A scientific research image repetition intelligent detection method based on an AI platform

Abstract

The application relates to the technical field of scientific research image processing, and specifically discloses a scientific research image repeatability intelligent detection method based on an AI platform. The AI platform is used to establish a published scientific research paper sample set, extract scientific research images of each paper, and separate background areas. Noise points in the background areas are detected, and noise point position coordinates are obtained. A noise point distribution topology graph is constructed, and an image physical fingerprint vector is generated. A large language model is used to extract context information from sample texts and encode the context information into a context semantic vector. The physical fingerprint vector and the context semantic vector are summarized to establish a scientific research image feature reference database. Physical fingerprint vectors, context semantic vectors and background area quality scores of to-be-detected papers are extracted, compared with the reference database, and background noise point similarity and context information difference are calculated. The contradiction degree score is calculated by comprehensively considering the two, the judgment standard is automatically adjusted according to the background area quality score, and a visual detection report is generated for images exceeding the judgment standard.

Description

Technical Field

[0001] This invention belongs to the field of scientific image processing technology, and relates to an intelligent detection method for the repetition of scientific images based on an AI platform. Background Technology

[0002] Scientific images are crucial carriers of experimental data, observational results, and model outputs in academic papers, and are widely used in disciplines such as materials science, chemistry, biology, medicine, and physics. To ensure the credibility of research findings and the integrity of academic records, it is necessary to perform duplication detection on scientific images to identify the repeated use of the same image within a single paper and across different papers.

[0003] Mainstream methods for detecting image duplication in scientific research focus on the main content region of the image, namely the visual representation of experimental data and observation results. In the preprocessing stage, these methods typically perform denoising, filtering, and background subtraction on non-subject areas, treating them as useless interference and filtering them out, retaining only the main subject area for subsequent feature extraction and similarity comparison. This approach means that existing technologies systematically discard image identity information that the background area might carry.

[0004] Furthermore, existing methods only compare visual features between images, completely disregarding the contextual information associated with them. When the main body of an image changes due to operations such as cropping or scaling, detection methods relying solely on main body region features struggle to effectively identify the common origin relationship between images. Simultaneously, for detected similar image pairs, existing methods cannot incorporate image association information to provide auxiliary judgment criteria. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention provides an intelligent detection method for the repetition of scientific research images based on an AI platform, which is used to solve the above-mentioned technical problems.

[0006] To achieve the above and other objectives, the technical solution adopted by the present invention is as follows:

[0007] This invention provides an intelligent method for detecting the repetition of scientific research images based on an AI platform. The method includes the following steps:

[0008] Step S1: Establish a sample set of published research papers, extract all research images from each paper in the sample set, perform format unification and type recognition processing on each image, separate the main content in the image, exclude non-background elements, and obtain the image background area.

[0009] Step S2: Noise detection is performed on the background region of each image to obtain the position coordinates of each noise point; a noise distribution topology map in the form of a Delaunay triangulation is constructed using these coordinates as vertices, and the side length, area, and angle distribution of each triangle in the triangulation, as well as the adjacency relationship between each vertex, are recorded; graph structure encoding is performed on the above topology map to generate a fixed-length image physical fingerprint vector;

[0010] Step S3: Using the large language model built into the AI ​​platform, extract five types of contextual information corresponding to each image from the paper sample text and encode them into a fixed-length contextual semantic vector.

[0011] Step S4: Summarize the physical fingerprint vectors and contextual semantic vectors of all images in the sample set to establish a scientific research image feature reference database;

[0012] Step S5: Input the paper to be tested, extract the physical fingerprint vector and context semantic vector of each image in the paper to be tested in the manner described, and calculate the background region quality score; compare the physical fingerprint vector and context semantic vector of the image of the paper to be tested with the data in the reference database one by one, and calculate the background noise similarity and context information difference.

[0013] Step S6: Calculate the degree of contradiction by combining the similarity of background noise in the image of the paper to be tested and the difference in context; adjust the judgment criteria according to the background region quality score, and generate a visual detection report.

[0014] Another aspect of the present invention provides an intelligent detection device for the repetition of scientific research images based on an AI platform, including a processor, a memory, and a communication bus;

[0015] The memory stores a computer-readable program that can be executed by the processor;

[0016] The communication bus enables communication between the processor and the memory;

[0017] When the processor executes the computer-readable program, it performs steps in the AI-based intelligent detection method for scientific image repetition as described in any one of the present invention.

[0018] As described above, the intelligent detection method for scientific image repetition based on an AI platform provided by the present invention has at least the following beneficial effects:

[0019] This invention performs image recognition on scientific images in research papers, automatically identifies different types of scientific images, and separates the main content from the background area according to the image type. This enables the subsequent detection process to extract features from the background area of ​​the scientific image, avoiding the detection limitations caused by judging the repetition rate solely based on the main content of the image, and improving the ability to acquire effective features in the process of detecting repetition of scientific images.

[0020] This invention processes the background region of an image to extract the location and distribution features of background noise, and constructs a spatial topological structure and image physical fingerprint based on the background noise, thereby achieving a quantitative characterization of the physical features of scientific research images. Even when the main image region has been cropped, scaled, or its brightness adjusted, the distribution features of the background noise can still be used to assist in determining the physical homology between images, thus improving the ability to identify hidden, repeating images.

[0021] This invention further utilizes an AI platform to perform semantic recognition on the corresponding paper text of scientific research images, extracting contextual information such as processing conditions, parameter indicators, experimental environment, research objects, and time parameters, and forming contextual semantic features. During the detection process, the similarity of background noise obtained from image processing and the difference in contextual information obtained from semantic recognition are jointly analyzed. When the background noise of two images is highly similar but the contextual information is significantly different, the corresponding contradiction level is increased, enabling the identification and alarm of abnormal repetition cases where "physical images are highly homogeneous but experimental semantics are inconsistent," thereby improving the accuracy and pertinence of scientific research image duplication detection. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the steps in the method of the present invention.

[0024] Figure 2 This is a schematic diagram of the noise topology graph construction and physical fingerprint generation logic of the present invention.

[0025] Figure 3 This is a schematic diagram illustrating the logic of similarity calculation and contradiction determination in this invention.

[0026] Figure 4 This is a system architecture diagram of the present invention. Detailed Implementation

[0027] The following description, in conjunction with the implementation of this invention, is merely an example and illustration of the concept of this invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention.

[0028] The proprietary English terms involved in this invention and their Chinese explanations are as follows:

[0029] Delaunay triangulation: a set of connected but non-overlapping triangles, the circumcircles of which do not contain any other points of the region;

[0030] Bowyer-Watson point-by-point insertion algorithm: It is a point-by-point insertion method that first generates an initial grid from a given set of points, and then adds points to the grid one by one according to the Delaunay partitioning principle, and reconnects them to generate a new grid, until all points have been added;

[0031] CRAFT Text Detector: A PyTorch-based text detector that can effectively detect text regions in images;

[0032] Canny edge detection algorithm: a standard algorithm widely used in edge detection, whose goal is to find an optimal edge detection solution or to find the location with the strongest gray-level intensity change in an image;

[0033] Two-Pass algorithm: A method in image processing used to find and label connected regions in an image;

[0034] BIO ternary tagging system: a three-dimensional sequence tagging method, where B indicates that a word is the beginning of an entity, I indicates that a word is the middle part of an entity, and O indicates that a word does not belong to any entity;

[0035] BERT, a pre-trained language model, is a Transformer-based pre-trained language model proposed by Google in 2018. Its core innovation lies in deep bidirectional training using unlabeled text through masked language modeling and next-sentence prediction tasks, enabling the model to simultaneously understand the contextual information on both sides of a word.

[0036] The softmax function is a normalization exponential function that can compress a vector containing arbitrary real numbers into another real vector of the same dimension, such that each element is in the range (0,1) and the sum of all elements is 1.

[0037] Jaccard similarity coefficient: A statistical index used to measure the similarity between two sets. It is defined as the ratio of the size of the intersection to the size of the union of the two sets and is used to characterize the dissimilarity between the sets.

[0038] Example 1:

[0039] Please see Figure 1-3 As shown, a method for intelligent detection of repetition in scientific research images based on an AI platform includes:

[0040] Step S1: Establish a sample set of published research papers, extract all research images from each paper in the sample set, perform format unification and type recognition processing on each image, separate the main content in the image, exclude non-background elements, and obtain the image background area.

[0041] In one specific embodiment, a sample set of published scientific papers is established, covering journal articles from multiple disciplines including biology, medicine, chemistry, materials science, and physics. The papers originate from publicly available databases such as PubMedCentral, arXiv, and journal websites, and are in PDF format. The AI ​​platform uses its built-in document parsing engine to structurally analyze each PDF paper, extracting all embedded image files, the corresponding captions for each image, and the method description text for the section containing the image.

[0042] In this embodiment, the paper sample data comes from publicly authorized datasets, open-access paper resources, and legally authorized scientific image data. When constructing the reference database, complete paper image files are not stored directly; instead, only image feature information for duplication detection is extracted, including background noise physical fingerprint vectors, context semantic vectors, and related index information, thereby avoiding long-term storage of the original paper image content.

[0043] Each extracted scientific image underwent a standardized format processing. Images in different compression formats such as PNG, JPG, and TIFF were uniformly converted to BMP lossless bitmap format. During the conversion process, lossy compression was avoided to preserve the complete distribution characteristics of the original background noise. Color images were converted to grayscale images using the following formula:

[0044] ;

[0045] Where R(i,j), G(i,j), and B(i,j) are the brightness values ​​of the pixel at coordinate (i,j) in the red, green, and blue channels, respectively, with grayscale as the dimension, derived from the original photoelectric conversion values ​​of the image sensor; I(i,j) is the converted grayscale value, with grayscale as the dimension; the weighting coefficients 0.299, 0.587, and 0.114 are derived from the weighting of the human eye's sensitivity to different wavelengths of light as defined in the ITU-T BT.601 standard; the function of this formula is to synthesize the three-channel color information into single-channel grayscale information, reducing the computational complexity of subsequent processing while preserving the image structural features.

[0046] This embodiment does not perform spatial scaling on the scientific research images, preserving the pixel matrix in the original image file to avoid altering the fixed-pattern noise characteristics generated by the imaging device through interpolation, resampling, or other operations. Addressing the resolution differences between scientific research images from different sources, in the subsequent noise feature extraction stage, only the position coordinates of effective noise points are proportionally normalized, mapping the pixel coordinates to a dimensionless unified coordinate space. This processing method does not change the grayscale value of the corresponding pixel of the noise, but only changes the representation of the spatial location of the noise, enabling the spatial distribution topology of noise in images of different resolutions to be compared at a uniform scale.

[0047] The system performs type recognition processing on standardized images. Based on the pixel distribution characteristics of the images, the system classifies scientific images into four categories: strip images, microscopic images, scatter images, and cloud-like images. Type recognition uses a lightweight image classification model built into the AI ​​platform. The input is the standardized grayscale image, and the output is the probability distribution of the four categories. The category with the highest probability is taken as the image type label. The role of this classifier is to provide type information for subsequent subject-to-background separation, enabling the system to automatically select an appropriate separation strategy.

[0048] This embodiment employs a lightweight convolutional neural network-based image classification model to identify the type of input scientific research images. The classification model uses the MobileNet-V2 network structure, reducing the number of model parameters and computational complexity through depthwise separable convolutions. The model includes multiple inverse residual modules, each consisting of pointwise convolutions, depthwise convolutions, and linear pointwise convolutions, used to extract multi-scale spatial texture and structural features from the input scientific research images.

[0049] The image classification model takes a 224×224 pixel scientific image as input and outputs a probability distribution for the corresponding image category. The model is trained using a self-built scientific image classification dataset, which includes four types of scientific images: strip, microscopic, scatter, and cloud image, containing approximately 20,000 sample images. The training, validation, and test sets are divided in an 8:1:1 ratio.

[0050] During model training, the AdamW optimizer was used to iteratively update the network parameters. The cross-entropy loss function was used to measure the difference between the model's predicted class probability distribution and the true class label. The training batch size was set to 32, and the initial learning rate was set to... The training rounds were set to 50. The model outputs the predicted probability for each category and determines the category with the highest probability as the category label for the current scientific image, so as to enable differentiated processing for different types of images in the future.

[0051] After completing the type recognition, the corresponding subject and background separation strategy is selected according to the image type.

[0052] For strip-type images, grayscale projection is used for separation: the average grayscale value of each row of pixels is calculated in the horizontal direction to obtain the horizontal projection curve.

[0053] ;

[0054] Where W is the image width, measured in pixels, derived from image size information; I(x,y) is the image pixel grayscale value, measured in grayscale levels; H(y) is the average grayscale value of the y-th row, measured in grayscale levels. This represents the summation of all pixels in the y-th row. The mean grayscale value of each column of pixels is calculated vertically to obtain the vertical projection curve:

[0055] ;

[0056] Where H is the image height, measured in pixels. This represents the summation of all pixels in column x. The striped regions exhibit distinct peaks and troughs on the projection curve. The strip boundary positions are determined by detecting the zero-crossing points of the first derivative of the projection curve. The area between adjacent boundaries is taken as the main image region, and the rest as the image background region. The purpose of these two formulas is to compress the grayscale distribution of a two-dimensional image into a one-dimensional projection curve, enabling rapid location of the stripes through one-dimensional signal analysis.

[0057] For microscopic images, Otsu's thresholding method combined with morphological operations is used for separation: Inter-class variance is calculated.

[0058] ;

[0059] Where t is the candidate segmentation threshold, in grayscale; ω1(t) is the ratio of foreground pixels to total pixels, in dimensionless proportion; ω2(t) is the ratio of background pixels to total pixels, in dimensionless proportion; μ1(t) is the mean grayscale value of foreground pixels, in grayscale; μ2(t) is the mean grayscale value of background pixels, in grayscale; and σ²(t) is the inter-class variance, in squares of grayscale. Iterating through t from 0 to 255, the value of t that maximizes σ²(t) is taken as the optimal segmentation threshold. This formula determines the best segmentation point by maximizing the grayscale difference between the foreground and background. Pixels with grayscale values ​​greater than t are classified as the main image region, and pixels with grayscale values ​​less than or equal to t are classified as the background region. Morphological opening and closing operations are then performed sequentially on the segmented binary image. The opening operation is an erosion-dilation operation, meaning the segmented binary image is first eroded using the structuring element SE, and then the eroded result is dilated using the same structuring element SE. The closing operation is a dilation-erosion operation, meaning the segmented binary image is first dilated using the structuring element SE, and then the dilated result is eroded using the same structuring element SE. Here, SE is a 3×3 pixel circular structuring element. The opening operation removes isolated noise points from the binary image, while the closing operation fills in small holes within the main image region.

[0060] For scattered images, the Hough transform is used to detect coordinate bounding boxes: the image edge detection results are mapped to Hough space, and the equation of the straight line is expressed as:

[0061] ;

[0062] Where (x, y) are the pixel coordinates of the image edge points, with units of pixels; θ hough Let be the angle between the normal to the line containing the pixel and the x-axis, in radians, ranging from 0 to π; ρ is the distance from the origin to the line containing the pixel, in pixels. The cumulative votes for each cell in the Hough space are counted, and the two horizontal and two vertical lines with the highest vote counts form the bounding box of the coordinate axes. The area inside the box is the main image region, and the area outside the box is the image background region. This formula transforms the line detection problem in image space into a peak detection problem in Hough space, providing robust detection capabilities for discontinuous or partially occluded coordinate axes.

[0063] For cloud-like images, the boundaries of the data display area are determined using a color scale and legend region detection: First, the Canny edge detection algorithm is used to extract the image edges, and the gradient magnitude is calculated using the following formula:

[0064] ;

[0065] Where i represents the row number of the image pixel, i.e., the vertical coordinate, ranging from 1 to H, where H is the height of the image, measured in pixels. j represents the column number of the pixel, i.e., the horizontal coordinate, ranging from 1 to W, where W is the width of the image, measured in pixels. G(i,j) represents the gradient magnitude of the image at the pixel in the i-th row and j-th column. x (i,j) and G y (i, j) represent the Sobel gradient components of the pixel in the horizontal and vertical directions, respectively. The units of the four gradient components are all gray levels per pixel, representing the change in gray level per unit pixel distance. The formula for calculating the gradient direction is:

[0066] ;

[0067] Where θ(i,j) is the gradient direction of the pixel, and its dimension is radians.

[0068] The Canny edge detection algorithm uses the two parameters mentioned above for edge extraction as follows: First, it calculates the gradient magnitude G(i,j) and gradient direction θ(i,j) for each pixel. Second, it performs non-maximum suppression along the gradient direction of each pixel, that is, it checks whether the gradient magnitude of the pixel is the maximum value in its neighborhood along the gradient direction. If it is, it is retained; otherwise, it is suppressed to zero, thus refining the blurry edge into a sharp edge with a width of one pixel. Then, it sets a high threshold T. high and low threshold T low T high and T low The method for determining T is an automatic setting method based on the gradient magnitude histogram. First, the gradient magnitude distribution histogram of all pixels in the entire image is statistically analyzed, and then the cumulative distribution function of the gradient magnitude is calculated. high The gradient magnitude at the 70th percentile of the cumulative distribution function is taken, which is the minimum gradient value of the top 30% of pixels with the largest gradient magnitude in the image. This value is determined because true edges in scientific images are usually located in the high quantile range of the gradient magnitude distribution; choosing the 70th percentile strikes a balance between capturing subject edges and suppressing weak edges. low Take T high The ratio is determined based on the fact that in the original settings of the Canny algorithm, the ratio of the high threshold to the low threshold is usually 2:1. A 2:1 ratio means that half of the edges can be effectively connected. If the ratio is higher than this, weak edges will not be connected enough, resulting in broken edges. If the ratio is lower than this, too many noisy edges will be preserved, leading to an increase in false detections. The units of both thresholds are gray levels per pixel, consistent with the units of the gradient magnitude.

[0069] Gradient magnitude higher than T high Pixels marked as strong edge points are directly retained, while those with gradient magnitudes lower than T are...low Pixels with gradient values ​​between the two are marked as non-edge points and discarded directly. Pixels with gradient values ​​between the two are marked as weak edge points, and are only retained if a weak edge point is adjacent to a strong edge point; otherwise, they are discarded. The purpose of the above steps is to extract edge pixels with clear boundaries from the image while suppressing noise and non-edge regions. Connectivity analysis is performed on the edge detection results, that is, adjacent edge pixels are grouped into the same connected region, the minimum bounding rectangle of each connected region is calculated, and the rectangular regions containing the color scale and legend are selected based on the size and position of the rectangles. The outer boundaries of these rectangular regions are taken as the boundaries of the data display area.

[0070] In cloud-map images, the color scale is typically a long, narrow rectangle located on one side or bottom of the image, containing a color gradient bar from smallest to largest and corresponding numerical labels. The legend is a rectangle located in a corner of the image, containing different colors or patterns and their corresponding explanations. Rectangles are selected based on their aspect ratio, area, and position: the color scale rectangle typically has an aspect ratio greater than 3 and less than 0.33, an area occupying 2% to 10% of the total image area, and is located near the right or bottom edge of the image; the legend rectangle occupies 1% to 5% of the total image area and is located near the four corners of the image. Rectangles meeting these criteria are designated as the color scale and legend areas. The boundary line of these rectangles closest to the image center is used as the dividing line between the data display area and the non-data area. The area containing the color gradient cloud map inside the dividing line is marked as the main image area, and the blank area outside the dividing line, as well as the color scale and legend areas themselves, are marked as the image background area.

[0071] After separating the subject from the background, non-background elements in the background area are identified and excluded. The scale bar is identified using a slender rectangular shape feature matching method. The detection conditions are that the rectangle's aspect ratio is greater than 5 and its area is between 0.01% and 0.5% of the total image area. The aspect ratio is dimensionless, and the area dimension is pixel square. The aspect ratio threshold of 5 is determined based on the fact that scale bars in scientific images typically appear as extremely narrow and long rectangles. If the threshold is set too low, other rectangular elements may be misidentified as scale bars; if it is set too high, larger scale bars may be missed. The lower limit of 0.01% and the upper limit of 0.5% are determined based on the statistical experience that scale bars usually occupy a small but not excessively small area in the image. The lower limit excludes small noise elements from being misidentified, and the upper limit excludes large elements such as coordinate axes and graph frames from being misidentified.

[0072] The text annotation recognition employs a CRAFT text detector, outputting the bounding box coordinates of each text region in pixels. The CRAFT text detector uses VGG16 as its backbone network, takes a grayscale image as input, and outputs the character region score and character connectivity score for each pixel. After thresholding, the bounding box coordinates of each text region are output. This detector is pre-trained on the SynthText dataset and fine-tuned on the ICDAR2015 dataset, with a character region score threshold of 0.7.

[0073] The identification of arrow guide lines employs Hough transform line segment detection, clustering detected line segments according to length and angle. Length is measured in pixels, and angle in radians. During clustering, a length threshold is set to 2% of the image diagonal length; segments shorter than this are considered noise and discarded. This threshold is determined based on the assumption that arrow guide lines typically possess sufficient visual saliency, and excessively short segments may be accidental edge fragments. For angle clustering, segments with similar directions are grouped together, with the angle difference within each group not exceeding π / 12 radians (15 degrees). This threshold is based on the premise that arrows in the same image usually have consistent directions or are at right angles.

[0074] The identification of frame boundary lines uses edge line detection, selecting straight line segments at the four boundaries of the image whose length exceeds 80% of the image width and height. This 80% threshold is determined based on the fact that frame boundary lines usually run through the entire edge of the image. If the threshold is set too low, table lines or dividing lines inside the image may be misidentified as frame lines; if it is set too high, the boundary lines may be missed due to annotations.

[0075] The identified interference element regions are uniformly marked and excluded from the background region, resulting in a clean background region.

[0076] Simultaneously, a background area quality score is calculated using the following formula:

[0077] ;

[0078] Among them, R area σ represents the ratio of the background area to the total image area. It is a dimensionless ratio derived from the ratio of the number of pixels in the background area to the total number of pixels in the image. gray σ represents the standard deviation of the grayscale values ​​of the background pixels, measured in grayscale levels. It is calculated statistically from the grayscale values ​​of each pixel in the background region of the image and reflects the uniformity of the background grayscale. max This is a reference value for the normalized standard deviation of grayscale, with the dimension being grayscale levels. It is taken as half of the grayscale range of 256, i.e., 128, as the reference benchmark. This value is determined based on the theoretical maximum standard deviation of 128 for an 8-bit grayscale image. gray A value close to 128 indicates extremely uneven background grayscale distribution, with uniformity approaching 0; Nnoise To determine if the response value of the background region of an image exceeds a preset threshold T after applying a 3×3 Laplacian high-pass filter. lap The number of pixels, measured in pixels, with a preset threshold T. lap The threshold value is set to 10, with the dimension being grayscale. This preset threshold is determined based on the fact that the Laplacian response value of background noise is typically between 10 and 50; signals below 10 may be weak fluctuations in a uniform background rather than true noise. ideal The ideal noise level reference value is set to a constant 500, with the unit being the number of pixels. This value is determined based on historical statistical results for different types of scientific research images. Its purpose is to measure the deviation between the effective noise level and the typical noise level in the current background region. Too few noise points result in insufficient nodes in the topological graph, leading to decreased fingerprint discrimination; too many noise points may contain a large amount of interference noise introduced by non-hardware factors. JPEG γ1 represents the JPEG compression artifact evaluation value, a dimensionless ratio calculated as the ratio of the gray-level jump amplitude at the boundary of an 8×8 pixel block to the gray-level fluctuation amplitude within the block. 0 indicates no compression, and 1 indicates severe compression. JPEG compression divides the image into 8×8 pixel blocks and performs discrete cosine transform; gray-level discontinuities at the block boundaries are typical artifacts of lossy compression. γ1, γ2, γ3, and γ4 are weighting coefficients for each indicator, each defaulting to 0.25, with dimensionless weights. The equal weighting is determined based on the fact that the four indicators have roughly equal impact on background quality in practical applications. Q is the overall quality score, also dimensionless. A higher score indicates better background quality and more reliable physical fingerprint extraction. The background quality scoring formula is designed because background quality is influenced by multiple factors, and a single indicator cannot comprehensively measure its suitability for physical fingerprint extraction. Therefore, a comprehensive evaluation is needed from four dimensions: area, uniformity, noise quantity, and compression artifacts. The formula employs a four-factor weighted summation structure, allowing for individual control and adjustment of each factor's contribution to the final score. All four components are normalized to a range of 0 to 1, ensuring that each dimension has the same numerical scale during weighted summation and preventing any single dimension from dominating the score due to dimensional differences. The design of equal weighting of the four components is based on the fact that, in practical applications, the four dimensions contribute roughly equally to background quality, eliminating the need to particularly emphasize any one dimension. When a dimension exhibits significant defects, its corresponding component score is low, lowering the overall score and ensuring that the quality score promptly reflects problems in the background area. This score is used in subsequent step S6 to dynamically adjust the judgment threshold. When quality is poor, the judgment threshold is automatically raised to avoid misjudgments due to insufficient noise information; when quality is good, the judgment threshold is automatically lowered to improve detection sensitivity, forming a closed-loop adaptive feedback mechanism.

[0079] When R area Below 0.1, C JPEGWhen the R value is higher than 0.8, the image is marked as a low-quality image, and the confidence level of its subsequent detection results will be reduced accordingly. area The threshold of 0.1 is determined based on the fact that when the background area accounts for less than 10%, the number of extractable noise points is too small, and the number of nodes in the topology graph is insufficient to form discriminative structural features. JPEG The threshold of 0.8 is determined because when the compression mark is higher than 0.8, the original noise distribution has been severely damaged by lossy compression, the extracted physical fingerprint no longer reflects the hardware characteristics of the imaging device, and the reliability of the detection results is significantly reduced.

[0080] Step S2: Perform noise detection on the background region of each image, obtain the position coordinates of each noise point, and convert the pixel coordinates into normalized spatial coordinates according to the image width and height; construct a noise distribution topology map in the form of a Delaunay triangulation using the normalized spatial coordinates as vertices, record the side length, area and angle distribution of each triangle in the triangulation and the adjacency relationship between each vertex, perform graph structure encoding on the above topology map, and generate a fixed-length image physical fingerprint vector.

[0081] During the manufacturing process of image sensors, the photoelectric conversion efficiency of each pixel unit exhibits nanometer-level process deviations, resulting in fixed-pattern noise; power supply ripple generates pixel-level response fluctuations in the sensor readout circuit. The spatial distribution patterns of these two types of noise are determined by the atomic-level random doping distribution and parasitic parameters of circuit wiring during the sensor chip manufacturing process, and cannot be reproduced through post-processing or transmitted through simple image copying. Therefore, the spatial distribution of background noise can serve as a physical fingerprint at the hardware level of the imaging device.

[0082] Preferably, the step of performing noise detection on the background region of each image and obtaining the position coordinates of each noise point includes the following specific steps:

[0083] In one specific embodiment, noise detection employs the difference of Gaussian method, first analyzing the background region image I. bg Gaussian filtering is performed to obtain a smooth background image I smooth The filtering formula is:

[0084] ;

[0085] Among them I bg (i,j) represents the grayscale value of the original background region of the image, with the dimension being grayscale level; smooth (i,j) represents the grayscale value of the background region of the smoothed image, with the dimension being grayscale levels. L(u,v) is a two-dimensional Gaussian kernel function, expressed as:

[0086] ;

[0087] Where σ is the standard deviation of the Gaussian kernel, with a value of 1.0 and a unit of pixels. The basis for determining σ to be 1.0 is that the typical spatial scale of background noise is about 1 to 3 pixels. If σ is too large, it will blur small noise and lead to missed detections; if σ is too small, it will not be able to effectively smooth the background texture and lead to false detections. u and v are the offset coordinates of pixels within the Gaussian kernel relative to the kernel center, with a unit of pixels. Both of them are integers from -2 to 2, corresponding to a 5×5 filtering window. Specifically, u represents the horizontal offset of a pixel within the kernel relative to the kernel center. u=-2 represents the second pixel to the left of the kernel center, u=0 represents the kernel center itself, and u=2 represents the second pixel to the right of the kernel center. v represents the vertical offset of a pixel within the kernel relative to the kernel center. v=-2 represents the second pixel above the kernel center, v=0 represents the kernel center itself, and v=2 represents the second pixel below the kernel center. The 5×5 kernel window size is determined because when σ=1.0, the weights of the Gaussian function outside the 2σ range (i.e., more than 2 pixels from the center) have decayed to approximately 1.8% of the center weights, and their contribution to the filtering result is negligible. Therefore, a window with a radius of 2σ can completely cover the effective weight distribution of the Gaussian kernel, while avoiding unnecessary computation by using an excessively large window. The original background image and the smoothed background image are then subjected to pixel-by-pixel interpolation, and the residual signal is extracted as the noise signal. The formula is as follows:

[0088] ;

[0089] Where N(i,j) is the image noise signal value, with grayscale as the unit. A positive value indicates that the pixel is brighter than the surrounding background, and a negative value indicates that the pixel is darker than the surrounding background. The function of this formula is to eliminate the slow gradient trend of the background by difference, retaining only the pixel-level fast fluctuation component, i.e., noise.

[0090] Adaptive threshold binarization is performed on the noisy signal N(i,j) to convert it into a binary image double(i,j). The binarization formula is: when |N(i,j)|≥T adapt When |N(i,j)| is true, double(i,j) = 1; when |N(i,j)| is false, double(i,j) = 1. <T adapt When double(i,j)=0. Where T adapt The adaptive threshold is calculated using the following formula:

[0091] ;

[0092] μ N Let σ be the global mean of the image noise signal N(i,j), with dimensions in gray levels. NLet be the global standard deviation of the noise signal N(i,j), with the dimension of gray level. The coefficient 2 is determined based on the fact that in a Gaussian distribution, the probability of a value falling outside the mean by more than or equal to two standard deviations is approximately 4.5%. Classifying these significantly deviating pixels as noise achieves a balance between capturing real noise and suppressing random fluctuations.

[0093] Connectivity analysis is performed on the binary image double(i,j) using a Two-Pass algorithm. This algorithm assigns temporary labels in the first scan and merges equivalent labels in the second scan, marking regions where all adjacent pixels have a value of 1. In the first scan, the image is traversed row by row. For each pixel with a value of 1, its left and top neighbors are checked. If a neighboring pixel is already labeled, the current pixel is assigned the same label; otherwise, a new label is assigned. In the second scan, equivalent labels are merged, unifying pixels belonging to the same connected region but assigned the same label. The following features are extracted for each connected region: centroid coordinates (x,j). c ,y c The spatial coordinates of this noise point are calculated using the following formula:

[0094] ;

[0095] Where A is the area of ​​the connected region, measured in pixels squared, and calculated from the number of pixels contained in the connected region, x p and y p Let be the coordinates of the p-th pixel within the connected component, with units of pixels. This represents the summation over all pixels within the connected region. The centroid coordinate formula uses the arithmetic mean of the coordinates of all pixels within the connected region as the representative location of image noise, achieving sub-pixel level positioning accuracy.

[0096] The detected image noise is filtered out, removing abnormal spots caused by dust particles and image defects, while retaining effective noise generated by the hardware characteristics of the imaging device. The filtering criteria are as follows:

[0097] A min ≤A≤A max ,

[0098] Where A is the noise area, A min The first preset threshold, i.e., the lower limit of the area, has a value of 2 and is measured in pixels squared. max The second preset threshold, i.e., the upper limit of the area, is set to 0.05% of the total number of pixels in the image, with the dimension being the square of pixels. A min The basis for selecting 2 is that connected components with an area less than 2 pixels are usually single-pixel random thermal noise fluctuations, which are inherent characteristics of the sensor's dark current and should be retained rather than filtered out; A maxThe determination of 0.05% of the total image pixels is based on the fact that when the area of ​​a connected component exceeds this threshold, it exceeds the typical size range of hardware noise in the imaging device. This is highly likely to be abnormal spots introduced by external factors such as dust, bubbles, or staining solution deposits, and should be filtered out. Connected components with an area A less than the first preset threshold A are considered to be... min The noise is identified as sensor thermal noise fluctuation. This type of noise usually manifests as tiny bright spots of a single or dual pixel. It is generated by the random fluctuation of dark current during the imaging process of CCD and CMOS sensors. It is an inherent feature of the imaging device hardware. Although the area is small, the number is large and the distribution is stable. It is an important source of information for the physical fingerprint of background noise, so it is retained.

[0099] The area of ​​the connected components A is greater than the second preset threshold A. max Noise is identified as abnormal spots caused by dust particles or image defects. Dust particles originate from tiny deposits on the surface of the imaging chip or impurities introduced during the gel staining process. Image defects arise from dead pixels in the image sensor, pixel anomalies caused by data transmission packet loss, or compression artifacts resulting from multiple format conversions. These abnormal spots far exceed the typical size range of hardware noise in terms of area, exhibit irregular clump-like shapes, and their positions are not fixed in different batches of images taken with the same device, resulting in poor reproducibility. They lack the stability and uniqueness required for physical fingerprints and are therefore filtered out. The connected region area A is set to within a first preset threshold A. min With the second preset threshold A max The noise points are identified as valid noise generated by the hardware characteristics of the imaging device. This type of noise mainly consists of sensor fixed-mode noise and power supply ripple noise, with moderate area and stable spatial distribution, and is the core data for constructing the background noise distribution topology map. The valid image noise points retained after filtering constitute the spatial distribution point set of background noise, where each valid image noise point records its spatial coordinates (x, y) and area A.

[0100] Preferably, the step of constructing a background noise distribution topology map in the form of Delaunay triangulation using the spatial coordinates of image noise as vertices includes the following specific steps:

[0101] In one specific embodiment, a Delaunay triangulation-style noise distribution topology is constructed using the spatial coordinates of valid image noise points as vertices. Before constructing the triangulation, the spatial coordinates of the valid image noise points are normalized to map them to a uniform scale space, thereby eliminating the influence of different image resolutions on the noise topology analysis. The normalization scheme is as follows:

[0102] ;

[0103] ;

[0104] Where x and y are the horizontal and vertical coordinates of the effective noise in the original image, in pixels; W is the width of the original image, in pixels; H is the height of the original image, in pixels; and x' and y' are the normalized spatial coordinates of the noise, dimensionless.

[0105] The triangulation network is constructed using the Bowyer-Watson point-by-point insertion algorithm. The specific steps are as follows: First, a sufficiently large super triangle is constructed to contain all valid noise coordinates. Since the noise coordinates have been transformed to a [0,1]×[0,1] normalized space, a fixed geometric region covering the normalized coordinate range is used when constructing the super triangle. Its three vertices are set to (-1,-1), (2,-1), and (0.5,2) respectively, ensuring coverage of all normalized noise coordinates. Then, valid noise points are inserted into the triangulation network one by one. Each time a point P is inserted, all triangles in the current triangulation network are traversed, and triangles whose circumcircles contain point P are identified and marked as bad triangles. After removing all bad triangles, a polygonal hole is left in the triangulation network. Point P is then connected sequentially to each vertex on the boundary of the hole to form a new triangle, which is then added to the triangulation network.

[0106] After all image noise is inserted, all triangles connected to the three vertices of the super triangle are deleted, and the remaining triangulation is the Delaunay triangulation. The empty circle condition is the core constraint of Delaunay triangulation, meaning that the circumcircle of any triangle in the triangulation does not contain any other noise points. This condition ensures the uniqueness and structural stability of the triangulation under a given vertex set. The formulas for calculating the coordinates of the circumcircle center a, b and the radius R are as follows:

[0107] ;

[0108] ;

[0109] ;

[0110] Where x1, y1, x2, y2, x3, and y3 are the coordinates of the three vertices of the triangle, which are dimensionless; the formula for calculating the center of the circumcircle in this embodiment is an algebraic form based on the expanded coordinates of the triangle vertices, which is normalized by the directed area parameter and is essentially equivalent to the traditional formula for the center of the circumcircle, only expressed in a matrix form for computer program implementation.

[0111] S directed Geometrically, it is four times the directed area formed by the three vertices of a triangle arranged in order. It is dimensionless, and the formula for calculation is:

[0112] ;

[0113] Sdirected The larger the absolute value of S, the larger the area of ​​the triangle. directed The symbol S indicates the orientation of the three vertices. directed A value greater than zero indicates that the vertices are arranged counterclockwise, S directed A value less than zero indicates that the vertices are arranged clockwise. When S directed When the radius is zero, it indicates that the three vertices are collinear and cannot form a valid triangle; in this case, the radius of the circumcircle approaches infinity. In the formula for calculating the coordinates of the circumcircle's center, S... directed As the denominator, the triangle's area must be non-zero, meaning the three points are not collinear. This is naturally guaranteed by the empty circle condition in actual Delaunay triangulation. directed Its function is to transform the algebraic combination of the coordinates of the vertices of a triangle into a scalar measure of the area of ​​the triangle, providing a normalized benchmark for the calculation of the coordinates of the center.

[0114] The above formulas are used to analytically calculate the center and radius of the circumcircle of a triangle using the coordinates of its three vertices, to check if the empty circle condition is met. The Bowyer-Watson algorithm efficiently constructs a Delaunay triangulation that satisfies the empty circle condition using an incremental approach. Compared to the brute-force method of exhaustively checking all circumcircles, it reduces the computational complexity from O(n log n) to O(n log n). 4 The value is reduced to O(nlogn).

[0115] After the Delaunay triangulation is constructed, the following four types of structural features are extracted from the triangulation: adjacency matrix M, which records the connection relationships between vertices, M... i,j =1 indicates that vertex i and vertex j are connected by an edge, M i,j =0 indicates no connected edges; the adjacency matrix is ​​a symmetric matrix with dimensionless Boolean values; the edge length distribution vector E calculates the Euclidean distance over all edges.

[0116] ;

[0117] Where x i y i x j y j Let the coordinates of adjacent vertices be dimensionless. Divide the range of side lengths into 20 equidistant intervals and count the number of sides in each interval, forming a 20-dimensional side length distribution vector; the area distribution vector is S. Calculate the area of ​​each triangle using Heron's formula.

[0118] ;

[0119] Where a, b, and c are the lengths of the three sides of the triangle, dimensionless; p is the semi-perimeter, dimensionless; S is the area of ​​the triangle, dimensionless. The area range is divided into 20 equidistant intervals, and the number of triangles in each interval is counted to form a 20-dimensional area distribution vector; the angle distribution vector θ is used to calculate the minimum angle for each triangle using the law of cosines.

[0120] ;

[0121] Where a is the longest side of the triangle, b and c are the other two sides, and are dimensionless. θ min The minimum angle, measured in radians, has an output range of 0 to π for the arccos function. Dividing the angle range from 0 to π / 3 into 20 equidistant intervals, the number of triangles within each interval is counted, forming a 20-dimensional angle distribution vector. These four structural features collectively describe the topological structure information of the spatial distribution of image noise: the adjacency matrix describes the adjacency relationship between image noise points; the side length distribution describes the distance pattern between image noise points; the area distribution describes the density variation of image noise points; and the angle distribution describes the uniformity of the triangular mesh shape.

[0122] Preferably, the step of performing graph structure encoding on the extracted topological graph structure features to generate a fixed-length image physical fingerprint vector includes the following specific steps:

[0123] In one specific embodiment, the graph structure encoding employs a three-layer graph convolutional network (PCNN). This PCNN is trained unsupervised, using a contrastive loss function, which maximizes the similarity of node features from the same image and minimizes the similarity of node features from different images. The network weights are initialized using a Xavier uniform distribution, with a learning rate of 1×10⁻³, an Adam optimizer, a batch size of 32, and 50 training epochs. The initial weight matrix W... 0 The dimensions of W¹ and W² are 60×128, 128×256, and 256×512, respectively;

[0124] The training dataset for the graph convolutional network consists of noise topology graph samples generated from the background regions of scientific research images, with each scientific research image corresponding to one graph structure sample. The training dataset includes scientific research image samples from different imaging devices, different acquisition environments, and different scientific research fields. The imaging devices include different models of microscopic imaging devices, scanning devices, and digital acquisition devices.

[0125] When constructing the training dataset, scientific images acquired by different imaging devices are first collected, and noise is extracted from the background regions of each image. Based on the positional relationships of the noise points, corresponding topological connections are established, generating graph structure samples containing node features and edge connections. Each graph structure sample includes a set of noise nodes, inter-node connections, and node feature information, where node features are composed of the spatial distribution features of the noise neighborhood. To enable the graph convolutional network to learn the inherent noise features of the imaging devices, positive and negative sample data are generated during training by constructing graph structure sample pairs. Specifically, multiple scientific images acquired by the same imaging device are first acquired, and the noise topology graphs corresponding to any two images are combined into a positive sample pair; simultaneously, scientific images acquired by different imaging devices are acquired, and the noise topology graphs corresponding to different devices are combined into a negative sample pair.

[0126] For positive sample pairs, to improve the model's ability to learn the inherent noise characteristics of the equipment, the positive sample images are not limited to the same research content, but include images obtained by the same imaging equipment at different times, with different acquisition objects, or under different acquisition environments, so that the model can learn the stable noise topological features caused by the differences in the sensor response of the imaging equipment. For negative sample pairs, images from different models, batches, or sensor equipment are selected and combined to make the background noise topological structure differ in terms of equipment origin.

[0127] By constructing positive and negative samples as described above, the graph convolutional network maximizes the feature similarity between noise topology maps of the same imaging device during training, while expanding the feature differences between noise topology maps of different imaging devices, thereby learning background noise physical fingerprint features that can characterize the uniqueness of the imaging device.

[0128] The training dataset includes a training set, a validation set, and a test set. The training set is used to update the parameters of the graph convolutional network model, the validation set is used to adjust the hyperparameters during network training, and the test set is used to evaluate the model's ability to recognize the topological features of noise from different imaging devices. In this embodiment, the number of training samples is no less than 10,000, with positive and negative sample pairs constructed according to a preset ratio to ensure a balance in the number of samples of different categories during model training.

[0129] Using the adjacency matrix M as the topological connectivity of the graph, the edge length distribution, area distribution, and angle distribution features of each vertex are concatenated to form the initial node feature vector H. 0 H 0 The dimension is 60, which is composed of 20-dimensional side length features, 20-dimensional area features and 20-dimensional angle features.

[0130] The node feature update formula for the l-th layer graph convolution is:

[0131] ;

[0132] in, Let M be the adjacency matrix after adding self-loops. M is the original adjacency matrix, where M[i][j] = 1 indicates that vertex i and vertex j are connected by an edge, and M[i][j] = 0 indicates that they are not connected by an edge. The matrix is ​​a dimensionless Boolean value. I is the identity matrix, a square matrix with diagonal elements of 1 and all other elements of 0. Adding self-loops allows nodes to retain their own characteristics while aggregating neighborhood information, preventing their information from being diluted during propagation. D is... Degree matrix:

[0133] ;

[0134] Where D represents the degree of the i-th node, i.e., the number of edges connected to that node, and is a dimensionless integer. All off-diagonal elements of D are 0. The inverse square root matrix of D is obtained by raising each diagonal element of D to the power of -1 / 2. The formula is as follows:

[0135] ;

[0136] Its function is to perform symmetric normalization on the adjacency matrix, so that the aggregation operation has a normalization effect and avoids nodes with high degree from occupying too much weight in feature aggregation.

[0137] H l Let be the feature matrix of the l-th layer nodes, with dimensions N×d. l The superscript l is the layer index, where l takes the values ​​0, 1, or 2. W l Let be the learnable weight matrix of the l-th layer, with dimension d. l ×d l+1 The weights are determined through backpropagation training and are dimensionless parameters. ReLU is the modified linear unit activation function, and its functional form is:

[0138] ;

[0139] Its purpose is to introduce non-linear expressive capabilities into the network while alleviating the gradient vanishing problem.

[0140] The dimensions of the three-layer graph convolution are set as follows: the first layer has a 60-dimensional input and a 128-dimensional output; the second layer has a 128-dimensional input and a 256-dimensional output; and the third layer has a 256-dimensional input and a 512-dimensional output. As the number of convolutional layers increases, the receptive field of each node expands layer by layer. The first layer aggregates 1-hop neighbor information, the second layer aggregates 2-hop neighbor information, and the third layer aggregates 3-hop neighbor information. By the third layer, the features of each node have fused with the structural information of most nodes in the entire topology graph.

[0141] After three layers of convolutional propagation, graph-level pooling is performed on all node features. Since the background region size and effective noise number vary across different images, the number of nodes output by the graph convolutional network is not fixed. Global average pooling is used to compress any number of node features into a fixed-length vector, ensuring that the physical fingerprint vector output from images with different noise levels maintains a consistent dimension. Global average pooling aggregates all node features into a fixed-dimensional image physical fingerprint vector, as shown in the formula:

[0142] ;

[0143] Where N is the total number of effective noise points, H³[i] is the 512-dimensional feature vector of the i-th node in the third layer, and F is the final generated 512-dimensional image physical fingerprint vector with dimensionless feature values.

[0144] Global average pooling compresses the features of any number of nodes into a fixed-length graph-level representation, ensuring that the output physical fingerprint vector maintains a consistent dimension even if different images have varying numbers of effective noise points. This facilitates subsequent similarity comparison calculations. The physical fingerprint vector encodes the global topological structure information of the spatial distribution of background noise, possessing device-level uniqueness and batch-specific imaging irreproducibility. It will serve as an image identifier for subsequent cross-paper similarity comparisons.

[0145] Step S3: Using the large language model built into the AI ​​platform, extract five types of contextual information corresponding to each image from the paper sample text and encode them into a fixed-length contextual semantic vector.

[0146] In a specific embodiment, the text paragraphs associated with each image are located from the paper text extracted in step S1, including the image caption text and the method description text of the chapter containing the image. The image caption text is usually located below and to the side of the image, containing the image number and a brief description; the method description text is usually located in the Materials and Methods section and the Results section of the paper, in paragraphs adjacent to the image reference, containing a detailed description of the experimental conditions. These text paragraphs are merged into a sequence of text to be extracted, and then input into the large language model built into the AI ​​platform for information extraction.

[0147] The large language model performs two tasks on the input text: named entity recognition and relation extraction.

[0148] Named entity recognition employs a sequence labeling model based on the Transformer architecture. It outputs an entity type label for each lexical unit in the text sequence. The label set includes five categories: processing condition entities, parameter index entities, experimental environment entities, research object entities, and time parameter entities. The entity labeling format follows the BIO ternary labeling system, where B represents the starting lexical unit of the entity, I represents an internal lexical unit of the entity, and O represents a non-entity lexical unit.

[0149] This embodiment uses a BERT plus CRF sequence labeling architecture as the sequence labeling model. The bottom encoder uses a BERT-base-uncased model, containing a 12-layer Transformer encoder, 12 attention heads, and 768-dimensional hidden states, with a total of 110 million parameters. A fully connected linear layer is placed above the BERT layer to map the 768-dimensional hidden states to emission scores of the number of label categories, and then a CRF layer is connected to model the transition constraints between labels using a transition matrix. The model was fine-tuned using domain-adaptive methods on PubMed and arXiv academic text corpora, with a learning rate of 2×10⁻⁻. 5 The training consisted of three epochs, employing the AdamW optimizer and the cross-entropy loss function. The training data comprised an annotated corpus of scientific research papers, using the BIO ternary annotation format. There were five entity types, and the training set contained approximately 20,000 text paragraphs. During decoding, the CRF layer used the Viterbi algorithm to find the globally optimal label sequence. Consecutive BIO sequences of the same type of tokens formed complete entities, and the entity text was a concatenation of the original text containing the corresponding tokens.

[0150] Relation extraction employs an attention-based relation classifier to predict the semantic relation type for any two identified entities. Relation types include modification relations, numerical association relations, object attribution relations, and temporal association relations. The extraction results are output in the form of triples, with the format being entity1-relation type-entity2.

[0151] The attention-based relation classifier used in this embodiment consists of an input layer, a self-attention layer, a bilinear attention layer, and an output layer. The input layer concatenates the BERT hidden state vectors of the two entities as the initial input for the relation representation. The self-attention layer uses a multi-head self-attention mechanism to enhance the context of the input vector, with 8 attention heads, each with a dimension of 64. The bilinear attention layer interactively calculates the scores for all relation types by performing a bilinear transformation on the enhanced representations of the two entities. The enhanced representation vectors of the two entities have a dimension of 768×768. The output layer performs softmax normalization on the scores of all relation types and takes the relation type with the highest probability as the final judgment result. Relation types include four categories: modification relations, numerical association relations, object affiliation relations, and temporal association relations. The training data shares the same labeled corpus as the named entity recognition dataset, with additional annotations for relation types between entities on top of the entity annotations. The learning rate during training is set to 1×10⁻. 5 The cross-entropy loss function and AdamW optimizer were used, and the training rounds were 5.

[0152] The large language model automatically assembles the extracted entities and relations into structured quintuple information. Each quintuple contains five fields: processing condition field, which stores the original text of the extracted processing condition entity; parameter index field, which stores the original text of the extracted parameter index entity and its corresponding value; experimental environment field, which stores the original text of the extracted experimental environment entity; research object field, which stores the original text of the extracted research object entity; and time parameter field, which stores the original text of the extracted time parameter entity and its value.

[0153] The text of each field in the structured quintuple information above is encoded into semantic embedding vectors. Encoding uses the pre-trained language model BERT. The text of each field is input into the BERT input layer, processed through multiple Transformer encoders, and the hidden state vector corresponding to the CLS marker in the last layer is taken as the semantic representation of that field. The CLS marker is a special placeholder at the beginning of the BERT input sequence, and its final hidden state is used to aggregate the semantic information of the entire sequence. The BERT input representation consists of three embedding components added element-wise: a lexical embedding component, which maps the segmented input text to a sequence of lexical IDs; a positional embedding component, which encodes the position number of the lexical in the sequence; and a segmentation embedding component, used to distinguish different sentences. All three embedding components are 768-dimensional vectors, and the element-wise addition yields the input embedding matrix, which has a dimension of L×768, where L is the length of the segmented lexical sequence.

[0154] BERT's encoding process is implemented through multiple bidirectional Transformers. Each Transformer layer contains two sub-layers: a multi-head self-attention mechanism and a feedforward neural network. Each sub-layer is followed by a residual connection and a layer normalization operation. The formula for calculating multi-head self-attention is:

[0155] ;

[0156] ;

[0157] ;

[0158] in, K, V, and d represent the query matrix, key matrix, and value matrix, respectively, all obtained from the input embedding matrix through different linear transformations; k This represents the dimension of the key vector, with a value of 64. is a scaling factor with dimensions of a dimensionless constant. Its function is to scale the dot product result to a reasonable range and prevent the gradient of the softmax function from being too small; h is the number of attention heads, with a value of 12. W i K W i Vand W O Both are learnable weight matrices with dimensions of 768×64 and 768×768 respectively, and are dimensionless parameters; concat represents matrix concatenation operation.

[0159] After 12 layers of Transformer encoding, the last hidden state vector corresponding to the CLS tag is taken as the semantic embedding vector of the field. The text of each field in the quintuple is encoded separately by BERT, resulting in five 768-dimensional semantic embedding vectors. These vectors are then concatenated in the order of processing conditions, parameter indicators, experimental environment, research object, and time parameters, resulting in a total dimension of 5 × 768 = 3840 dimensions. To maintain consistency with the dimension of the physical fingerprint vector, a fully connected layer linearly maps the 3840-dimensional vector to a 512-dimensional fixed-length context semantic vector. The mapping formula is as follows:

[0160] ;

[0161] Where v1 to v5 are 768-dimensional semantic embedding vectors encoded by BERT, representing the five fields of processing conditions, parameter indicators, experimental environment, research objects, and time parameters, respectively, with dimensionless feature values; [v1||v2||v3||v4||v5] represents a 3840-dimensional vector obtained by concatenating the five vectors in sequence; W proj The learnable projective weight matrix has a dimension of 512×3840 and is a dimensionless parameter; b proj V is a learnable bias vector with 512 dimensions and dimensionless parameters; ctx The final output is a 512-dimensional context semantic vector with dimensionless feature values. The projection mapping of the fully connected layer compresses the high-dimensional concatenated vector into a compact representation with the same dimension as the physical fingerprint vector. Simultaneously, it automatically learns the weight distribution among the five fields through learnable parameters, enabling the model to adaptively highlight the semantic dimensions more critical for repetition determination while suppressing interference from redundant or irrelevant information. This context semantic vector encodes the experimental condition semantic information associated with the image and will serve as another identity dimension for subsequent cross-modal alignment and contradiction calculation.

[0162] Step S4: Summarize the physical fingerprint vectors and contextual semantic vectors of all images in the sample set to establish a scientific research image feature reference database.

[0163] In one specific embodiment, the physical fingerprint vector and contextual semantic vector obtained after processing each image of each paper in the sample set in steps S2 and S3 are summarized to establish a scientific research image feature reference database. The reference database is stored using a relational database management system, and each record in the database uniquely corresponds to one scientific research image.

[0164] Each record in the database contains the following fields: Image ID field, storing the unique identifier of the image, formatted as: Paper DOI + Image Sequence Number, data type variable-length string with a maximum length of 256 characters. This ID is generated based on the Paper's Digital Object Identifier (DOI), concatenating an underscore and the image's sequence number within the paper to ensure unique IDs for different images from different papers; Source Paper Information field, storing the paper's title, author list, journal name, publication year, and DOI number, data type variable-length string with a maximum length of 2048 characters; Physical Fingerprint Vector field, storing the 512-dimensional image physical fingerprint vector generated in step S2, data type 512-dimensional single-precision floating-point array, where each element is a dimensionless feature value, and its value range is constrained to be non-negative by the ReLU activation function output. Real numbers; Processing condition original text field, storing the original text of the processing condition entity extracted in step S3, data type is variable-length string; Parameter index original text field, storing the original text and value of the parameter index entity extracted in step S3, data type is variable-length string; Experimental environment original text field, storing the original text of the experimental environment entity extracted in step S3, data type is variable-length string; Research object original text field, storing the original text of the research object entity extracted in step S3, data type is variable-length string; Time parameter original text field, storing the original text and value of the time parameter entity extracted in step S3, data type is variable-length string; Context semantic vector field, storing the 512-dimensional context semantic vector generated in step S3, data type is 512-dimensional single-precision floating-point array, the dimension of each element is a dimensionless feature value.

[0165] To improve retrieval efficiency, vector indexes are built for the physical fingerprint vector field and the context semantic vector field. The indexes employ a hierarchical navigable small-world graph algorithm within the approximate nearest neighbor search algorithm. The construction process involves using the physical fingerprint vectors of all images as nodes and calculating the cosine similarity between any two vectors using the following formula:

[0166] ;

[0167] Among them, F a and F b Let be the 512-dimensional physical fingerprint vectors of images a and b, respectively. · represents the vector inner product, and ||·|| represents the L2 norm of the vector. The dimensions are dimensionless eigenvalues. The cosine similarity ranges from [-1, 1]. Since the elements of the physical fingerprint vector are non-negative real numbers, the actual range is [0, 1]. The closer the value is to 1, the more consistent the directions of the two vectors, meaning the more similar the background noise topology of the two images.

[0168] Based on the cosine similarity, the M most similar nodes are selected as neighbors for each node, with M being 16. The value of M depends on the average connectivity of each node in the graph structure; too small a M results in excessively long search paths, reducing retrieval efficiency, while too large a M leads to excessive index space consumption and increased construction time. All nodes are connected according to their neighbor relationships to form a hierarchical navigable graph structure. The upper layer uses long-distance connections for fast jumps, while the lower layer uses short-distance connections for fine-grained searches. After the index is built, for a given query vector, starting from the top-level entry point, a greedy search is performed layer by layer downwards for the neighbor nodes with the highest cosine similarity until the nearest neighbor set at the bottom level is reached. The retrieval complexity is O(log n), where N is the total number of images in the database.

[0169] Meanwhile, the storage of original image files and background region mask files employs a combination of file system and database indexing. Original image files and background region mask files are stored in designated directories on the local disk, with filenames named after their image IDs. The directory structure is organized hierarchically according to the paper's DOI. Database records include separate fields for the storage paths of the original image files and the background region mask files, storing the complete paths of the corresponding files as variable-length strings. This storage method separates the storage of large binary files from the structured index, ensuring both lightweight and efficient database queries and facilitating rapid retrieval of image files by path during report generation.

[0170] Once the database is established, its overall size depends on the number of papers in the sample set and the number of images per paper. Taking a sample set containing 10,000 papers with an average of eight images per paper as an example, the reference database stores approximately 80,000 records. The total storage space for physical fingerprint vectors and contextual semantic vectors is approximately 80,000 records × 1024 dimensions × 4 bytes per dimension, equivalent to about 320 megabytes. Adding the storage overhead for text fields and index structures, the total database size is approximately one gigabyte, which can be fully accommodated and efficiently run on the local storage devices of a conventional server or high-performance workstation. This reference database provides a structured benchmark data source for the paper comparison in the subsequent step S5.

[0171] Step S5: Input the paper to be tested, extract the physical fingerprint vector and context semantic vector of each image in the paper to be tested in the manner described, and calculate the background region quality score; compare the physical fingerprint vector and context semantic vector of the image of the paper to be tested with the data in the reference database one by one, and calculate the background noise similarity and context information difference.

[0172] Preferably, the input is a paper to be tested, and the physical fingerprint vector and contextual semantic vector of each image in the paper to be tested are extracted according to the method described above, and the background region quality score is calculated, including the following specific steps:

[0173] In one specific embodiment, the user uploads the PDF of the paper to be tested to the AI ​​platform. The AI ​​platform calls its built-in document parsing engine to perform structural analysis on the PDF of the paper to be tested, extracting all embedded scientific images, as well as the caption text for each image and the method description text of the chapter in which the image is located. Following the methods described in steps S1 to S3, image preprocessing, physical fingerprint extraction, and context information extraction are sequentially performed on each image in the paper to be tested, resulting in a physical fingerprint vector, a context semantic vector, and a background region quality score for each image.

[0174] Preferably, the physical fingerprint vector and contextual semantic vector of the image of the paper to be tested are compared one by one with the data in the reference database to calculate the background noise similarity and contextual information difference, including the following specific steps:

[0175] In one specific embodiment, the physical fingerprint vector of each image in the paper under test is compared one by one with the physical fingerprint vectors of all images in the reference database to calculate the background noise similarity. The similarity is measured using cosine similarity, and the calculation formula is as follows:

[0176] ;

[0177] Among them, F A Let F be the 512-dimensional physical fingerprint vector of the image A of the paper to be tested. B To reference the 512-dimensional physical fingerprint vector of image B in the database, let · denote the vector inner product operation, and ||F|| denote the L2 norm of vector F, calculated using the following formula:

[0178] ;

[0179] Where F[i] represents the i-th dimension component of vector F, with dimensionless eigenvalues. Cosine similarity Sim fp The value range of is [-1, 1]. Since the output of each component of the physical fingerprint vector after the ReLU activation function is a non-negative real number, the actual value range is [0, 1]. fp The closer the value is to 1, the more consistent the directions of the two vectors, meaning the more similar the background noise topology of the two images, and the greater the likelihood that they physically came from the same shooting session; Sim fp The closer the value is to 0, the more significant the difference in background noise distribution between the two images, and the higher the probability that they were taken from different locations. Cosine similarity is determined because, compared to Euclidean distance, which is sensitive to the absolute value of the vector and easily affected by overall image brightness shifts, cosine similarity only measures the consistency of vector direction and has a natural robustness to changes in the overall intensity of noise signals.

[0180] Simultaneously, the contextual semantic vector of each image in the paper under test is compared one by one with the contextual semantic vector of the corresponding image in the reference database to calculate the contextual information difference. The difference is measured using a five-dimensional weighted semantic distance metric, and the calculation formula is as follows:

[0181] ;

[0182] Where i is the context information dimension index, and i ranges from 1 to 5, corresponding to the five dimensions of processing conditions, parameter indicators, experimental environment, research objects, and time parameters, respectively; w i The preset weights for the i-th dimension are dimensionless, with default values ​​of w1=0.25, w2=0.25, w3=0.20, w4=0.20, and w5=0.10. The weight allocation is based on the fact that the treatment conditions and parameter indicators usually have the most significant impact on the experimental results, followed by the experimental environment and research subjects, while the time parameter has a relatively smaller impact; d i Let be the semantic distance in the i-th dimension, with a dimensionless distance value ranging from 0 to 1.

[0183] Semantic distance d in each dimension i Calculations are performed based on data type. For the processing condition dimension, if the processing conditions of two images belong to different categories, then d1 is set to 1; if they belong to the same category but have different parameters, then d1 is calculated based on the proportion of parameter differences, i.e., d1 is the proportion of the difference in the text description to the total description length. For the parameter index dimension, if the parameter is numerical, then d2 is the normalized absolute difference, calculated using the following formula:

[0184] ;

[0185] Where p A and p B These are the parameter values ​​for the two images, with the same units as the corresponding parameters, max(p) A ,p B The function of ) is to normalize the difference to a value between 0 and 1, with ε taking a value of 10. -6 , guarantee p A and p B When both are 0, the denominator is not 0; in this case, the numerator |p A -p B The value of ε is also zero, and the calculated result of d2 is 0, indicating that the parameter values ​​of the two images are exactly the same and there is no difference, which is consistent with the actual semantics. The basis for determining the value of ε is that this value is much smaller than the measurement precision of any meaningful parameter in scientific research experiments, and its impact on the normalization calculation results under normal conditions is negligible, while effectively avoiding division by zero anomalies; if the parameter is textual description type, then d2 is taken as the cosine distance of the corresponding semantic embedding vector, that is

[0186] ;

[0187] Where v A and v B These refer to the semantic embedding vectors of images A and B in terms of parameter metrics, respectively. For the experimental environment dimension, d3 is calculated based on the shortest path distance between concepts in a pre-defined subject knowledge base: d3 = pathLen / maxDepth, where pathLen is the number of edges in the shortest path between the two concepts in the knowledge base hierarchy (dimensionless integer), and maxDepth is the maximum depth of the knowledge base hierarchy (dimensionless integer). This normalization method keeps the value of d3 between 0 and 1. For the research object dimension, d4 is calculated by subtracting the Jaccard similarity coefficient of the names of the two research objects from 1, using the following formula:

[0188] ;

[0189] Where S A and S B These are the word sets after word segmentation of the names of two research objects, |S A ∩S B |S represents the size of the intersection of the two sets. A ∪S B | represents the size of the union of the two sets, both of which are dimensionless integers. For the time parameter dimension, d5 is the normalized time difference, calculated using the following formula:

[0190] ;

[0191] Where t A and t B These are the time values ​​for the two images, with dimensions consistent with the corresponding time units.

[0192] Contextual information difference (Diff) ctx The value ranges from 0 to 1. A value closer to 1 indicates a greater difference in the experimental conditions claimed by the two images, while a value closer to 0 indicates greater similarity in the experimental conditions. This formula quantifies the degree of difference between the two images in terms of experimental conditions into a unified scalar index through weighted fusion of multi-dimensional semantic distance, providing semantic difference input for the contradiction calculation in step S6. The background noise similarity Sim between each test image and the reference image is... fp Difference between contextual information and Diff ctx The records are paired to form a comparison result set, which is used for determining the degree of contradiction in step S6.

[0193] Step S6: Calculate the degree of contradiction by combining the similarity of background noise in the image of the paper to be tested and the difference in context; adjust the judgment criteria according to the background region quality score, and generate a visual detection report.

[0194] In one specific embodiment, background noise similarity is read pair by pair from the comparison result set. Difference between contextual information and The degree of contradiction is scored based on the two factors, and the calculation formula is as follows:

[0195] ;

[0196] in, The background noise similarity between image A of the paper to be tested and image B of the reference database is dimensionless and ranges from 0 to 1. Contextual information difference measure is used to assess the degree of difference between two images in terms of structural content. A preset threshold is set for the similarity of background noise, which is dimensionless.

[0197] The The value is based on the following: Images acquired by the same imaging device under the same shooting conditions typically maintain a high level of background noise physical fingerprint similarity, while the background noise similarity between images from different imaging devices or different sources is typically low. Therefore, by experimentally analyzing the background noise similarity distribution between images acquired by the same device and images acquired by different devices, the critical value within the distinguishing range of the two is taken as the value. The basis for this setting. In this embodiment, we take... =0.85, a value chosen based on the fact that in actual experimental statistics, the physical fingerprint cosine similarity of images from the same capture is usually no less than 0.85, while the similarity of images from different captures is usually no higher than 0.70. 0.85 falls within the safe dividing line between these two values, effectively distinguishing between physically similar and physically dissimilar image pairs. When Sim... fp Below When the two images are in a state of flux, it indicates that they are very likely from different shooting sessions at the physical level, and there is no physical basis for reuse. In this case, CCF-ME is directly set to 0, indicating no contradiction. When Sim... fp Reaching or exceeding When the two images are highly consistent at the physical level, Sim indicates that they are physically identical. fp and Multiplying them yields the CCF-ME. The CCF-ME value ranges from 0 to 1. The closer the value is to 1, the more similar the physical fingerprints are and the greater the difference in context. That is, the two images are physically highly homologous but have obvious logical contradictions in semantics, and the higher the risk of abnormal duplication.

[0198] The judgment criteria are dynamically adjusted based on the background area quality score. The adjustment formula for the judgment criteria is:

[0199] ;

[0200] Among them, Tbase The basic threshold is set at 0.60, and the unit is a dimensionless ratio. T base The decision to set the threshold to 0.60 is based on the fact that when the physical fingerprint similarity is 0.85 and the contextual difference is 0.70, the CCF-ME score is 0.595, which is close to 0.60. This combination represents a typical abnormal repetition scenario, hence the base threshold is set to 0.60. ref As the quality reference score, the smaller value between the quality scores of the background region of the test image and the reference image is taken. The calculation formula is as follows:

[0201] ;

[0202] Q A and Q B These are the background region quality scores for the test image and the reference image, respectively. The scores are dimensionless and range from 0 to 1. α(Q) ref ) is the threshold adjustment coefficient function, and its value selection rules are as follows:

[0203] ;

[0204] Among them, Q ref The quality threshold of 0.5 is determined based on the fact that when the background area quality score is below 0.5, there are usually problems such as insufficient background area, uneven grayscale, or severe compression artifacts, which reduces the reliability of noise information. In this case, the threshold is increased by 20%, i.e., multiplied by 1.2, to reduce the risk of misjudgment due to insufficient noise information; Q ref When the threshold is not less than 0.5, the background quality is good and the noise information is sufficient and reliable. At this point, lowering the threshold by 10%, i.e., multiplying it by 0.9, can improve the detection sensitivity for weak similar signals. The selection logic for the threshold adjustment coefficients of 1.2 and 0.9 is based on the fact that too large an adjustment range will lead to a sharp deterioration in the missed detection or false detection rate, while too small an adjustment range will lose its adaptive significance. An adjustment range of 15% to 25% is an empirically reasonable range.

[0205] Then, the contradiction score CCF-ME is compared with the adjusted judgment criterion T. final Compare them. The comparison rules are as follows:

[0206] ;

[0207] When CCF-ME exceeds T final If the image pair is deemed to have abnormal duplication, it is marked as a suspected duplicate image pair. Simultaneously, a comprehensive confidence score is calculated using the following formula:

[0208] ;

[0209] In this system, β1, β2, and β3 are weighting coefficients, with default values ​​of β1=0.3, β2=0.3, and β3=0.4, satisfying β1+β2+β3=1. These are dimensionless weights. The highest weight of 0.4 for β3 is determined because CCF-ME itself integrates the product of similarity and difference information, making it the most direct anomaly detection indicator, and therefore should be given the highest weight. The selection logic for β1 and β2, each at 0.3, is based on the fact that similarity and difference are components of CCF-ME, and their independent information still has auxiliary reference value for the confidence score. conf The value ranges from 0 to 1, and the higher the score, the higher the reliability of the abnormal repetition judgment.

[0210] A visual detection report is generated for all image pairs marked as suspected duplicates. The original image files and background region masks of two images are retrieved from the reference database. The two original images are scaled to the same size and then stitched side-by-side to generate a thumbnail comparison image. Background noise distribution topology data of the two images are retrieved from the reference database. Noise vertices and Delaunay triangulation edges of the test image are drawn in red, while noise vertices and triangulation edges of the reference image are drawn in blue. These are superimposed on the same coordinate system to generate a topology overlay comparison image, visually demonstrating the consistency and local differences in noise distribution between the two images. Five types of contextual information from the reference database are read from the original text of the two images. The comparison results for processing conditions, parameter indicators, experimental environment, research objects, and time parameters are listed item by item. Dimensions with differences are highlighted, generating a comparison table of differences in contextual information item by item.

[0211] The above thumbnail side-by-side comparison, topology overlay comparison, and contextual information difference comparison table, along with the background noise similarity Sim, are combined. fp Contextual information difference (Diff) ctx The degree of contradiction (CCF-ME) and the overall confidence score. conf The data is integrated into a structured, visualized detection report and output. The report is sorted from highest to lowest score based on the degree of contradiction, allowing users to prioritize reviewing image pairs with the highest risk of anomalous duplication.

[0212] Example 2:

[0213] Please see Figure 4 As shown, a scientific image repeatability intelligent detection device based on an AI platform includes a processor, a memory, and a communication bus.

[0214] The memory stores a computer-readable program that can be executed by the processor;

[0215] The communication bus enables communication between the processor and the memory;

[0216] When the processor executes the computer-readable program, it performs steps in the AI-based intelligent detection method for scientific image repetition as described in any one of the present invention.

[0217] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0218] It should be understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0219] It should be understood that determining B based on A does not mean determining B solely based on A; it also means determining B based on A and / or other information.

[0220] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0221] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent detection of repetition in scientific research images based on an AI platform, characterized in that, Includes the following steps: Step S1: Establish a sample set of published research papers, extract all research images from each paper in the sample set, perform format unification and type recognition processing on each image, separate the main content in the image, exclude non-background elements, and obtain the image background area. Step S2: Noise detection is performed on the background region of each image to obtain the position coordinates of each noise point, and the pixel coordinates are converted into normalized spatial coordinates according to the image width and height; a noise distribution topology map in the form of a Delaunay triangulation is constructed using the normalized spatial coordinates as vertices, and the side length, area and angle distribution of each triangle in the triangulation and the adjacency relationship between each vertex are recorded. The above topology map is encoded in graph structure to generate a fixed-length image physical fingerprint vector. Step S3: Using the large language model built into the AI ​​platform, extract five types of contextual information corresponding to each image from the paper sample text and encode them into a fixed-length contextual semantic vector. Step S4: Summarize the physical fingerprint vectors and contextual semantic vectors of all images in the sample set to establish a scientific research image feature reference database; Step S5: Input the paper to be tested, extract the physical fingerprint vector and context semantic vector of each image in the paper to be tested in the manner described, and calculate the background region quality score; compare the physical fingerprint vector and context semantic vector of the image of the paper to be tested with the data in the reference database one by one, and calculate the background noise similarity and context information difference. Step S6: Calculate the degree of contradiction by combining the similarity of background noise in the image of the paper to be tested with the difference in context; The judgment criteria are adjusted based on the background area quality score, and a visual inspection report is generated.

2. The intelligent detection method for the repetition of scientific research images based on an AI platform according to claim 1, characterized in that, Step S2 includes the following steps: Gaussian filtering is applied to the background region of the image to obtain a smooth background image. The background region of the image and the smooth background image are then subjected to pixel-by-pixel difference operation to extract the residual signal as the image noise signal. After adaptive threshold binarization of the image noise signal, connected component analysis is used to locate the centroid coordinates of each connected component as the spatial coordinates of the corresponding image noise, and the area of ​​each connected component is recorded simultaneously. Based on this, noise points with a connected region area smaller than the first preset threshold are judged as sensor thermal noise fluctuations and retained, while noise points with a connected region area greater than the second preset threshold are judged as abnormal spots caused by dust particles and image defects and filtered out. Finally, the spatial coordinate set of effective image noise points generated by the hardware characteristics of the imaging device and the corresponding connected region area are output. A noise distribution topology graph in the form of a Delaunay triangulation is constructed using the effective image noise spatial coordinate set as vertices. The side length, area, and angle distribution of each triangle in the triangulation, as well as the adjacency relationship between each vertex, are recorded. The above topology graph is then encoded using graph structure to generate a fixed-length image physical fingerprint vector.

3. The intelligent detection method for the repetition of scientific research images based on an AI platform according to claim 2, characterized in that, The construction of a noise distribution topology map in the form of a Delaunay triangulation using the effective image noise spatial coordinate set as vertices includes the following steps: Using the set of spatial coordinates of effective noise points as input, and taking the two-dimensional spatial coordinates of each noise point as a vertex of a two-dimensional plane, an initial vertex set is constructed. Based on the initial vertex set, the Bowyer-Watson point-by-point insertion algorithm is used to generate the Delaunay triangulation network, ensuring that the circumcircle of any triangle in the triangulation network does not contain other vertices, maximizing the minimum angle of each triangle, thereby obtaining the image space partitioning structure. Structural features are extracted from the Delaunay triangulation. An adjacency matrix is ​​established using the image noise number as the row and column index. The side length distribution vector of each triangle, the area distribution vector of each triangle, and the angle distribution vector of the minimum angle of each triangle are recorded. The adjacency matrix, side length distribution vector, area distribution vector, and angle distribution vector are used as structural features of the background noise distribution topology. After multi-layer propagation and pooling operations of the graph convolutional network, a fixed-dimensional image physical fingerprint vector is generated.

4. The intelligent detection method for the repetition of scientific research images based on an AI platform according to claim 2, characterized in that, The process of extracting structural features and generating an image physical fingerprint vector via graph structure encoding includes the following steps: The adjacency matrix is ​​used as the topological connection relationship of the graph, and the side length distribution vector, area distribution vector and angle distribution vector are concatenated as the initial feature vector of each node to construct graph structure data with node attributes. The graph structure data is input into a graph convolutional network, and neighborhood information is aggregated through multiple layers of graph convolution operations. Each layer of graph convolution operation performs a weighted summation of the features of each node and the features of its neighboring nodes, and then transforms them through a nonlinear activation function to expand the receptive field of the nodes layer by layer. Graph pooling is performed on the node features after multi-layer graph convolution operations to aggregate the features of all nodes into a fixed-dimensional graph-level representation vector, which is a fixed-length image physical fingerprint vector.

5. The intelligent detection method for the repetition of scientific research images based on an AI platform according to claim 1, characterized in that, Step S3 includes the following steps: Extract the text paragraphs associated with each image from the main text of each paper sample, including the image caption text, the method description text of the paragraph containing the image, and the result analysis text containing the description of experimental conditions; The extracted text paragraphs are input into the large language model built into the AI ​​platform to perform named entity recognition and relation extraction. The model identifies processing condition entities, parameter index entities, experimental environment entities, research object entities, and time parameter entities in the text, and extracts the modification and association relationships between each entity, outputting structured quintuple information. The text entities of each dimension in the structured quintuple information are input into a pre-trained language encoder to generate semantic embedding vectors of each dimension, and the semantic embedding vectors of each dimension are concatenated into a fixed-length context semantic vector.

6. The intelligent detection method for the repetition of scientific research images based on an AI platform according to claim 1, characterized in that, Step S5 includes the following steps: The paper document to be tested is analyzed, all images in the paper are extracted, the format of each scientific research image is standardized and the type is identified, the main content and background area are separated, non-background elements are excluded, and the background area of ​​the image is obtained. The background region quality score is calculated based on factors including the area ratio of the background region to the entire image, the standard deviation of the grayscale of the background region pixels, the number of effective noise points detected in the background region, and whether there are any lossy compression artifacts in the image. Noise detection is performed on the background region to obtain the position coordinates of each noise point, and the pixel coordinates are converted into normalized spatial coordinates according to the image width and height. The noise distribution topology map in the form of Delaunay triangulation is constructed using the normalized spatial coordinates as vertices, and a fixed-length image physical fingerprint vector is generated through graph structure encoding. We use a large language model to extract five types of contextual information from the associated text: processing conditions, parameter indicators, experimental environment, research object, and time parameters for each image, and encode them into a fixed-length contextual semantic vector. The physical fingerprint vector and contextual semantic vector of the image of the paper to be tested are compared one by one with the data in the reference database to calculate the background noise similarity and contextual information difference.

7. The intelligent detection method for the repetition of scientific research images based on an AI platform according to claim 6, characterized in that, Calculating the background noise similarity and contextual information difference includes the following steps: The physical fingerprint vector and context semantic vector of the stored images are read one by one from the reference database, and the corresponding source paper information, image number and five types of context information are read in the original text. The physical fingerprint vector of each image in the paper to be tested is compared with the physical fingerprint vector of each image in the reference database using cosine similarity calculation to obtain the background noise similarity. The contextual semantic vector of each image in the paper to be tested is compared with the contextual semantic vector of the corresponding image in the reference database in a dimension-by-dimensional manner: the semantic distance of the processing condition dimension, the numerical difference of the parameter index dimension, the semantic distance of the experimental environment dimension, the name similarity of the research object dimension, and the numerical difference of the time parameter dimension are calculated respectively; the difference values ​​of the above five dimensions are weighted and summed according to the preset weights and then normalized to obtain the contextual information difference degree. The background noise similarity and contextual information difference between each test image and the reference image are recorded in pairs to form a comparison result set.

8. The intelligent detection method for the repetition of scientific research images based on an AI platform according to claim 1, characterized in that, Step S6 includes the following steps: The background noise similarity and context information difference are read from the comparison result set one by one, and it is determined whether the background noise similarity exceeds the preset threshold of background noise similarity. If it does not exceed the threshold, it indicates that the two images are taken from different locations at the physical level, and are directly judged as having no abnormal repetition, with the degree of contradiction set to zero. If it exceeds the threshold, the background noise similarity and context information difference of the image pair are multiplied to obtain the degree of contradiction. Read the background region quality scores of the image to be tested and the reference image, and take the lower value of the two as the quality reference score of the image pair; when the quality reference score is lower than the preset quality threshold, the judgment standard is raised to reduce the risk of misjudgment due to insufficient background noise information; when the quality reference score is higher than the preset quality threshold, the judgment standard is lowered to improve the detection sensitivity. The degree of contradiction is compared with the adjusted judgment criteria. When the contradiction exceeds the judgment criteria, the image pair is judged to have abnormal repetition and is marked as a suspected duplicate image pair. For all image pairs marked as suspected duplicates, the original image files and background region masks of the two images are retrieved from the reference database to generate thumbnail side-by-side comparison images; the background noise distribution topology maps of the two images are overlaid and drawn, with different colors used to distinguish the noise vertices and triangular mesh edges of the two images to generate an overlaid comparison image of the topology maps; the original text of five types of contextual information of the two images is read from the reference database, and the comparison results of processing conditions, parameter indicators, experimental environment, research objects and time parameters are listed item by item, and the dimensions with differences are marked to generate a contextual difference table; The above thumbnail comparison, topology overlay, and context difference table, along with the background noise similarity score, context information difference score, contradiction degree score, and comprehensive confidence score, are integrated into a visual detection report and output.

9. A scientific research image repetition intelligent detection device based on an AI platform, characterized in that, Includes processor, memory, and communication bus; The memory stores a computer-readable program that can be executed by the processor; The communication bus enables communication between the processor and the memory; When the processor executes the computer-readable program, it performs the steps of the AI-based intelligent detection method for scientific image duplication as described in any one of claims 1 to 8.

Images