Thin layer chromatogram spectrum duplicate checking method and system based on artificial intelligence

Abstract

This invention relates to the field of image processing technology, providing an artificial intelligence-based method and system for plagiarism detection in thin-layer chromatography images. The method involves: converting, denoising, and normalizing the thin-layer chromatography image to form a standardized input image; extracting and fusing multi-level features through a multi-layer convolutional network to generate an overall feature vector and a difference heatmap; performing multi-band wavelet decomposition on the denoised grayscale image, calculating the energy difference of JPEG compression noise in each sub-band, and generating a suspected tampering heatmap; detecting the solvent front and spot distance, calculating the ratio shift value, and performing physical self-consistency verification to generate a physical inconsistency marker; weighted fusing four similarity indicators with the centroid offset and physical inconsistency markers to form a comprehensive compliance score; comparing with historical spectra in the LIMS system, and outputting normal review, re-examination, or data interception through a three-level threshold. This invention achieves collaborative plagiarism detection through visual comparison, tamper detection, and physical verification.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for deduplication of thin-layer chromatograms based on artificial intelligence. Background Technology

[0002] Thin-layer chromatography (TLC) is a rapid separation and detection method widely used in pharmaceutical analysis, traditional Chinese medicine identification, food safety, and chemical industries. The experimental results are typically recorded as images of the developed thin-layer plate, i.e., TLC chromatograms. In TLC chromatograms, the ratio shift (Rf) value characterizes the location of spots, and the color, size, and separation clarity of the spots directly reflect the chemical properties and content relationships of the components in the sample. Verification and management of TLC chromatograms refers to the process of verifying the authenticity, completeness, and compliance of the TLC image data generated or recorded in the experiment, as well as its systematic storage, retrieval, and traceability. Its core tasks include: authenticity verification; tamper detection; reuse identification; and historical chromatogram management.

[0003] With the increasing demands for laboratory informatization and data integrity, TLC maps, as an important part of original records, have become a crucial link in quality control and data anti-counterfeiting. However, existing technologies still have significant shortcomings in using artificial intelligence to intelligently verify TLC maps.

[0004] Patent CN202210240410A discloses a thin-layer chromatography component analysis method based on adaptive weight fusion. This method extracts feature vectors from the image to be analyzed using a preset filter, matches these feature vectors with template feature vectors of the target sample to obtain a first similarity score, and simultaneously inputs the image to be analyzed into a trained first neural network to obtain spectral bands. A second similarity score is then obtained based on the calculated relative positions between these bands. Finally, the two similarity scores are input into a trained second neural network for fusion analysis, outputting the component analysis results. While this approach introduces the concept of similarity fusion, its similarity sources are limited to two dimensions: feature vector matching scores and relative positions between spectral bands. Furthermore, its core purpose is component identification rather than data authenticity verification and tamper detection. In addition, this method only involves two-dimensional similarity fusion and does not involve a multi-dimensional comprehensive scoring mechanism.

[0005] The intelligent thin-layer chromatography (TLC) comparison technology, jointly developed by 3DWorld and Shenzhen Institute for Drug Control, was officially launched in March 2026. This technology uses deep learning to build an AI verification model for TLC, extracting 23 core features such as spot grayscale, Rf value, and morphology from chromatograms via convolutional neural networks. It generates a visual comparison result using a difference heatmap, enabling full lifecycle traceability management for each product and deeply integrating with the LIMS system. This technology represents the latest advancement in the application of artificial intelligence in TLC chromatogram verification, effectively identifying visual similarity and duplication issues between chromatograms.

[0006] However, the aforementioned existing technologies still share the following common drawbacks: First, their core technical paths are all based on a single "feature extraction-similarity comparison" approach. This means that by extracting the visual features of the chromatogram and comparing it with historical chromatograms, they can only determine whether there is visual similarity between the chromatograms and cannot identify digital image tampering operations such as copying, splicing, or local modification. Second, none of the above schemes utilize the inherent physical constraints of thin-layer chromatography (i.e., the ratio of the spot center to the origin to the solvent front to the origin) to verify the physical consistency of the data. This allows counterfeiters to evade detection by modifying the position of local spots or adjusting the Rf value. Third, the judgment output of the existing technologies is based on a single threshold mode, which cannot classify risks according to the severity of tampering traces and is difficult to meet the differentiated requirements of LIMS systems for compliance auditing.

[0007] Therefore, there is currently a lack of a multi-dimensional thin-layer chromatogram duplication detection scheme that can simultaneously take into account visual similarity comparison, digital tampering trace detection and physical consistency verification. This is the technical problem that this invention aims to solve. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method for duplicate detection of thin-layer chromatograms based on artificial intelligence, comprising the following steps: Step 1: The original thin-layer chromatography image is scaled proportionally and normalized by edge filling, then converted to grayscale and denoised to obtain a denoised grayscale image that highlights the outline and positional features of the spots; the denoised grayscale image is normalized by the hyperbolic tangent activation function to form a standardized input image. Step 2: The standardized input image extracts and fuses multi-level features through a multi-layer convolutional network with built-in artificial intelligence algorithms to generate the overall feature vector of the current map; at the same time, it outputs a heatmap of the difference between the current map and the historical map. Step 3: Perform multi-band wavelet decomposition on the denoised grayscale image to obtain multiple frequency band sub-bands; calculate the JPEG compressed noise energy distribution of the current image and historical images in each frequency band sub-band, and calculate the difference degree of the JPEG compressed noise energy distribution for each frequency band sub-band; when the difference degree of any frequency band sub-band exceeds a preset threshold, generate a suspected tampering heatmap including the tampering boundary position; and calculate the Euclidean distance between the grayscale centroid coordinates of the bright area in the suspected tampering heatmap and the geometric center of the heatmap, as the centroid offset of the suspected tampering heatmap; Step 4: Detect the distance from the solvent front to the origin and the distance from the center of each spot to the origin in the denoised grayscale image. Calculate the ratio of the distance from the center of each spot to the origin to the distance from the solvent front to the origin. Use this ratio as the ratio shift value of the spot and verify whether the ratio shift values ​​of all spots meet the preset physical self-consistency conditions. When it is detected that the distance from the solvent front to the origin cannot be uniquely determined, or the linear correlation coefficient between the spot center distance and the solvent front distance is lower than the preset threshold, or the ratio shift value of any spot exceeds the range of 0 to 1, it is determined that the spectrum has suspected ratio shift value forgery or local tampering, and a physical inconsistency mark is generated. Step 5: Perform multi-dimensional calculations on the overall feature vector of the current map and the overall feature vector of the historical map: calculate the error similarity index using the average mean square error, calculate the directional similarity index using the average cosine similarity, calculate the linear correlation index using the average correlation coefficient, and calculate the structural similarity index using the average structural similarity; then, weight and fuse these four indicators with the centroid offset of the suspected tampering heatmap obtained in Step 3 and the physical inconsistency markers from Step 4 according to preset weight coefficients to form a comprehensive compliance score; Step Six: Compare the comprehensive compliance score with the historical maps stored in the LIMS system in batches to obtain the compliance value for each comparison; input the compliance value into a preset three-level threshold judge: when the compliance value is higher than the second threshold, it is judged as having no risk of duplication and no trace of tampering, and a normal audit process output is formed; when the compliance value is between the first and second thresholds, it is judged as having suspected tampering, and a heat map of the tampered area and a report on the consistency of the ratio shift value are generated and output to the review queue for manual review; when the compliance value is lower than the first threshold, it is judged as highly suspected of duplication or systematic tampering, and the associated experimental data is directly intercepted and frozen, triggering the compliance audit event record.

[0009] Optionally, the process of normalizing the pixel values ​​of the denoised grayscale image using the activation function includes the following steps: Using the denoised grayscale image as the initial image, a linear stretching transformation is performed on the grayscale value of each pixel to map the original integer value range from 0 to 255 to the floating-point range from 0 to 1, thus obtaining a preliminary normalized image. Each pixel value in the pre-normalized image is input into the hyperbolic tangent activation function for non-linear mapping, so that the output value falls within the range of negative one to positive one, thus obtaining the activation-mapped image. Calculate the mean and standard deviation of all pixels in the activation mapping image. Subtract the mean from each pixel value and divide by the standard deviation to make the pixel distribution of the final image satisfy the statistical characteristics of zero mean and unit variance, thus obtaining the standardized input image.

[0010] Optionally, a multi-layer convolutional network is used to perform shallow feature extraction to obtain a low-level feature map containing the location, area, and gray intensity of the spots. The low-level feature map is then processed by a middle convolutional layer to form a mid-level morphological feature map containing the shape and edge contour of the spots. The mid-level morphological feature map is then processed by a deep convolutional layer to form a feature vector of the overall chromatographic distribution.

[0011] Optionally, the bright areas of the difference heatmap indicate significant differences, while the dark areas indicate similarities.

[0012] Optionally, the process of generating the overall feature vector of the current map includes the following steps: Shallow feature maps, mid-level feature maps, and deep feature vectors are extracted from a multi-layer convolutional network. The shallow and mid-level feature maps are each passed through a global average pooling layer to compress the spatial dimension of each feature map into a single value, thus transforming the shallow feature map into a shallow feature vector and the mid-level feature map into a mid-level feature vector. The shallow feature vector, mid-level feature vector, and deep feature vector are then placed side by side to form three independent primary feature vector groups, denoted as the feature vector group. The shallow, middle, and deep feature vectors in the feature vector group are concatenated end to end in channel order to form a combined long vector whose dimension is the sum of the dimensions of the three. This combined long vector is input into the first fully connected layer, which performs a non-linear mapping through a linear transformation and the ReLU activation function to output an intermediate fused feature vector with reduced dimension. The intermediate fused feature vector is input into the second fully connected layer. The fully connected layer performs a linear transformation but does not add an activation function to output an original feature vector with a preset fixed dimension. Then, the original feature vector is subjected to L2 norm normalization, that is, the square root of the sum of the squares of all its elements is calculated, and then each element is divided by the square root. Finally, an overall feature vector with a consistent numerical range and a unique direction is obtained. The overall feature vector is the overall feature representation of the current thin-layer chromatogram.

[0013] Optionally, the process of outputting a dimension-reduced intermediate fused feature vector includes the following steps: The shallow feature vector, middle feature vector and deep feature vector in the feature vector group are concatenated end to end in channel order to form a combined long vector with a dimension equal to the sum of the dimensions of the three features, which is denoted as the combined long vector. The combined long vector is input into the first fully connected layer. The fully connected layer performs a linear transformation on the input vector, weights and sums each output neuron with all input neurons and adds a bias to obtain the linearly transformed intermediate vector. Then, each element in the intermediate vector is processed by the ReLU activation function, keeping all elements greater than zero unchanged and setting elements less than or equal to zero to zero, and finally outputting a non-linear intermediate fusion feature vector with reduced dimension. The intermediate fused feature vector is input into the second fully connected layer, which performs a linear transformation again and outputs an original feature vector with a preset fixed dimension; the original feature vector is an input normalized to the L2 norm.

[0014] Optionally, the error similarity index and the direction similarity index are calculated as follows: the overall feature vector of the current map and the overall feature vector of the historical map are taken as two input vectors, denoted as the input pair; the input pair is simultaneously fed into the mean square error calculation module and the mean cosine similarity calculation module: in the mean square error calculation module, the square of the difference between the two vectors is calculated element by element and then the average is obtained to obtain the error similarity index; in the mean cosine similarity calculation module, the dot product of the two vectors is calculated and divided by the product of their magnitudes to obtain the direction similarity index; the error similarity index and the direction similarity index are merged into a two-dimensional intermediate vector.

[0015] Optionally, the linear correlation index and structural similarity index are calculated as follows: The input pair is simultaneously fed into the average correlation coefficient calculation module and the average structural similarity calculation module: In the average correlation coefficient calculation module, the covariance of the two vectors is calculated by dividing the product of their respective standard deviations to obtain the linear correlation index; In the average structural similarity calculation module, the two vectors are treated as one-dimensional signal sequences, and the similarity of brightness, contrast and structure is calculated separately and then averaged to obtain the structural similarity index; The linear correlation index and the structural similarity index are merged into a two-dimensional intermediate vector.

[0016] Optionally, the process of forming the weighted sum of the four indicators includes the following steps: concatenating the error similarity index calculated from the average mean square error, the directional similarity index calculated from the average cosine similarity, the linear correlation index calculated from the average correlation coefficient, and the structural similarity index calculated from the average structural similarity in a preset fixed order to generate an array containing four numerical components; reading the four weight coefficients corresponding to error similarity, directional similarity, linear correlation, and structural similarity respectively from the system configuration parameters, with the sum of the four weight coefficients being one; performing multiplication operations on each of the four components in the array, multiplying each component by its corresponding weight coefficient to obtain four weighted values; performing a summation operation on the four weighted values, and using the summation result as the weighted sum of the four indicators.

[0017] An AI-based thin-layer chromatography chromatogram plagiarism detection system, used for AI-based thin-layer chromatography chromatogram plagiarism detection methods, includes: The image preprocessing module is used to scale and normalize the original thin-layer chromatography image proportionally and fill the edges, and then perform grayscale conversion and background noise reduction to obtain a denoised grayscale image that highlights the outline and positional features of the spots. The denoised grayscale image is further processed by the hyperbolic tangent activation function to normalize the pixel values ​​and form a standardized input image. The feature extraction module is used to extract multi-level features from the standardized input image through a multi-layer convolutional network with built-in artificial intelligence algorithms and fuse them to generate the overall feature vector of the current map; at the same time, it outputs a heatmap of the difference between the current map and the historical map. The tampering detection module includes: (a) Multi-band compression trace detection submodule, used to perform multi-band wavelet decomposition on the denoised grayscale image, calculate the noise energy distribution difference in each band, locate the tampering boundary area, generate a suspected tampering heat map; and calculate the Euclidean distance between the grayscale centroid of the bright area in the heat map and the geometric center of the heat map as the centroid offset. (b) Physical self-consistency verification submodule, used to detect the distance from the solvent front to the origin and the distance from the center of each spot to the origin in the noise-reduced grayscale image, calculate the ratio of the distance from the center of each spot to the origin to the distance from the solvent front to obtain the spot ratio shift value, and verify whether the ratio shift values ​​of all spots meet the physical self-consistency condition, and generate physical inconsistency markers. The similarity metric module is used to perform multi-dimensional calculations on the overall feature vector of the current map and the overall feature vector of the historical map, and obtain four similarity indices through average mean square error, average cosine similarity, average correlation coefficient and average structural similarity respectively. The comprehensive scoring module is used to weight and fuse the four similarity indicators with the centroid offset and physical inconsistency markers of the suspected tampering heatmap to form a comprehensive compliance score. The LIMS integration and three-level judgment module is used to compare the comprehensive compliance score with the historical maps stored in the LIMS system in batches, and call the three-level threshold judge: when the compliance value is higher than the second threshold, the normal review process is output; when the compliance value is between the first and second thresholds, a review reminder is output and a report on the consistency of the tampered area heat map and the ratio shift value is generated; when the compliance value is lower than the first threshold, the related experimental data is directly intercepted and frozen, triggering the compliance audit event record. The display module is used to display the tampered area heatmap, the ratio shift consistency report, and the difference heatmap in the form of overlay layers.

[0018] This invention solves the problems of low efficiency and high subjectivity in manual verification. In existing technologies, manual comparison is not only time-consuming and inefficient, but the judgment results also heavily depend on the experience and condition of the inspectors, lacking unified standards and prone to missed or incorrect judgments. This invention uses artificial intelligence algorithms to achieve automatic batch comparison of spectra, with unified judgment standards, improving efficiency by over 90%, and eliminating biases caused by human factors.

[0019] This invention solves the problem that simple hash comparison cannot identify covert tampering. In existing technologies, comparisons based on file hash values ​​can only identify completely identical chromatogram files, failing to handle normal experimental differences such as shooting angle, lighting, and minor variations in spots, and also unable to detect subtly altered chromatograms. This invention employs a multi-level feature extraction model based on convolutional neural networks to encode and compare key features of thin-layer chromatography, such as spot position, Rf value, color intensity, and distribution. This enables accurate identification of duplicate or tampered chromatograms while tolerating reasonable experimental errors.

[0020] This invention further solves the problem that existing deep learning-based plagiarism detection schemes cannot detect traces of digital image tampering. Existing technologies (such as CN202410296628) can only determine visual similarity between images and cannot identify digital tampering that has been performed through copying, splicing, or local modification. This invention performs multi-band wavelet decomposition on the denoised grayscale image, analyzes the differences in the energy distribution of JPEG compression noise in each sub-band, generates a suspected tampering heatmap, and calculates the centroid offset. This allows for precise location of tampering boundary regions, and even if the forger adjusts the position of local spots or Rf values, traces of tampering can be detected from inconsistencies in the noise history.

[0021] This invention solves the problem that existing technologies fail to utilize the physical constraints of thin-layer chromatography, making it difficult to detect Rf value forgery. For the first time, this invention transforms the physical definition of the Rf value (the ratio of spot distance to solvent front distance) into a verification condition. By detecting the distance from the solvent front to the origin and the center distance of each spot, the Rf value is calculated and a linear correlation coefficient is verified. When the Rf value exceeds the range of 0-1, or the linear correlation coefficient is below a threshold, or the solvent front cannot be uniquely determined, a physical inconsistency marker is generated, thereby effectively identifying Rf value forgery or data consistency anomalies.

[0022] This invention solves the problem of existing technologies having a single judgment mode and being unable to classify risks. Existing general image similarity algorithms and thin-layer chromatography plagiarism detection schemes all use a single threshold output (similar / dissimilar), which cannot differentiate processing based on the severity of tampering. This invention sets up a three-level threshold judge: when the compliance score is below the first threshold, it outputs a normal review process; when it is between the first and second thresholds, it outputs a review and generates a heatmap of the tampered area and a consistency report of the ratio shift value; when it is above the second threshold, it directly intercepts and freezes the associated experimental data, triggering a compliance audit event record, thus realizing intelligent risk classification and control.

[0023] This invention solves the problems of data lack of traceability and insufficient compliance. Existing technical solutions generally lack complete process records and cannot provide a traceable chain of verification evidence, making it difficult to meet the requirements of laboratory data compliance audits. This invention links and archives all deduplication processes, tampering detection results, physical consistency verification reports, and detection tasks in the LIMS system, forming a complete electronic record that ensures data traceability and compliance.

[0024] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0025] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0026] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the artificial intelligence-based thin-layer chromatography chromatogram duplication detection method in this invention; Figure 2 This is a diagram illustrating the process of normalizing pixel values ​​in a noise-reduced grayscale image using an activation function in this invention. Figure 3This is a diagram illustrating the process of generating the overall feature vector of the current map in this invention. Figure 4 This is a block diagram of the artificial intelligence-based thin-layer chromatography chromatogram duplication detection system in this invention; Figure 5 This is a standard thin-layer chromatogram; Figure 6 This is an abnormal thin-layer chromatogram. Detailed Implementation

[0027] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0028] This invention is applied to the compliance audit of thin-layer chromatography data in a Laboratory Information Management System (LIMS). After completing the thin-layer chromatography experiment, the inspector acquires the spectrum using a thin-layer imager and uploads it to the LIMS system. This invention automatically performs image preprocessing, feature extraction, tampering detection, physical consistency verification, and similarity quantification, ultimately outputting a compliance judgment result.

[0029] The overall algorithm process is divided into six stages: Image preprocessing stage (step 1) converts the original image into a standardized input image; Feature extraction stage (step 2) generates an overall feature vector and outputs a difference heatmap through a multi-layer convolutional network; Tampering detection stage (step 3) locates the tampered area through wavelet decomposition and JPEG noise analysis; Physical consistency verification stage (step 4) identifies data forgery through Rf value self-consistency detection; Multi-dimensional similarity calculation and weighted fusion stage (step 5) comprehensively scores the four similarity indicators with tampering traces and physical verification results; Three-level threshold determination stage (step 6) outputs three results based on the score: normal review, re-examination, or data interception.

[0030] This invention is applicable to quality control and data traceability management of thin-layer chromatography analysis in the fields of pharmaceuticals, food, and cosmetics.

[0031] Example 1: A method for checking duplicate thin-layer chromatograms based on artificial intelligence.

[0032] This embodiment provides an artificial intelligence-based method for deduplication of thin-layer chromatography images. This method is applied to a LIMS system for compliance auditing of uploaded thin-layer chromatography images. The following is combined with... Figure 1 The flowchart shown will be explained in detail.

[0033] Step 1: Image Preprocessing. Acquire raw thin-layer chromatography images (e.g., captured at 300 dpi resolution using a thin-layer chromatography imager and saved as JPEG format). First, scale the image proportionally to a uniform size (e.g., scale the longest side to 1024 pixels, and scale the short sides proportionally), then pad the edges (zero padding) to obtain a 1024×1024 pixel square image. Next, convert the image to grayscale (using a weighted average method: Gray=0.299R+0.587G+0.114B), and apply a Gaussian filter (window size 3×3, standard deviation σ=0.5) for background smoothing and noise reduction to obtain a denoised grayscale image. This Gaussian filter effectively suppresses sensor thermal noise while preserving the energy distribution characteristics of JPEG compressed noise, meeting the sensitive requirements of subsequent tampering detection steps (Step 3) regarding noise energy distribution differences.

[0034] Next, as Figure 2 As shown, the pixel value normalization process of the denoised grayscale image is performed using the hyperbolic tangent activation function: First, the grayscale value of each pixel is linearly mapped from integers of 0 to 255 to floating-point numbers of 0 to 1; then, the tanh function is applied for non-linear mapping so that the output value falls within the interval [-1, 1]; finally, the mean μ and standard deviation σ of all pixels are calculated, and x' = (x - μ) / σ is applied to each pixel value x to obtain a normalized input image with zero mean and unit variance. This normalization process helps to accelerate the convergence of the subsequent neural network.

[0035] Step 2: Feature Extraction and Overall Feature Vector Generation. The standardized input image is fed into a pre-trained multi-layer convolutional neural network (this embodiment uses a modified ResNet-50 architecture, removing the final fully connected layer and adding bypass outputs at different levels). The network extracts low-level feature maps through shallow convolutional kernels (e.g., the first set of convolutional layers, receptive field 3×3), which encode the location (coordinates), area (number of connected pixels), and grayscale intensity (statistical distribution of pixel values) of the spots. Mid-level convolutional layers (e.g., layers 10-20) extract intermediate morphological feature maps, encoding the shape (circularity, aspect ratio) and edge contours (gradient orientation histogram) of the spots. Deep convolutional layers (e.g., the last few layers) extract high-level semantic feature vectors, encoding the overall chromatographic distribution pattern (e.g., spatial arrangement order and relative distance of the spots).

[0036] In this embodiment, the multi-level feature fusion is specifically implemented as follows: Figure 3As shown, the shallow feature map is compressed into a shallow feature vector (e.g., 256 dimensions in this embodiment) using global average pooling, the mid-level feature map is compressed into a mid-level feature vector (e.g., 512 dimensions) using global average pooling, and the feature vector output from the deep convolution (e.g., 1024 dimensions) is placed side by side. Then, these three vectors are concatenated end to end to form a combined long vector with dimensions 256 + 512 + 1024 = 1792, which is then passed through two fully connected layers (the first layer outputs 512 dimensions with ReLU activation function; the second layer outputs 256 dimensions with no activation function), and finally L2 norm normalization is performed to obtain the overall feature vector of the current map (256 dimensions). At the same time, based on the Grad-CAM technique, the feature map of the last convolutional layer is weighted and summed with the classification weights to generate a difference heatmap, which is superimposed on the original image for output. The bright areas indicate that the current map and the historical map have significant feature differences at that position, and the dark areas indicate high similarity.

[0037] Step 3: Tampering Trace Detection. This step is used to detect whether the image has undergone tampering operations such as copying, splicing, or local modification. JPEG compression leaves a unique noise energy distribution pattern in the high-frequency subbands of the image. When different regions of the image have different compression histories, their noise energy distribution will be inconsistent. Specifically, a three-level Daubechies-8 wavelet decomposition is performed on the denoised grayscale image. Each decomposition generates one low-frequency approximate subband (LL) and three high-frequency detail subbands (horizontal LH, vertical HL, and diagonal HH). The three-level decomposition generates a total of 1 low-frequency subband and 9 high-frequency subbands (3×3). For each high-frequency subband, its noise energy histogram is calculated (e.g., the subband coefficient values ​​are quantized to 32 intervals, and the number of pixels in each interval is counted). In this embodiment, the noise energy distribution is represented by this histogram.

[0038] Calculate the noise energy histograms of the current and historical spectra in the corresponding subbands, and calculate the Kullback-Leibler divergence (KL divergence) between them as the dissimilarity. The calculation formula is: D KL =∑p(x)·log(p(x) / q(x)), where p(x) is the normalized histogram of the current spectral subband, and q(x) is the normalized histogram of the historical spectral subband. A preset threshold is used (set to 0.15 in this embodiment, a value obtained through extensive experimental statistics). When the KL divergence of any subband exceeds 0.15, the spatial frequency region corresponding to that subband is marked, and the tampering boundary region is reconstructed through inverse wavelet transform, generating a suspected tampering heatmap (the higher the pixel value in the heatmap, the greater the suspicion of tampering).

[0039] The calculation process for the centroid offset of the suspected tampered heatmap is as follows: First, the suspected tampered heatmap... Figure 2Value-based calculus (in this embodiment, the threshold is 0.6 times the maximum value of the heatmap; in practical applications, it can be adjusted according to image quality) is used to obtain the suspected tampering area (bright area). Then, the gray-level centroid coordinates (G) of all pixels in this area are calculated. x G y The calculation formula is: G x =∑(x·I(x,y)) / ΣI(x,y), G y =Σ(y·I (x,y) ) / ΣI (x,y) , where I (x,y) This represents the grayscale value of the heatmap. Simultaneously, the coordinates of the geometric center of the heatmap (C) are calculated. x C y = (width / 2, height / 2). Center offset d = sqrt((G) x -C x ) 2 +(G y -C y ) 2 The offset, measured in pixels, reflects the degree of deviation of the tampered area from the image center and serves as a feature component for subsequent comprehensive scoring.

[0040] Step 4: Physical Consistency Verification. The shift value Rf in thin-layer chromatography satisfies the physical definition: Rf i =d i / d f , where d i Let d be the distance from the center of the i-th spot to the origin. f Rf is the distance from the solvent front to the origin. In a legitimate, unaltered spectrum, the Rf of all spots is... i It should be between 0 and 1, and d i With d f There is a strong linear relationship between them (coefficient of determination R). 2 Approximately 1). The specific implementation process is as follows: Perform the following detection on the denoised grayscale image: Distance detection from the solvent front to the origin: The Canny edge detection operator (low threshold 50, high threshold 150) is used to extract image edges. Then, the Hough transform is applied (accumulator threshold set to 100, minimum line length set to 1 / 3 of the image width) to detect horizontal lines. The longest and most complete continuous horizontal edge is found in the upper half of the image (vertical coordinate 0 to 1 / 2 of the height). The vertical coordinate of this edge represents the distance d from the solvent front to the origin. f (The origin is the position of the bottom sample line of the image, which is pre-located through template matching). If the confidence score difference between two or more candidate edges is less than 10%, a warning "The solvent front cannot be uniquely determined" will be output.

[0041] Blob center distance detection: Adaptive thresholding (Otsu method) is applied to the denoised grayscale image, and all connected components with an area greater than 0.005% of the image width (e.g., approximately 50 pixels in a 1024×1024 image) are labeled as candidate blobs. For each candidate blob, its geometric center coordinates are calculated, and the ordinate of the center (y-axis, origin at the bottom of the image) is used as d. i Discard values ​​with a ordinate greater than or equal to d. f For invalid spots, or those less than or equal to 0, retain the ordinate between 0 and d. f Spots between the lines are considered valid spots; if the vertical coordinate is greater than the image height, it is also considered invalid.

[0042] The ratio shift value is calculated as follows: For each valid spot, calculate Rf. i =d i / d f .

[0043] The self-consistency check is: (i) if any Rf i <0 or Rf i >1, determine physical inconsistency; (ii) for all valid spots (d i ,Rf i* d f Perform linear regression (forced to pass through the origin) and calculate the coefficient of determination R. 2 If R 2 <0.95, determine physical inconsistency; (iii) if step (i) or (ii) is triggered, or if there is a "solvent front cannot be uniquely determined" warning, generate a physical inconsistency flag (flag value = 1), otherwise the flag is 0.

[0044] Step 5: Multi-dimensional similarity calculation and weighted fusion. The overall feature vector of the current map (dimension 256) and the overall feature vector of the historical map (same dimension) are compared using a four-dimensional similarity calculation: Mean Squared Error (MSE): The average of the squared differences of each element is used to obtain the error similarity index. In order to map MSE to the interval [0,1] and make it positively correlated with similarity, a transformation function is defined: S1=1 / (1+MSE), where S1 is the error similarity index. The larger the value of S1, the more similar the two feature vectors are (S1=1 when MSE=0, and S1 approaches 0 as MSE increases).

[0045] Mean cosine similarity: Calculate the dot product of two vectors and divide it by the product of their magnitudes to obtain the directional similarity index (S2, the closer the value is to 1, the more similar they are).

[0046] Average correlation coefficient: Calculate the product of covariance and standard deviation to obtain the linear correlation index (S3, the closer the value is to 1, the more similar the correlation).

[0047] Average Structural Similarity (SSIM): Treating two vectors as one-dimensional signal sequences, we calculate the similarity of brightness, contrast, and structure respectively, and then take the average to obtain the structural similarity index (S4, the closer the value is to 1, the more similar they are).

[0048] Weighted Fusion: The four indicators mentioned above are weighted and summed with the centroid offset d from step three and the physical inconsistency flag from step four to form a comprehensive compliance score. In this embodiment, the preset weighting coefficients are: ω1=0.2 (MSE, the higher the consistency after transformation, the better), ω2=0.2 (cosine), ω3=0.2 (correlation coefficient), ω4=0.2 (SSIM), ω5=0.1 (centroid offset, normalized to [0,1]), ω6=0.1 (physical inconsistency flag). Then Score=S1×ω1+S2×ω2+S3×ω3+S4×ω4+(1-d')×ω5+(1-flag)×ω6, where d' is the normalized value of the centroid offset (divided by the length of the image diagonal), and the Score value ranges from 0 to 1. The higher the value, the more compliant the map.

[0049] Step Six: Three-Level Threshold Determination and Output. The overall compliance score is compared in batches with historical data stored in the LIMS system to obtain the score value for each comparison. Based on two thresholds pre-determined through ROC curve analysis (in this embodiment, the first threshold Th1 = 0.75 (corresponding to a false positive rate ≤ 1%) and the second threshold Th2 = 0.92 (corresponding to a false negative rate ≤ 0.5%), a three-level determination is performed: If the score is greater than 0.92, it is determined that there is no risk of duplication and no trace of tampering. The normal review process is output, and the system automatically approves the graph.

[0050] If 0.75 ≤ Score ≤ 0.92: This is considered suspicious for tampering. The system generates a heatmap of the tampered area (overlaid on the original image, with red marking areas of inconsistent compression history, orange marking areas of Rf abnormality, and purple marking overlapping areas) and a ratio shift consistency report (containing the Rf values ​​of each spot). i Percentage deviation from historical data points, and R-squared value of linear regression. 2 The value is output to the review queue for a second manual review.

[0051] If the score is less than 0.75, it is determined to be highly suspected of being a duplicate or systematic tampering. The system will directly intercept the graph, freeze the associated experimental data records, and trigger a compliance audit event (recording the operator, time, graph hash value, judgment basis, etc.). At the same time, an alarm will be sent to the quality administrator.

[0052] Explanation of difference heatmaps and suspected tampering heatmaps: Difference heatmaps are used to assist manual review of overall feature differences between graphs, while suspected tampering heatmaps are used to locate areas of digital tampering. They can be displayed overlay, but they represent different technical information and serve different detection purposes.

[0053] Example 2: System Architecture. Corresponding to the above method examples, this invention also provides an artificial intelligence-based thin-layer chromatography chromatogram deduplication system. For example... Figure 4 As shown, the system includes the following modules: Image preprocessing module: Implements the function of step one, converting the original image into a standardized input image.

[0054] Feature extraction module: Implements the functions of feature extraction and overall feature vector generation in step two, and outputs a difference heatmap.

[0055] The tampering trace detection module includes (a) a multi-band compressed trace detection submodule, which realizes wavelet decomposition, noise energy distribution calculation, tampering suspected heat map generation and centroid offset calculation in step three; and (b) a physical self-consistency verification submodule, which realizes solvent front detection, spot recognition, Rf calculation and physical inconsistency marker generation in step four.

[0056] Similarity metric module: Implements the calculation of the four similarity metrics in step five.

[0057] Comprehensive scoring module: Implements the weighted fusion in step five and outputs a comprehensive compliance score.

[0058] LIMS Integration and Level 3 Judgment Module: Implements batch comparison and level 3 threshold judgment in step six, outputs corresponding results and triggers audit events.

[0059] Display module: Used to display the tampered area heatmap, the shift value consistency report, and the difference heatmap in the form of overlay layers.

[0060] Example 3: Deduplication Verification Based on Real Thin-Layer Chromatography. To fully demonstrate the feasibility and technical effectiveness of this invention in a real experimental scenario, this example uses two actual thin-layer chromatographic identification chromatograms (…). Figure 5 , Figure 6 ) as input samples. Figure 5 and Figure 6 All samples were obtained from thin-layer chromatography identification experiments of honeysuckle and wild honeysuckle medicinal materials. The spotting information and image characteristics are as follows.

[0061] Figure 5This is a standard thin-layer chromatogram. The spotting order from left to right is as follows: 1: LZ20233421001 test solution + Lonicera japonica saponin B reference solution (5 μl each); 2: LZ20233421007 test solution + Lonicera japonica saponin B reference solution (5 μl each); 3: LZ20233421001 test solution (10 μl); 4: LZ20233421007 test solution (10 μl); 5: Lonicera japonica reference material solution (10 μl); 6: Lonicera japonica reference material solution (10 μl); 7: Lonicera japonica saponin B reference solution (10 μl). In this chromatogram, spot 7 is only Lonicera japonica saponin B reference standard and should appear as a single, clear spot. Visual inspection was performed. Figure 5 All spots were in normal position and shape, with no additional abnormal spots.

[0062] Figure 6 This is an abnormal thin-layer chromatogram. Figure 6 and Figure 5 The order of the sampling is exactly the same, but Figure 6 An extra spot appeared at position 7 (above the spot of Lonicera japonica saponin B reference standard). This anomalous spot was simulated through digital image editing (copying and pasting a similar spot area from another normal pattern to this position), and therefore its JPEG compression history is inconsistent with the surrounding area, rather than simply being spot contamination.

[0063] right Figure 5 The plagiarism analysis of (normal spectrum) is as follows: Step 1: Image preprocessing, ... Figure 5 The original image (resolution 300dpi) is scaled proportionally to 1024×1024 pixels, and after grayscale conversion and Gaussian filtering (σ=0.5) for noise reduction, a denoised grayscale image is obtained; then, the hyperbolic tangent activation function is applied for normalization to form a standardized input image.

[0064] Step 2: Feature extraction. The standardized input image is fed into the modified ResNet-50 network to extract and fuse multi-level features, generating an overall feature vector (256 dimensions). Simultaneously, Grad-CAM is used to output a difference heatmap. Comparing the current map (Honeysuckle 5 control material) with the historical standard map (Mountain Honeysuckle control material), the heatmap shows that, except for Mountain Honeysuckle 6, the difference values ​​in all other regions are less than 0.1 (normalized value), which is in line with expectations.

[0065] Step 3: Tampering trace detection. Perform three-level Daubechies-8 wavelet decomposition on the denoised grayscale image and calculate the JPEG compression noise energy distribution of each sub-band. Figure 5The KL divergence of all subbands is less than 0.15 (maximum value is 0.07), no inconsistencies in compression history were detected, no suspected tampering heatmaps were generated, and the centroid offset is 0.

[0066] Step 4: Physical consistency verification, measuring the distance d from the solvent front to the origin. f (d) f (≈820 pixels). Identify spots for each track and calculate Rf. i =d i / d f The Rf value of the reference spot No. 7, Lonicera japonica saponin B, was 0.42, which meets the standard value.

[0067] Rf of all spots i All are in the range of 0 to 1, and with (d i ,Rf i ·d f The coefficient of determination for linear regression is 0.992 (>0.95), indicating a unique solvent front. The physical inconsistency flag is set to 0.

[0068] Step 5: Multi-dimensional similarity calculation and weighted fusion. The overall feature vector of the current map is compared with the standard map to obtain: S1 (error similarity, S1 = 1 / (1+MSE)) = 0.97, S2 (cosine similarity) = 0.99, S3 (correlation coefficient) = 0.98, S4 (SSIM) = 0.96. Using weights (ω1~ω6) = (0.2, 0.2, 0.2, 0.2, 0.1, 0.1), the score is: Score = 0.97×0.2 + 0.99×0.2 + 0.98×0.2 + 0.96×0.2 + 1×0.1 + 1×0.1 = 0.98.

[0069] Step 6: Level 3 threshold determination, Score = 0.98 > second threshold (0.92), determined as "no risk of duplication and no trace of tampering", output normal review process.

[0070] right Figure 6 The plagiarism analysis of the abnormal pattern (with an extra spot at position 7) is as follows: Steps one and two: with Figure 5 The same applies, so I won't repeat myself.

[0071] Step 3: Tamper trace detection. Figure 6Wavelet decomposition and noise energy distribution calculation are performed. In the spatial region corresponding to orbit 7 (the location of the extra spots), a high-frequency subband KL divergence of 0.23 (>0.15) is detected, indicating that the compression history of this region is inconsistent with other regions. The system generates a suspected tampering heatmap (stored internally as a two-dimensional heatmap data, where abnormal areas are highlighted). For ease of explanation in this embodiment, the heatmap image display is omitted, and only the location coordinates of the abnormal area are output: the geometric center of the abnormal spot is located at pixel coordinates (320, 450), with a radius of 12 pixels. The centroid offset is calculated to be 0.15 (normalized value) based on this heatmap data.

[0072] Step 4: Physical consistency verification. Detect d f =820 pixels. Spot identification was performed on orbital 7, revealing two spots: a normal spot (Rf = 0.42) and an abnormal spot (Rf = 0.68). The abnormal spot's Rf value exceeded the expected range for this sample (0.42 ± 0.05), and its d... i With d f In the linear regression, this outlier caused the coefficient of determination to drop to 0.86 (<0.95), and the physical inconsistency flag was set to 1.

[0073] Step 5: Weighted fusion. Among the four similarity indices, due to the influence of abnormal spots, S1 drops to 0.72, S2 drops to 0.85, S3 drops to 0.81, and S4 drops to 0.79. Therefore, Score = 0.72×0.2 + 0.85×0.2 + 0.81×0.2 + 0.79×0.2 + (1-0.15)×0.1 + (1-1)×0.1 = 0.634 + 0.085 = 0.719.

[0074] Step 6: Level 3 threshold determination. Score = 0.719 < first threshold (0.75), which is determined to be "highly suspected of being duplicated or systematically tampered with". The system directly intercepts the graph, freezes the associated experimental data, triggers a compliance audit event, and outputs the suspected tampering mark and the coordinates of the abnormal area (above track 7, near the center coordinates (320, 450)) for review.

[0075] The comparison results are as follows: Figure 5 (Normal spectra) were deemed compliant and released. Figure 6 (The extra spot on number 7) was accurately identified as suspected tampering and intercepted.

[0076] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of equivalents of this invention, this invention is also intended to include these modifications and variations.

Claims

1. An artificial intelligence-based thin layer chromatogram map duplicate checking method, characterized in that, Includes the following steps: Step 1: The original thin-layer chromatographic image is scaled proportionally and normalized with edge filling, and then grayscale and background noise reduction are performed to obtain a noise-reduced grayscale image that highlights the outline and positional features of the spots. The denoised grayscale image is normalized by the hyperbolic tangent activation function to form a standardized input image. Step 2: The standardized input image extracts and fuses multi-level features through a multi-layer convolutional network with built-in artificial intelligence algorithms to generate the overall feature vector of the current map; at the same time, it outputs a heatmap of the difference between the current map and the historical map. Step 3: Perform multi-band wavelet decomposition on the denoised grayscale image to obtain multiple frequency band sub-bands; Calculate the JPEG compressed noise energy distribution of the current spectrum and historical spectrum in each frequency band subband, and calculate the difference of JPEG compressed noise energy distribution in each frequency band subband; when the difference of any frequency band subband exceeds a preset threshold, generate a suspected tampering heat map including the tampering boundary position; The Euclidean distance between the gray-scale centroid coordinates of the bright area in the suspected tampering heatmap and the geometric center of the heatmap is calculated and used as the centroid offset of the suspected tampering heatmap. Step 4: Detect the distance from the solvent front to the origin and the distance from the center of each spot to the origin in the denoised grayscale image. Calculate the ratio of the distance from the center of each spot to the origin to the distance from the solvent front to the origin. Use this ratio as the ratio shift value of the spot and verify whether the ratio shift values ​​of all spots meet the preset physical self-consistency conditions. When it is detected that the distance from the solvent front to the origin cannot be uniquely determined, or the linear correlation coefficient between the spot center distance and the solvent front distance is lower than the preset threshold, or the ratio shift value of any spot exceeds the range of 0 to 1, it is determined that the spectrum has suspected ratio shift value forgery or local tampering, and a physical inconsistency mark is generated. Step 5: Perform multi-dimensional calculations on the overall feature vector of the current map and the overall feature vector of the historical map: calculate the error similarity index using the average mean square error, calculate the directional similarity index using the average cosine similarity, calculate the linear correlation index using the average correlation coefficient, and calculate the structural similarity index using the average structural similarity; then, weight and fuse these four indicators with the centroid offset of the suspected tampering heatmap obtained in Step 3 and the physical inconsistency markers from Step 4 according to preset weight coefficients to form a comprehensive compliance score; Step Six: Compare the comprehensive compliance score with the historical maps stored in the LIMS system in batches to obtain the compliance value for each comparison; input the compliance value into a preset three-level threshold judge: when the compliance value is higher than the second threshold, it is judged as having no risk of duplication and no trace of tampering, and a normal audit process output is formed; when the compliance value is between the first and second thresholds, it is judged as having suspected tampering, and a heat map of the tampered area and a report on the consistency of the ratio shift value are generated and output to the review queue for manual review; when the compliance value is lower than the first threshold, it is judged as highly suspected of duplication or systematic tampering, and the associated experimental data is directly intercepted and frozen, triggering the compliance audit event record.

2. The artificial intelligence-based thin layer chromatogram duplication checking method according to claim 1, characterized in that, The process of normalizing pixel values ​​in a denoised grayscale image using an activation function includes the following steps: Using the denoised grayscale image as the initial image, a linear stretching transformation is performed on the grayscale value of each pixel to map the original integer value range from 0 to 255 to the floating-point range from 0 to 1, thus obtaining a preliminary normalized image. Each pixel value in the pre-normalized image is input into the hyperbolic tangent activation function for non-linear mapping, so that the output value falls within the range of negative one to positive one, thus obtaining the activation-mapped image. Calculate the mean and standard deviation of all pixels in the activation mapping image. Subtract the mean from each pixel value and divide by the standard deviation to make the pixel distribution of the final image satisfy the statistical characteristics of zero mean and unit variance, thus obtaining the standardized input image. 3.The artificial intelligence-based thin layer chromatogram duplication checking method of claim 1, wherein, A multi-layer convolutional network performs shallow feature extraction to obtain a low-level feature map containing the location, area, and gray intensity of the spots. The low-level feature map is processed by a middle convolutional layer to form a medium-level morphological feature map containing the shape and edge contour of the spots. The medium-level morphological feature map is then processed by a deep convolutional layer to form a feature vector of the overall chromatographic distribution. 4.The artificial intelligence-based thin layer chromatogram duplication checking method of claim 1, wherein, In a difference heatmap, bright areas indicate significant differences, while dark areas indicate similarities.

5. The method for detecting duplicate thin-layer chromatograms based on artificial intelligence as described in claim 1, characterized in that, The process of generating the overall feature vector of the current map includes the following steps: Shallow feature maps, mid-level feature maps, and deep feature vectors are extracted from a multi-layer convolutional network. The shallow and mid-level feature maps are each passed through a global average pooling layer to compress the spatial dimension of each feature map into a single value, thus transforming the shallow feature map into a shallow feature vector and the mid-level feature map into a mid-level feature vector. The shallow feature vector, mid-level feature vector, and deep feature vector are then placed side by side to form three independent primary feature vector groups, denoted as the feature vector group. The shallow, middle, and deep feature vectors in the feature vector group are concatenated end to end in channel order to form a combined long vector whose dimension is the sum of the dimensions of the three. This combined long vector is input into the first fully connected layer, which performs a non-linear mapping through a linear transformation and the ReLU activation function to output an intermediate fused feature vector with reduced dimension. The intermediate fused feature vector is input into the second fully connected layer. The fully connected layer performs a linear transformation but does not add an activation function to output an original feature vector with a preset fixed dimension. Then, the original feature vector is subjected to L2 norm normalization, that is, the square root of the sum of the squares of all its elements is calculated, and then each element is divided by the square root. Finally, an overall feature vector with a consistent numerical range and a unique direction is obtained. The overall feature vector is the overall feature representation of the current thin-layer chromatogram.

6. The method for detecting duplicate thin-layer chromatograms based on artificial intelligence as described in claim 5, characterized in that, The process of outputting a dimension-reduced intermediate fused feature vector includes the following steps: The shallow feature vector, middle feature vector and deep feature vector in the feature vector group are concatenated end to end in channel order to form a combined long vector with a dimension equal to the sum of the dimensions of the three features, which is denoted as the combined long vector. The combined long vector is input into the first fully connected layer. The fully connected layer performs a linear transformation on the input vector, weights and sums each output neuron with all input neurons and adds a bias to obtain the linearly transformed intermediate vector. Then, each element in the intermediate vector is processed by the ReLU activation function, keeping all elements greater than zero unchanged and setting elements less than or equal to zero to zero, and finally outputting a non-linear intermediate fusion feature vector with reduced dimension. The intermediate fused feature vector is input into the second fully connected layer, which performs a linear transformation again and outputs an original feature vector with a preset fixed dimension; the original feature vector is an input normalized to the L2 norm.

7. The method for detecting duplicate thin-layer chromatograms based on artificial intelligence as described in claim 1, characterized in that, The error similarity index and the direction similarity index are calculated as follows: The overall feature vector of the current map and the overall feature vector of the historical map are used as two input vectors, denoted as the input pair; the input pair is simultaneously fed into the mean square error calculation module and the mean cosine similarity calculation module: in the mean square error calculation module, the square of the difference between the two vectors is calculated element by element and then the average is obtained to obtain the error similarity index; in the mean cosine similarity calculation module, the dot product of the two vectors is calculated and divided by the product of their magnitudes to obtain the direction similarity index; the error similarity index and the direction similarity index are combined into a two-dimensional intermediate vector.

8. The method for detecting duplicate thin-layer chromatograms based on artificial intelligence as described in claim 1, characterized in that, The linear correlation index and structural similarity index are calculated as follows: The input pair is simultaneously fed into the average correlation coefficient calculation module and the average structural similarity calculation module. In the average correlation coefficient calculation module, the covariance of the two vectors is calculated and divided by the product of their respective standard deviations to obtain the linear correlation index. In the average structural similarity calculation module, the two vectors are treated as one-dimensional signal sequences, and the similarity of brightness, contrast, and structure is calculated separately and then averaged to obtain the structural similarity index. The linear correlation index and the structural similarity index are then combined into a two-dimensional intermediate vector.

9. The method for detecting duplicate thin-layer chromatograms based on artificial intelligence as described in claim 1, characterized in that, The process of forming the weighted sum of the four indicators includes the following steps: the error similarity index calculated by the average mean square error, the direction similarity index calculated by the average cosine similarity, the linear correlation index calculated by the average correlation coefficient, and the structural similarity index calculated by the average structural similarity are sequentially concatenated in a preset fixed order to generate an array containing four numerical components. Read the four weight coefficients corresponding to error similarity, directional similarity, linear correlation and structural similarity respectively from the system configuration parameters, and the sum of the four weight coefficients is one; Perform multiplication on each of the four components in the array, multiplying each component by its corresponding weight coefficient to obtain four weighted values; perform summation on the four weighted values, and use the summation result as the weighted sum of the four indicators.

10. A thin-layer chromatography chromatogram plagiarism detection system based on artificial intelligence, used to implement the thin-layer chromatography chromatogram plagiarism detection method based on artificial intelligence as described in any one of claims 1-9, characterized in that, include: The image preprocessing module is used to normalize the size of the original thin-layer chromatography image by scaling it proportionally and filling the edges, and then perform grayscale conversion and background noise reduction to obtain a noise-reduced grayscale image that highlights the outline and positional features of the spots. The denoised grayscale image is further processed by the hyperbolic tangent activation function to normalize the pixel values, forming a standardized input image; The feature extraction module is used to extract multi-level features from the standardized input image through a multi-layer convolutional network with built-in artificial intelligence algorithms and fuse them to generate the overall feature vector of the current map; at the same time, it outputs a heatmap of the difference between the current map and the historical map. The tampering detection module includes: (a) Multi-band compression trace detection submodule, used to perform multi-band wavelet decomposition on the denoised grayscale image, calculate the noise energy distribution difference in each band, locate the tampering boundary area, generate a suspected tampering heat map; and calculate the Euclidean distance between the grayscale centroid of the bright area in the heat map and the geometric center of the heat map as the centroid offset. (b) Physical self-consistency verification submodule, used to detect the distance from the solvent front to the origin and the distance from the center of each spot to the origin in the noise-reduced grayscale image, calculate the ratio of the distance from the center of each spot to the origin to the distance from the solvent front to obtain the spot ratio shift value, and verify whether the ratio shift values ​​of all spots meet the physical self-consistency condition, and generate physical inconsistency markers. The similarity metric module is used to perform multi-dimensional calculations on the overall feature vector of the current map and the overall feature vector of the historical map, and obtain four similarity indices through average mean square error, average cosine similarity, average correlation coefficient and average structural similarity respectively. The comprehensive scoring module is used to weight and fuse the four similarity indicators with the centroid offset and physical inconsistency markers of the suspected tampering heatmap to form a comprehensive compliance score. The LIMS integration and three-level judgment module is used to compare the comprehensive compliance score with the historical maps stored in the LIMS system in batches, and call the three-level threshold judge: when the compliance value is higher than the second threshold, the normal review process is output; when the compliance value is between the first and second thresholds, a review reminder is output and a report on the consistency of the tampered area heat map and the ratio shift value is generated; when the compliance value is lower than the first threshold, the related experimental data is directly intercepted and frozen, triggering the compliance audit event record. The display module is used to display the tampered area heatmap, the ratio shift consistency report, and the difference heatmap in the form of overlay layers.

Images