplicate between the current image pairs, generate a reverse evidence map based on visual normalization.

[0078] This embodiment aims to generate intuitive and irrefutable visual evidence, solving the problem of evidence images becoming distorted and fragmented due to geometric transformations in traditional algorithms, including: S441: Calculate the inverse matrix of the homography matrix in the current image pair, and use the inverse matrix to reverse map the matched target image region back to the source image coordinate system.

[0079] S442: Based on the modal information during matching, perform visual attribute synchronization on the reverse-mapped image.

[0080] In this embodiment, if the matching is based on the "inverse color mode," the system will automatically perform bitwise inversion on the image after inversion mapping; if it is based on the "Gamma mode," Gamma enhancement will be performed simultaneously. Finally, the source image (as the background) and the processed target image (as the foreground) are superimposed with complementary red and green colors. The overlapping area is displayed in yellow, and the difference area is displayed in red / green separation colors, which visually shows the traces of tampering.

[0081] For all image pairs in the standard query image set that are determined to be duplicates, steps S441 to S442 above are executed independently to generate corresponding reverse evidence images.

[0082] S45: For each image in the standard query image set, perform internal clone detection based on high-frequency fingerprints.

[0083] This embodiment, as a supplement to the pairwise comparison between images described above, also includes detecting "copy-paste" tampering within the same image, including: S451: Use the Maximum Stable Extreme Region (MSER) ​​algorithm to extract connected patch regions in the current image and calculate the Hu moment invariant for each region.

[0084] S452: Filter regions with similar shapes, extract high-frequency noise residuals, and calculate the cross-correlation coefficient of noise fingerprints between regions.

[0085] This embodiment uses Gaussian blur difference to extract and remove the "high-frequency noise residuals" (i.e., sensor thermal noise and background noise) from the main content of the image. The normalized cross-correlation coefficient (NCC) of the noise residuals of two similarly shaped regions is calculated. If the correlation coefficient exceeds a threshold (e.g., 0.78), it is determined to be an internal clone. This is because in naturally captured images, the random noise fingerprints of different regions cannot be completely identical; only through copying and pasting can noise fingerprint duplication occur.

[0086] For each image in the standard query image set, steps S451 to S452 are performed independently to complete the internal clone detection.

[0087] Example 2 One embodiment of the present invention provides a high-throughput identification system for Western Blot images, comprising: The acquisition module is configured to: acquire the Western Blot image set to be deduplicated, and perform adaptive normalization preprocessing on each image in the image set to obtain a standard query image set; The extraction module is configured to: construct a multimodal adversarial feature space and extract features of each standard query image in the standard query image set under different modalities; The packaging module is configured to package each modal feature of each standard query image into a geometric variant tensor and convert it into a high-dimensional feature tensor. The duplicate detection module is configured to: perform pairwise asymmetric intelligent routing on the images in the standard query image set based on the high-dimensional feature tensor to determine the query set and index set in each pair; calculate all Euclidean distances in a high-throughput manner using GPU-accelerated elastic matrix scanning to select the best candidate geometric transformation parameters from the query set to the index set in each pair; perform vector flow consistency cleaning and homography matrix health verification on the candidate geometric transformation parameters to determine whether there are duplicate regions between the query set and the index set in each pair; if duplicates are determined to exist, generate a reverse evidence map based on visual normalization to show pixel-level aligned tampering traces.

[0088] Example 3 One embodiment of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned high-throughput identification method for Western Blot images.

[0089] Example 4 In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned Western Blot image high-throughput identification method.

[0090] Example 5 One embodiment of the present invention provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the aforementioned Western Blot image high-throughput identification method.

[0091] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0093] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A high-throughput identification method for Western Blot images, characterized in that, include: Obtain the set of Western Blot images to be checked for deduplication, and perform adaptive normalization preprocessing on each image in the set to obtain the standard query image set; Construct a multimodal adversarial feature space and extract features of each standard query image in the standard query image set under different modalities; Each modal feature of each standard query image is packaged into a geometric variant tensor and then converted into a high-dimensional feature tensor. Based on the high-dimensional feature tensor, pairwise asymmetric intelligent routing is performed on the images in the standard query image set to determine the query set and index set in each pair; through GPU-accelerated elastic matrix scanning, all Euclidean distances are calculated in a high-throughput manner at once to select the best candidate geometric transformation parameters from the query set to the index set in each pair. The candidate geometric transformation parameters are subjected to vector flow consistency cleaning and homography matrix health verification to determine whether there are duplicate regions between the query set and index set in each pair. If duplication is detected, a reverse evidence image based on visual normalization is generated, showing pixel-level aligned tampering traces.

2. The high-throughput identification method for Western Blot images as described in claim 1, characterized in that, The multimodal adversarial feature space includes texture mode, inverted color mode, ghost layer mode, morphological speckle mode, and dynamic Gamma penetration mode; The ghost layer mode extracts high-frequency residual skeletons through Gaussian blur and difference from the original image, adaptive threshold binarization, and morphological opening and closing operations, which are used to detect edge traces covered by blurring.

3. The high-throughput identification method for Western Blot images as described in claim 1, characterized in that, The asymmetric intelligent routing includes: using GPU to quickly calculate the bidirectional feature pre-detection scores of the two parties to be compared, and forcing the party with lower feature coverage to be used as the query set and the party with higher feature coverage to be used as the index set, in order to optimize the cache hit rate of matrix multiplication.

4. The high-throughput identification method for Western Blot images as described in claim 1, characterized in that, The GPU-accelerated elastic matrix scan includes: stacking the features of all geometric variants belonging to the same image into a continuous high-dimensional feature tensor along the 0th dimension in the GPU memory; calculating the Euclidean distance between all variants and the query set in parallel through a single matrix multiplication (torch.cdist); and then filtering valid matching points through a ratio test.

5. The high-throughput identification method for Western Blot images as described in claim 1, characterized in that, The vector flow consistency cleaning includes: calculating the displacement vectors of all matching point pairs, constructing an angle histogram and locking the main peak direction, removing outliers whose direction deviates from the main peak by more than a preset threshold or whose displacement length is abnormal, and ensuring that the remaining matching points follow the same overall motion trend.

6. The high-throughput identification method for Western Blot images as described in claim 1, characterized in that, The homography matrix health check includes: projecting the four boundary corners of the standard query image onto the target plane through the homography matrix to form a quadrilateral, and checking whether the quadrilateral satisfies the convex hull constraint, side length ratio constraint, and interior angle constraint; if it does not satisfy these constraints, it is determined to be an invalid distortion; if it does satisfy these constraints, the target image is further mapped back to the source image coordinate system using the inverse matrix, and the inverse color or Gamma parameter is automatically synchronized according to the matching modality to generate a pixel-level aligned red-green complementary color superposition evidence image.

7. A high-throughput identification system for Western blotting images, characterized in that, include: The acquisition module is configured to: acquire the Western Blot image set to be deduplicated, and perform adaptive normalization preprocessing on each image in the image set to obtain a standard query image set; The extraction module is configured to: construct a multimodal adversarial feature space and extract features of each standard query image in the standard query image set under different modalities; The packaging module is configured to package each modal feature of each standard query image into a geometric variant tensor and convert it into a high-dimensional feature tensor. The duplicate detection module is configured to: perform pairwise asymmetric intelligent routing on the images in the standard query image set based on the high-dimensional feature tensor to determine the query set and index set in each pair; and calculate all Euclidean distances in a high-throughput manner by using GPU-accelerated elastic matrix scanning to select the best candidate geometric transformation parameters from the query set to the index set in each pair. The candidate geometric transformation parameters are subjected to vector flow consistency cleaning and homography matrix health verification to determine whether there are duplicate regions between the query set and index set in each pair. If duplication is detected, a reverse evidence image based on visual normalization is generated, showing pixel-level aligned tampering traces.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements a high-throughput identification method for Western Blot images as described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement a Western Blot image high-throughput identification method as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform a Western Blot image high-throughput identification method as described in any one of claims 1-6.

Images