A Visual Identification Method and System for Card Authenticity Based on Floral Pattern Anti-counterfeiting Features

Abstract

This invention relates to the field of computer vision and image recognition technology, and discloses a method and system for visually identifying the authenticity of cards based on floral anti-counterfeiting features. The method includes: macro imaging and adaptive preprocessing of the floral area, extraction of the floral line skeleton based on curvature flow analysis, hierarchical encoding of floral features and construction of multidimensional fingerprints, depth comparison based on Fourier descriptors and KL divergence, and counterfeit detection and comprehensive judgment based on generative adversarial networks. The accuracy rate of floral anti-counterfeiting feature recognition reaches 99.2%, the detection rate of high-quality counterfeit products reaches 98.5%, and the identification time per card is less than 3 seconds.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image recognition technology, specifically to a method and system for visually identifying the authenticity of cards based on floral anti-counterfeiting features. Background Technology

[0002] With the rapid development of the trading card market, the collectible value and trading price of high-end cards have continued to rise, with some rare cards reaching market values ​​of tens or even hundreds of thousands of RMB per card. At the same time, counterfeiting techniques are constantly evolving. High-quality counterfeit cards have achieved a high degree of similarity to genuine products in terms of material selection, printing processes, and surface treatment, posing a serious technical challenge to traditional visual identification methods. Currently, card authenticity authentication mainly relies on the human experience of professional appraisers, who make judgments by visually observing macroscopic features such as printing colors, paper texture, and surface gloss. However, this subjective experience-based authentication method has inherent drawbacks, including low efficiency, inconsistent standards, and a significantly higher misjudgment rate when facing sophisticated counterfeits, making it difficult to meet the actual needs of large-scale authentication.

[0003] In the anti-counterfeiting technology system of trading cards, floral patterns are a classic anti-counterfeiting element widely used in the printing process of high-end trading cards. Floral patterns are woven from precise geometric curves according to specific mathematical rules, and their subtle features such as line curvature, density distribution, and color gradation carry rich anti-counterfeiting information. Genuine trading cards typically use high-precision gravure printing or laser engraving processes for their floral patterns, achieving extremely high levels of craftsmanship in terms of line precision and color reproduction. Counterfeit cards, limited by equipment precision and technological levels, will inevitably show differences from genuine cards in the microscopic details of their floral patterns. Therefore, precise quantitative analysis and comparative testing of floral anti-counterfeiting features can provide objective and reliable technical evidence for authenticating trading cards.

[0004] Chinese patent CN119693961A discloses a method and system for judging the consistency and authenticity of seals based on content and anti-counterfeiting lines. This method acquires and preprocesses an image of the seal to be analyzed, uses target detection technology to locate the seal position, identifies the text content of the seal using an OCR algorithm, and compares the similarity of the anti-counterfeiting lines based on connected component analysis, thereby achieving the consistency and authenticity judgment of the seal. However, the technical solution in prior art 1 targets the authentication of seal-type targets. Its anti-counterfeiting line detection is based on simple connected component counting and area comparison, lacking the ability to analyze the multi-dimensional features of complex floral patterns. This solution does not involve the precise extraction of the floral line skeleton and vector structure restoration technology, and cannot achieve sub-pixel-level quantitative evaluation of the printing accuracy and registration deviation of the floral pattern. Furthermore, its authenticity judgment relies solely on a single similarity threshold comparison and does not introduce a deep learning-based counterfeit distribution deviation detection mechanism, limiting its ability to identify highly counterfeit products.

[0005] Furthermore, most existing anti-counterfeiting feature detection methods employ a global image similarity comparison strategy, which involves performing pixel-level or feature-level matching calculations between the image to be detected and the reference image at the overall level. When dealing with highly structured and hierarchical anti-counterfeiting elements such as floral patterns, this method is prone to insufficient detection sensitivity due to the global averaging of subtle local differences. Simultaneously, existing technologies generally lack the ability to specifically analyze process defects such as registration deviations, dot diffusion, and color drift that occur during the printing process, making it difficult to distinguish between normal printing tolerances of genuine products and systematic process deviations in counterfeit products.

[0006] Therefore, there is an urgent need to develop a multi-dimensional visual identification technology specifically for the anti-counterfeiting features of floral patterns on cards. This technology can perform in-depth analysis of floral patterns at multiple levels, from macro to micro. By combining advanced image analysis algorithms and deep learning models, it can achieve high-precision, high-efficiency, and high-robust automatic identification of genuine and counterfeit cards. Summary of the Invention

[0007] In order to overcome the technical problems of existing technologies, such as relying on human experience for card authenticity identification, lacking the ability to conduct multi-dimensional quantitative analysis of the anti-counterfeiting features of floral patterns, and insufficient identification accuracy when facing highly counterfeit products, this invention provides a visual identification method and system for card authenticity based on floral anti-counterfeiting features.

[0008] In a first aspect, the present invention provides a method for visually identifying the authenticity of cards based on the anti-counterfeiting feature of a floral pattern, comprising the following steps: Step S1, macro imaging and adaptive preprocessing of the floral pattern area: a high-resolution macro imaging unit is used to acquire images of the anti-counterfeiting floral pattern area on the card surface, and after white balance correction, illumination homogenization compensation, and multi-scale sharpening enhancement processing, the region of interest is cropped and sub-pixel level registered to obtain the floral pattern image to be analyzed; Step S2, extraction of the floral pattern line skeleton based on curvature flow analysis: the floral pattern image to be analyzed is subjected to multi-scale Gaussian difference filtering and skeleton thinning processing based on curvature flow evolution to extract the floral pattern line skeleton with a single pixel width, and topological connectivity repair and branch pruning optimization are performed; Step S3, floral pattern feature layer. The process involves hierarchical encoding and multidimensional fingerprint construction. Macroscopic layout features, mesoscopic structural features, and microscopic line features are extracted from the floral skeleton structure data and fused and encoded according to hierarchical weights to generate a multidimensional floral feature fingerprint. Step S4 involves depth comparison based on Fourier descriptors and KL divergence. The Fourier descriptor similarity and color space distribution KL divergence value between the multidimensional floral feature fingerprint and the genuine reference feature fingerprint are calculated to generate a depth comparison score. Step S5 involves forgery detection and comprehensive judgment based on generative adversarial networks. A pre-trained generative adversarial network forgery detection model outputs a forgery probability score, which is weighted and fused with the depth comparison score to generate a comprehensive authentication score. The authenticity judgment is then output based on a comparison of the comprehensive authentication score with a preset threshold.

[0009] Secondly, the present invention provides a card authenticity visual identification system based on floral anti-counterfeiting features, including a floral macro imaging and preprocessing module, a floral skeleton extraction module, a hierarchical coding and fingerprint construction module, a depth comparison module, and a counterfeit detection and comprehensive judgment module. The floral pattern macro imaging and preprocessing module acquires images of the floral pattern area on the card and performs adaptive preprocessing operations such as white balance correction, illumination homogenization compensation, and multi-scale sharpening enhancement to output a standardized floral pattern image for analysis. The floral pattern skeleton extraction module performs multi-scale Gaussian difference filtering and curvature flow-based skeleton thinning on the floral pattern image to extract the floral pattern line skeleton and completes topology repair and branch pruning. The hierarchical encoding and fingerprint construction module extracts layout features, structural features, and line features at macro, meso, and micro scales, respectively, and generates a multi-dimensional feature fingerprint of the floral pattern through hierarchical weight fusion encoding. The depth comparison module retrieves reference fingerprints from the genuine product feature database and performs Fourier descriptor similarity calculation and color space KL divergence comparison. The counterfeit detection and comprehensive judgment module outputs a counterfeit probability score through a generative adversarial network model and merges it with the depth comparison score to output the final identification conclusion. The above five modules form a data-driven cascaded processing link, with the output of the previous module serving as the key input of the next module. The comprehensive judgment result can be fed back to adjust the front-end imaging parameters and skeleton extraction parameters, forming a closed-loop optimized intelligent identification system.

[0010] The beneficial effects of this invention are as follows: By constructing a skeleton extraction algorithm for floral patterns based on curvature flow analysis, the original vector structure of the floral pattern design is accurately restored, achieving sub-pixel-level quantitative evaluation of printing accuracy and registration deviation; by establishing a three-level hierarchical coding system (macro-meso-micro), it supports cascaded authenticity screening from coarse to fine, significantly improving identification efficiency; by introducing a counterfeit detection module based on generative adversarial networks, it effectively identifies high-quality counterfeits that deviate from the distribution of genuine products. The accuracy rate of floral pattern anti-counterfeiting feature recognition reaches 99.2%, the detection rate of high-quality counterfeits reaches 98.5%, and the identification time per card is less than 3 seconds, effectively solving the problem of large-scale authenticity identification of high-end trading cards. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the method for visually identifying the authenticity of cards based on floral anti-counterfeiting features, provided in an embodiment of the present invention.

[0012] Figure 2 This is a schematic diagram of the structural architecture of the card authenticity visual identification system based on the floral anti-counterfeiting feature provided in an embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0014] Reference Figure 1 As shown in the figure, the visual identification method for card authenticity based on floral anti-counterfeiting features provided in this embodiment of the invention includes five core steps, from S1 to S5. These steps form a data-driven cascaded processing link, where the execution of each subsequent step depends on the output of the preceding step. Simultaneously, the judgment conclusion of step S5 can provide feedback to adjust the imaging parameters of step S1 and the skeleton extraction parameters of step S2, thereby constituting a closed-loop optimization mechanism to continuously improve identification accuracy. The specific implementation methods of each step are described in detail below.

[0015] Step S1: Macro imaging and adaptive preprocessing of the floral pattern area. In one embodiment of the present invention, the core task of step S1 is to acquire high-quality image data of the floral pattern anti-counterfeiting area on the surface of the card to be identified, and to eliminate interference factors such as uneven lighting, color shift and geometric distortion introduced during the acquisition process through a series of adaptive preprocessing operations, so as to lay a reliable data foundation for subsequent feature extraction and analysis.

[0016] Specifically, this embodiment employs a high-resolution macro imaging unit to acquire images of the surface of the card to be identified. Preferably, the optical resolution of the macro imaging unit is no less than 1200 dpi to ensure that the fine structure of the floral pattern, with line widths as low as 0.05 mm, can be captured. The imaging unit is equipped with a ring-shaped LED light source array, employing a diffuse illumination mode to effectively avoid high-gloss interference caused by local specular reflection on the card surface. In one embodiment of the invention, the color temperature of the LED light source is set to 5500K to simulate standard daylight conditions, ensuring that the color reproduction of the acquired image meets the accuracy requirements of subsequent color analysis. The imaging sensor uses a CMOS area array sensor with an effective pixel count of no less than 20 million pixels, and the single-frame acquisition time is controlled within 50 ms, supporting continuous high-speed acquisition to meet the throughput requirements of batch identification scenarios.

[0017] After acquiring the original image of the floral pattern, the system first performs a white balance correction. This embodiment uses the gray world hypothesis algorithm to automatically white balance the original floral image. By calculating the mean values ​​of the image in the R, G, and B channels and adjusting the gain of the R and B channels based on the G channel, color shifts caused by light source color temperature deviations are eliminated. Preferably, the mean deviation of each channel after white balance correction is controlled within ±2%.

[0018] Subsequently, the system performs illumination homogenization compensation processing. This embodiment employs homomorphic filtering to compensate for image illumination. Specifically, the image is transformed from the spatial domain to the logarithmic domain, then a high-pass filter is used to suppress low-frequency illumination components and enhance high-frequency detail components. Finally, the image is transformed back to the spatial domain to obtain the illumination-homogenized image. In one embodiment of the invention, the cutoff frequency of the high-pass filter is set to the spatial frequency corresponding to 1 / 20 of the image size, and the high-frequency gain coefficient... =1.5, low-frequency gain coefficient =0.5, to effectively suppress large-area lighting gradients while preserving the details of the floral pattern.

[0019] After illumination compensation, the system further performs multi-scale sharpening enhancement processing. This embodiment employs the Laplacian pyramid decomposition method, decomposing the image into 3-5 sub-band images of different resolutions, and applying sharpening enhancement of different intensities to each sub-band image. Preferably, the enhancement coefficient of each layer is adaptively adjusted within the range of 1.2-2.5 according to the signal-to-noise ratio of that layer. Lower-resolution layers use larger enhancement coefficients to highlight macroscopic structural contours, while higher-resolution layers use smaller enhancement coefficients to enhance line details while avoiding noise amplification. The enhanced sub-band images are then recombined into a complete, normalized floral image.

[0020] Finally, the system performs region-of-interest (ROI) cropping and subpixel-level registration operations on the standardized floral pattern image based on the geometric center localization result. This embodiment first uses Hough transform to detect the outer contour circle or ellipse of the floral pattern to determine the geometric center coordinates. The system first defines the bounding rectangle range; then, using the geometric center as a reference, it extends outwards by a preset margin to crop the region of interest; finally, a phase-correlation-based sub-pixel registration algorithm is used to precisely align the cropped image with the standard template, achieving a registration accuracy of 0.1 pixels. The resulting floral image has a uniform size, consistent lighting conditions, and precise geometric alignment, providing high-quality input data for feature extraction in subsequent steps. It is worth noting that the adaptive preprocessing strategy in this embodiment does not use fixed parameter configurations but dynamically adjusts them according to the actual conditions of the acquisition environment. In a preferred embodiment of this invention, the system first performs a rapid calibration shot on the standard grayscale color chart before each acquisition, automatically adjusting preprocessing parameters such as white balance gain, illumination compensation intensity, and sharpening enhancement coefficient based on the measurement deviation of each grayscale value on the grayscale color chart. This self-calibration mechanism ensures consistent preprocessing output quality under different working environments, effectively improving the system's environmental adaptability and the repeatability of the identification results. In addition, the preprocessing module is equipped with an image quality assessment function, which performs blur and contrast detection on the acquired original floral images. When the image quality is lower than the preset minimum acceptable standard, the re-acquisition process is automatically triggered to avoid introducing low-quality input data into the subsequent analysis process.

[0021] Step S2: Extraction of the skeleton of the floral pattern lines based on curvature flow analysis. In one embodiment of the present invention, the core task of step S2 is to accurately extract the central skeleton structure of the floral pattern lines from the image to be analyzed. The floral pattern is composed of a large number of intricate curves interwoven together. Directly analyzing the original pixel image makes it difficult to effectively characterize the geometric topological features of the lines. Therefore, this step uses a curvature flow-driven skeleton thinning algorithm to shrink the line regions of varying widths to skeleton lines of single pixel width, accurately restoring the original vector structure of the floral design.

[0022] First, the system performs multi-scale Gaussian difference filtering on the floral image to be analyzed to enhance the line edge response. This embodiment constructs a Gaussian difference filter bank containing three scale levels, where the first level's Gaussian kernel standard deviation... =0.5 corresponds to the detection of microscopic line edges, second level. =1.0 corresponds to edge enhancement for medium-width lines, third level. =2.0 corresponds to the enhancement of the macroscopic structural contour. The filter responses of each level are fused into a single edge-enhanced image through a weighted summation, with the fusion weights being respectively... =0.5、 =0.3 and =0.2, which makes the edges of the micro lines dominate the fusion result.

[0023] Based on the edge-enhanced image, the system performs skeleton thinning processing based on curvature flow evolution. The core idea of ​​the curvature flow evolution algorithm is to calculate the local curvature value of each pixel on the boundary of the line region. The curvature value is used as the driving force to shrink the boundary along the normal direction, so that the high curvature area (such as the two sides of the line) shrinks faster, while the low curvature area (such as the end point and intersection of the line) shrinks slower. Thus, the line area is gradually shrunk to the central axis position while keeping the line topology unchanged.

[0024] In one embodiment of the present invention, the specific implementation process of the curvature flow evolution algorithm is as follows: First, the edge enhancement image is binarized using the Otsu thresholding method to separate the floral line region and the background region; then, sub-pixel level contour extraction is performed on the boundary of the line region to obtain the parameterized representation of the boundary curve; next, an iterative shrinkage loop is entered. In each iteration, the local curvature value is calculated for each sampling point on the boundary curve, and the shrinkage displacement is determined according to the following curvature flow evolution equation: ,in: To parameterize the boundary curve in arc length parameter and time parameters The location coordinates, in pixels; This represents the local curvature value at that location, in pixels. The calculation method is to divide the cross product of the second derivative and the first derivative of the curve by the cube of the magnitude of the first derivative. Let be the dimensionless unit normal vector pointing inwards from the line region at this location. This equation shows that the velocity of the boundary point is proportional to its curvature and its direction is along the inward normal direction, thus achieving curvature-driven contraction of the boundary.

[0025] After each iteration, the system performs topology checks on the shrunken boundary curves to determine if self-intersections or breaks have occurred. This embodiment employs a scan-line-based topology consistency check method. After each iteration, it checks whether the Euler number of the boundary curves has changed. If the Euler number changes, it indicates that the topology has been disrupted. In this case, it reverts to the previous step, halves the shrinkage step size for that region, and re-executes the process. The iteration convergence threshold is set to 0.01 pixels, meaning that skeleton extraction is considered complete when the maximum displacement of all boundary points between two adjacent iterations is less than 0.01 pixels. The maximum number of iterations is set to 500 to prevent infinite loops in extreme cases.

[0026] After curvature flow evolution is complete, the system performs topology connectivity repair and branch pruning optimization on the initially extracted floral line skeleton. The topology repair stage mainly addresses the skeleton breakage problem caused by digitization errors or noise. This embodiment adopts an automatic connection strategy based on endpoint distance. When the distance between the endpoints of two skeleton line segments is less than a preset connection threshold... When the pixel value is 3, a smooth connection curve segment is generated between the two endpoints using cubic spline interpolation. The branch pruning stage is used to remove false short branches caused by noise. In this embodiment, branches with a length less than a preset pruning threshold are removed. End branches with a resolution of 10 pixels were identified as noise branches and deleted. The resulting floral skeleton structure data, after repair and optimization, fully and accurately reflects the topological geometric features of the floral pattern, including the direction of the lines, the distribution of intersections, and the overall symmetrical structure.

[0027] Step S3: Hierarchical Encoding and Multidimensional Fingerprint Construction of Floral Pattern Features. In one embodiment of the present invention, the core task of step S3 is to extract complementary feature representations from the floral pattern skeleton structure data output in step S2 at three different spatial scales: macroscopic, mesoscopic, and microscopic. These multi-scale features are then integrated into a compact multidimensional floral pattern feature fingerprint vector through a hierarchical weighted fusion encoding mechanism for subsequent authentication and detection. This hierarchical feature representation strategy stems from the inherent characteristics of the floral pattern anti-counterfeiting design—the floral pattern presents different structural information at different observation scales. At the macroscopic level, it reflects the symmetry and balance of the overall layout; at the mesoscopic level, it reflects the arc distribution and density variation of the line groups; and at the microscopic level, it reflects the color transition and outline integrity of individual lines.

[0028] In the macro-layout feature extraction stage, the system first performs a global topological analysis on the floral skeleton structure data. Specifically, this embodiment transforms the floral skeleton into a graph structure representation in the sense of graph theory, where the intersections of skeleton lines are the nodes of the graph, and the skeleton line segments between the intersections are the edges of the graph. Based on this graph structure, the system extracts the following macro-layout features: global topological connectivity features, including a histogram of node degree distribution, the number of connected components, and the Betti number of the skeleton; and a symmetry distribution index, which quantifies the degree of symmetry of the floral pattern by calculating the autocorrelation coefficients of the skeleton graph structure in multiple symmetry axis directions. Preferably, the symmetry calculation uses eight evenly distributed angular directions, that is, the symmetry autocorrelation value is calculated every 22.5°, ultimately obtaining an 8-dimensional symmetry feature vector. In addition, the macro-features also include the centroid coordinate offset and principal axis direction angle of the skeleton, used to quantify the printing positioning accuracy of the floral pattern on the card.

[0029] In the mesoscopic structural feature extraction stage, the system divides the floral skeleton into multiple local regions and analyzes the structural characteristics of line groups within each region. This embodiment employs a concentric ring-zone division strategy centered on the geometric center of the floral pattern, dividing the floral region into 4-6 rings. The width of each ring is adaptively determined based on the radial line density of the floral pattern. Within each ring, the system extracts a histogram of line curvature distribution. Specifically, it calculates the curvature value point-by-point for all skeleton line segments within the ring, and then statistically plots the curvature value distribution histogram. The histogram has 32 bins, covering a curvature range of [missing information]. ,in This represents the 95th percentile of the absolute value of curvature within the annulus. Simultaneously, the system extracts the density gradient vector, and characterizes the density gradient of the floral pattern from the center to the periphery by calculating the rate of change in line density between adjacent annulus zones. Line density is defined as the total length of the skeleton line segments per unit area, with units of pixels per pixel. The density values ​​of each annulus form a vector, and the difference between adjacent elements constitutes the gradient vector of the density gradient.

[0030] In the micro-line feature extraction stage, the system delves into the pixel-level details of individual lines for analysis. This embodiment performs color transition gradient analysis on the original floral image regions on both sides of the skeleton lines. Specifically, along the normal direction of the skeleton lines, a strip region 5-10 pixels wide is taken on each side of the line, and the gradient distribution of RGB color values ​​from the center of the line outwards within this strip region is calculated. Genuine cards typically exhibit smooth and regular color transitions in their floral lines, while counterfeit cards, due to insufficient printing precision, often show step-like or irregular fluctuations in color transitions. This embodiment discretizes the color transition gradient into a vector representation, using 10 sampling points along the normal direction, and calculating on the R, G, and B channels respectively, resulting in a 30-dimensional color transition gradient feature vector.

[0031] In addition, microscopic features also include a measure of contour closure integrity. Floral patterns contain numerous closed curve loops. Genuine cards exhibit strict closure of these loops, while counterfeit cards often show loops that are not closed or have gaps or burrs at the closure points. This embodiment calculates the gap distance between the start and end points of each loop by traversing all closed loops in the floral pattern framework, and then counts those gap distances exceeding a preset closure threshold. The loop percentage of 0.5 pixels is used as the contour closure integrity metric. Preferably, the closure integrity metric ranges from 0 to 1, where 1 indicates that all loops are strictly closed.

[0032] In the hierarchical fusion encoding stage, the system weights and concatenates the features from the three scales according to hierarchical weights, and then uses principal component analysis to reduce dimensionality to obtain the final multidimensional feature fingerprint vector of the floral pattern. Specifically, the macroscopic layout feature vector has a dimension of approximately 20-30, the mesoscopic structure feature vector has a dimension of approximately 128-192 (depending on the number of rings and the number of histogram bins), and the microscopic line feature vector has a dimension of approximately 30-50. These three are weighted according to hierarchical weights. =0.25、 =0.40 and =0.35 is used for weighted concatenation (where M represents macroscopic, S represents structure, and L represents line). The concatenated high-dimensional vector is then reduced to the target dimension using principal component analysis. =256 dimensions, forming a compact cluster of multidimensional feature fingerprints. Preferably, the cumulative variance contribution rate retained during the dimensionality reduction process is not less than 95% to ensure that the feature fingerprint after dimensionality reduction still retains sufficient discriminative power.

[0033] Step S4: Based on the deep comparison of Fourier descriptors and KL divergence, in one embodiment of the present invention, the core task of step S4 is to perform a precise quantitative comparison between the multi-dimensional feature fingerprint of the floral pattern output in step S3 and the reference feature fingerprint in the genuine product feature database. By calculating two complementary metrics, Fourier descriptor similarity and KL divergence in color space distribution, the degree of similarity between the card to be identified and the genuine product is comprehensively evaluated from two dimensions: geometric shape and color distribution. This two-dimensional comparison strategy can effectively deal with high-level counterfeiting methods that only make detailed imitations in geometric structure or only in color representation.

[0034] In terms of Fourier descriptor similarity calculation, the system first performs equal-interval resampling on the contour curve of the floral pattern skeleton. In this embodiment, sampling is performed along the skeleton contour curve at equal arc length intervals, with the number of sampling points... The number of samples was set to 512 to ensure sufficient sampling density for frequency domain analysis while preserving contour details. The resampled contour curves were then represented as a complex sequence. ,in and The first The x and y coordinates of each sampling point , The imaginary unit is used. Performing a discrete Fourier transform on this complex sequence yields the Fourier descriptor coefficient sequence: ,in: For the first The Fourier descriptor coefficients are complex values ​​with moduli of 1. This reflects the frequency of the profile curve. The amplitude of the harmonic components, in pixels; The frequency order has a range of values. ; The base of the natural constant; π is the mathematical constant of a circle. To achieve translation, rotation, and scaling invariance description of the contour, this embodiment normalizes the Fourier descriptor coefficients by removing the zeroth-order components. To eliminate the effect of translation, divide all coefficients by To eliminate the scaling effect, the modulus of each order coefficient is taken to eliminate the effects of rotation and starting point. The final result is retained before... =64th order normalized Fourier descriptor as contour morphological feature representation.

[0035] Based on this, the Fourier descriptor similarity between the card to be identified and the authentic reference is calculated. : ,in: The Fourier descriptor similarity score ranges from 0 to 1, with values ​​closer to 1 indicating greater geometric similarity. The first card to be identified The normalized Fourier descriptor modulus value is dimensionless; The first reference for genuine products The normalized Fourier descriptor modulus value is dimensionless; =64 represents the order of the retained Fourier descriptors. The physical meaning of this formula is to measure the degree of difference between two sets of Fourier descriptors using normalized Euclidean distance, converting it into a similarity value ranging from 0 to 1.

[0036] Regarding the KL divergence calculation for color space distribution, the system performs statistical analysis on the color distribution of the floral pattern area. This embodiment first converts the pixels in the floral pattern area from the RGB color space to the Lab color space, which better matches the perceptual characteristics of the human eye and can better reflect the visual differences between colors. Subsequently, the histogram distribution of pixel values ​​is statistically analyzed on the L, a, and b channels, with each channel's histogram bin number set to 64. The histograms of the three channels are combined to form a 192-dimensional color distribution vector. After normalizing the color distribution vector to a probability distribution, the system calculates the KL divergence between the card to be identified and the authentic reference card. ,in: The KL divergence value has a range of values. The smaller the value, the more similar the two distributions are; it is dimensionless. The color distribution of the card to be identified is in the first position. The probability values ​​on each bin satisfy the following condition. ; The color distribution of the genuine product is shown in section [number]. The probability values ​​on each bin satisfy the following condition. ; =192 represents the total number of bins in the histogram (64 bins each for L, a, and b); The function is the natural logarithm. To avoid division by zero errors caused by a probability value of zero, this embodiment performs Laplace smoothing on the histogram, that is, increments the count value of each bin by 1 and then renormalizes it.

[0037] The system then converts the KL divergence values ​​into color similarity values ​​ranging from 0 to 1. :

[0038] ,in: The similarity of color space distribution is given, with a value range of [value range missing]. The closer the value is to 1, the more similar the color distribution; This is the attenuation coefficient, and its value is selected based on ensuring that the color similarity corresponding to the typical KL divergence value (approximately 0.01-0.05) between genuine cards is greater than 0.95, preferably... =1.0; is the base of the natural constant.

[0039] Finally, the system performs a linear weighted fusion of Fourier descriptor similarity and color similarity to generate a depth comparison score. : ,in: This is the depth comparison score, with a value range of [value range missing]. ; =0.6 is the weight coefficient of Fourier descriptor similarity. This weight is high because geometric features have stronger discriminative power in the anti-counterfeiting of floral patterns. =0.4 is the weighting coefficient for color similarity; .

[0040] Step S5: Counterfeit Detection and Comprehensive Judgment Based on Generative Adversarial Networks. In one embodiment of the present invention, step S5 introduces a counterfeit detection mechanism based on generative adversarial networks (GANs), serving as a powerful supplement to the deterministic comparison method in step S4. While traditional feature-based counterfeit determination methods are highly effective in detecting legitimate counterfeits, their ability to distinguish genuine counterfeits decreases when faced with meticulously designed high-quality counterfeits that mimic the feature distribution of genuine products. Generative adversarial networks, by learning the inherent generation distribution of the floral pattern on genuine products, can identify anomalous samples that deviate from genuine products at the probability distribution level, thus providing a probabilistic detection method that complements deterministic comparison.

[0041] The generative adversarial network (GAN) forgery detection model used in this embodiment comprises two core components: a generator and a discriminator. The generator uses a U-Net architecture, with its encoder containing five downsampled convolutional layers. Each layer has a 4×4 kernel size and a stride of 2, with the number of feature channels in each layer being 64, 128, 256, 512, and 512, respectively. The encoder compresses the input floral image into a high-level semantic feature representation through layer-by-layer downsampling. The decoder contains five upsampled deconvolutional layers symmetrical to the encoder. Skip connections are used to concatenate and fuse the features from each layer of the encoder with the corresponding features from the corresponding layer of the decoder to preserve image detail. The generator's function is to attempt to reconstruct the input floral image. Its reconstruction quality directly reflects the degree of similarity between the input image and the distribution of genuine products. If the input is genuine, the generator can reconstruct it with high quality and a small reconstruction error; if the input is a forgery, its feature distribution deviates from the genuine product distribution learned by the generator, leading to a decrease in reconstruction quality and a larger reconstruction error.

[0042] The discriminator employs a PatchGAN structure with a receptive field of 70×70 pixels. The input is either the original floral image or an image reconstructed by the generator. The output is a matrix proportional to the spatial size of the input image, where each element represents the probability that a corresponding local region is a real image. The discriminator contains four convolutional layers with a 4×4 kernel size, a stride of 2, and feature channels of 64, 128, 256, and 512 respectively. The last convolutional layer outputs a single-channel probability map. The advantage of the PatchGAN structure lies in its ability to focus on local texture details, making it more sensitive to microscopic differences at the level of floral lines compared to traditional global discriminators.

[0043] During the calculation of the forgery probability score, the system simultaneously inputs the floral image to be analyzed into both the generator and the discriminator. The generator outputs the reconstructed image. Then, the system calculates the original image. With reconstructed images Pixel-by-pixel L1 reconstruction error: ,in: The average reconstruction error has a range of values. The unit is grayscale value; and In coordinates of the original image and the reconstructed image respectively The pixel grayscale value at that location; and These represent the height and width of the image, respectively, in pixels. Simultaneously, the discriminator outputs a discrimination score for the original image. Forgery probability score Determined jointly by reconstruction error and discrimination score: ,in: To determine the probability of falsification, the value range is: The closer the value is to 1, the more likely it is to be a counterfeit. The sigmoid activation function maps the input to... interval; =0.05 is the scaling factor for reconstruction error. Its value is chosen to ensure that the contribution of the reconstruction error (about 2-5 gray values) of a typical genuine product is within a reasonable range and is dimensionless. =2.0 is the scaling factor for the discriminant score, which is dimensionless; =-1.0 is the bias term, used to adjust the position of the decision boundary, and is dimensionless.

[0044] Finally, the system uses the depth alignment score obtained in step S4. The forgery probability score obtained in step S5 Weighted fusion is performed to generate a comprehensive identification score. : ,in: To comprehensively identify the score, the value range is: The closer the value is to 1, the more likely the card is to be genuine; =0.6 is the fusion weight of the depth comparison score. This weight is high because deterministic feature comparison has higher confidence when the genuine template is known. =0.4 represents the score fusion weight based on generative adversarial networks; The system will combine the identification score with a preset authenticity threshold. When comparing, When the card to be identified is determined to be genuine, If the card is immediately identified as counterfeit, it is determined to be genuine. Preferably, the threshold for determining authenticity is... The recommended value range is 0.65-0.85, which can be adjusted according to the anti-counterfeiting complexity of different card series and the tolerance for misjudgment in the authentication scenario. Generally, a higher threshold (such as 0.80) is used in the authentication of high-end rare cards to reduce the risk of misjudgment.

[0045] In a preferred embodiment of the present invention, step S5 further includes a closed-loop feedback adjustment mechanism. The system feeds back the comprehensive identification score and intermediate feature data during the identification process to the front-end step to dynamically optimize the processing parameters of subsequent identification tasks. Specifically, when the comprehensive identification score is in the uncertainty range near a threshold ( When the sample is identified as a boundary sample requiring verification, the system automatically adjusts the exposure parameters and focus distance of the macro imaging unit in step S1 to acquire higher-quality floral images during verification. Simultaneously, the system feeds back the frequency order information of the largest Fourier descriptor deviation during depth comparison to step S2, which is used to adaptively adjust the scale parameters of the multi-scale Gaussian difference filter, resulting in higher resolution for skeleton extraction within the frequency range where deviations are significant. This closed-loop feedback mechanism allows the entire identification system to continuously learn and optimize during operation, gradually improving the accuracy of identifying various forgery methods.

[0046] Reference Figure 2 As shown in the figure, this embodiment of the invention also provides a card authenticity visual identification system based on floral anti-counterfeiting features. This system includes a floral macro imaging and preprocessing module, a floral skeleton extraction module, a hierarchical coding and fingerprint construction module, a depth comparison module, and a counterfeit detection and comprehensive judgment module. These five functional modules correspond one-to-one with steps S1 to S5 in the aforementioned method embodiment. The modules are cascaded together through standardized data interfaces to form an end-to-end automated identification processing chain.

[0047] The floral macro imaging and preprocessing module corresponds to the function implementation in step S1. This module integrates a high-resolution macro imaging unit and an image preprocessing subunit. The hardware configuration of the macro imaging unit includes a CMOS area array sensor with an optical resolution of no less than 1200 dpi, a ring-shaped LED diffuse reflection light source array with a color temperature of 5500K, and an electric focusing mechanism. The image preprocessing subunit incorporates a white balance correction engine, a homomorphic filtering illumination compensation engine, a Laplacian pyramid multi-scale sharpening engine, and a floral region localization and sub-pixel registration engine based on Hough transform. This module receives physical cards as input, outputs standardized floral images to be analyzed, and receives imaging parameter adjustment instructions from the counterfeit detection and comprehensive judgment module through a feedback interface to achieve dynamic optimization of exposure parameters and focus distance. In one embodiment of the present invention, the floral macro imaging and preprocessing module supports a working efficiency of acquiring and processing no less than 20 cards per minute, meeting the throughput requirements of batch identification scenarios.

[0048] The floral skeleton extraction module corresponds to the function implementation in step S2. This module incorporates a multi-scale Gaussian difference filter bank and a skeleton refinement engine based on curvature flow evolution. =0.5、 =1.0 and =2.0 three-scale Gaussian difference filter, with fusion weights of 0.5, 0.3 and 0.2 for each layer response. The skeleton refinement engine implements the curvature flow evolution iterative algorithm as detailed in the method embodiment, with a convergence threshold of 0.01 pixels and a maximum number of iterations of 500. This module also includes a topology repair subunit and a branch pruning subunit, which respectively adopt an endpoint distance connection strategy (connection threshold) =3 pixels) and short branch deletion strategy (pruning threshold) =10 pixels). The module receives the image of the floral pattern to be analyzed as input, outputs complete floral skeleton structure data, and receives scale parameter adjustment information from the depth comparison module through the feedback interface to optimize the filter configuration.

[0049] The hierarchical coding and fingerprint construction module corresponds to the functional implementation of step S3. This module includes three parallel feature extraction subunits and one fusion coding subunit. The macroscopic feature extraction subunit is responsible for converting the floral skeleton into a graph structure representation and extracting global topological features and symmetry distribution indices. The mesoscopic feature extraction subunit is responsible for dividing the skeleton into regions according to a concentric ring partitioning strategy and calculating the line curvature distribution histogram and density gradient vector within each ring. The microscopic feature extraction subunit is responsible for performing color transition gradient analysis and contour closure integrity detection on individual lines. The fusion coding subunit receives the output features from the three extraction subunits and, according to... =0.25、 =0.40、 A weighted concatenation algorithm with a hierarchical weight of 0.35 is used, and principal component analysis is applied to reduce the dimensionality of the concatenated features to 256 dimensions, ultimately generating a multi-dimensional floral fingerprint. The output of this module is simultaneously connected to the deep comparison module and the update interface of the genuine product feature database, supporting the continuous enrichment of the genuine product feature library during the accumulation of authentication data.

[0050] The depth comparison module corresponds to the functionality of step S4. This module includes a genuine product feature database retrieval subunit, a Fourier descriptor calculation subunit, and a KL divergence calculation subunit. The genuine product feature database is indexed and organized according to card series, version, and rarity, supporting fast and accurate retrieval based on card model, with a retrieval response time controlled within 50ms. The Fourier descriptor calculation subunit implements the entire process of contour resampling, discrete Fourier transform, and normalized similarity calculation as detailed in the method embodiment. The KL divergence calculation subunit implements the entire process of Lab color space conversion, histogram statistics, and KL divergence calculation. The outputs of the two subunits are linearly weighted (… =0.6, =0.4) is merged into a deep comparison score, which is then passed to the forgery detection and comprehensive judgment module.

[0051] The forgery detection and comprehensive judgment module corresponds to the function implementation in step S5. This module includes a generative adversarial network (GAN) inference engine, a forgery probability calculation subunit, and a comprehensive judgment subunit. The GAN inference engine loads the pre-trained U-Net generator and PatchGAN discriminator model parameters, performs forward inference on the input floral image to be analyzed, and generates a reconstructed image and a discrimination score. Preferably, the inference engine is deployed using GPU acceleration, and the inference time for a single image is controlled within 100ms. The forgery probability calculation subunit calculates the forgery probability score based on the reconstruction error and the discrimination score using the sigmoid fusion formula. The comprehensive judgment subunit combines the depth comparison score and the forgery probability score according to... =0.6、 A weighted fusion of values ​​with a value of 0.4 is performed to obtain a comprehensive identification score, which is then compared with the authenticity determination threshold. The final card authenticity determination is output after comparison. This module also integrates a closed-loop feedback controller, which automatically sends parameter optimization instructions to the floral macro imaging and preprocessing module and scale adjustment suggestions to the floral skeleton extraction module when the identification result is in the uncertainty range, thereby achieving system-level closed-loop adaptive optimization.

[0052] The modules in the above system can be deployed on the same server and cascaded through function calls between software modules, or they can be deployed as a microservice architecture and communicate through network interfaces. In a preferred embodiment of the invention, the system adopts an edge computing plus cloud collaboration deployment mode. The floral macro imaging and preprocessing module and the floral skeleton extraction module are deployed on edge devices to achieve low-latency image acquisition and preliminary processing. The hierarchical coding and fingerprint construction module, the depth comparison module, and the forgery detection and comprehensive judgment module are deployed on the cloud server to utilize more powerful computing resources to complete deep feature analysis and generative adversarial network inference. Intermediate data is transmitted between the edge device and the cloud through an encrypted HTTPS channel, and the end-to-end identification latency of a single card is controlled within 3 seconds.

[0053] To fully verify the effectiveness of the method and system proposed in this invention, this embodiment conducted a systematic experimental evaluation on a self-built high-end trading card dataset. The test dataset contains 5,000 genuine card samples and 3,000 counterfeit card samples, covering mainstream trading card brands and series on the market. The counterfeit samples cover multiple levels of counterfeiting, from low-quality to high-quality. All images were acquired under standard lighting conditions using a 1200 dpi macro imaging unit consistent with the system of this invention.

[0054] Experimental results show that the method of this invention achieves an accuracy rate of 99.2% in recognizing the anti-counterfeiting features of the floral pattern on the overall test set, with a 99.5% accuracy rate for genuine products (true positive rate) and a 98.7% accuracy rate for counterfeit products (true negative rate). In the test on a subset of highly counterfeit products (a total of 800 images, marked as highly counterfeit and difficult to distinguish by professional appraisers), the detection rate reached 98.5%, and the false alarm rate was only 0.8%, which is significantly better than the traditional appraisal method that relies solely on human experience.

[0055] In a comparative experiment with the anti-counterfeiting line connected component analysis method disclosed in CN119693961A, the method of this invention improved the detection rate of high-quality counterfeit products by 15.3 percentage points (98.5% vs. 83.2%) and the overall identification accuracy by 6.8 percentage points (99.2% vs. 92.4%). This significant improvement is mainly attributed to the following technical advantages: the skeleton extraction algorithm based on curvature flow analysis can more accurately capture the geometric topological differences of the floral patterns compared to a simple connected component counting method; the hierarchical coding system provides more comprehensive discrimination information than single-scale analysis through the synergistic fusion of macro-, meso-, and micro-level features; and the generative adversarial network counterfeit detection module identifies abnormal samples that deviate from the distribution of genuine products at the probability distribution level, compensating for the insufficient sensitivity of deterministic comparison methods when facing sophisticated high-quality counterfeits.

[0056] In terms of processing efficiency, the system of this invention takes 2.1 seconds to identify a single card on a server equipped with an NVIDIA RTX 4090 GPU. This includes 0.4 seconds for macro imaging and preprocessing, 0.6 seconds for skeleton extraction, 0.3 seconds for hierarchical encoding, 0.2 seconds for depth comparison, and 0.6 seconds for forgery detection and comprehensive judgment. In batch identification mode, through a pipelined parallel processing mechanism, the system's steady-state throughput reaches 25 cards per minute, fully meeting the efficiency requirements of large-scale identification scenarios. The design goal of less than 3 seconds for single-card identification has been fully validated. When deployed and tested on a portable device equipped only with an integrated graphics card, the single-card identification time increases to approximately 8.5 seconds due to the lack of dedicated GPU acceleration, but this is still far superior to the average time for manual identification (typically 2-5 minutes per card), demonstrating the practical value of this invention on different hardware platforms.

[0057] Furthermore, the effectiveness of the closed-loop feedback optimization mechanism was experimentally verified. In a test involving the continuous identification of 1000 cards, the system with closed-loop feedback enabled improved the identification accuracy of the last 500 cards by 0.3 percentage points compared to the first 500 cards, and the success rate of boundary sample verification increased from 92.1% to 96.7%, demonstrating the system's ability to continuously optimize identification parameters through feedback learning during continuous operation.

[0058] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for visually identifying the authenticity of cards based on floral anti-counterfeiting features, characterized in that, Includes the following steps: Step S1, Macro Imaging and Adaptive Preprocessing of the Floral Pattern Area: A high-resolution macro imaging unit is used to acquire images of the floral pattern anti-counterfeiting area on the surface of the card to be identified, obtaining the original floral pattern image; white balance correction, illumination uniformity compensation, and multi-scale sharpening enhancement processing are sequentially performed on the original floral pattern image to generate a standardized floral pattern image; Based on the geometric center localization results of the floral pattern, the standardized floral image is cropped for the region of interest and registered at the sub-pixel level to obtain the floral image to be analyzed. Step S2, Extraction of the skeleton of the floral pattern lines based on curvature flow analysis: Perform multi-scale Gaussian difference filtering on the floral pattern image to be analyzed to enhance the edge response of the lines and generate an edge-enhanced image; Perform a skeleton thinning process based on curvature flow evolution on the edge-enhanced image, extracting the skeleton of the floral lines by iteratively shrinking the line region to a single pixel width. The topological connectivity of the floral pattern skeleton was repaired and the branches were pruned and optimized to obtain the floral pattern skeleton structure data. Step S3, Hierarchical encoding of floral pattern features and construction of multidimensional fingerprint: Extract macro layout features based on floral pattern skeleton structure data. Macro layout features include global topological connection relationship of skeleton and symmetry distribution index. Structural features are extracted at the mesoscale, including histograms of line curvature distribution and gradient vectors of density gradients; line features are extracted at the microscale, including color transition gradient distribution and contour closure integrity measures; macro-layout features, structural features, and line features are fused and encoded according to hierarchical weights to generate a multi-dimensional feature fingerprint of the floral pattern. Step S4, based on the depth comparison of Fourier descriptors and KL divergence: retrieve the genuine floral reference feature fingerprint of the same model as the card to be identified from the genuine feature database; Calculate the Fourier descriptor similarity of the floral pattern lines between the multidimensional feature fingerprint and the genuine floral pattern reference feature fingerprint, and calculate the KL divergence value of the color space distribution between the two; generate a depth alignment score based on the Fourier descriptor similarity and the KL divergence value.

2. The method for visually identifying the authenticity of cards based on floral anti-counterfeiting features according to claim 1, characterized in that, In step S1, the imaging resolution of the high-resolution macro imaging unit is no less than 1200 dpi, and the field of view covers the entire range of the anti-counterfeiting area of ​​the floral pattern. The multi-scale sharpening enhancement process adopts Laplacian pyramid decomposition, with 3-5 decomposition layers, and the enhancement coefficient of each layer is adaptively adjusted within the range of 1.2-2.

5.

3. The method for visually identifying the authenticity of cards based on the anti-counterfeiting feature of floral patterns according to claim 1, characterized in that, In step S2, the scale parameters of the multi-scale Gaussian difference filter include three scale levels: σ1=0.5, σ2=1.0, and σ3=2.0; the iterative convergence threshold for the skeleton refinement processing based on curvature flow evolution is set to 0.01 pixels, and the maximum number of iterations does not exceed 500.

4. The method for visually identifying the authenticity of cards based on floral anti-counterfeiting features according to claim 1, characterized in that, In step S4, the preset value range of the authenticity determination threshold is 0.65-0.85, the weight coefficient of the depth comparison score is 0.6, and the weight coefficient of the forgery probability score is 0.

4.

5. The method for visually identifying the authenticity of cards based on floral anti-counterfeiting features according to claim 1, characterized in that, In step S2, the skeleton thinning process based on curvature flow evolution includes: calculating the local curvature value for each line pixel in the edge enhancement image, and determining the shrinkage velocity vector of the pixel based on the curvature value; gradually shrinking the line region to the central axis position along the direction of the shrinkage velocity vector, while maintaining the topological connectivity of the lines during the shrinkage process; stopping the iteration when the skeleton displacement between two adjacent iterations is less than the convergence threshold, and outputting a single-pixel width floral line skeleton.

6. The method for visually identifying the authenticity of cards based on floral anti-counterfeiting features according to claim 1, characterized in that, In step S3, the weights of the hierarchy are allocated as follows: the weight of macro layout features is 0.2-0.3, the weight of structural features is 0.3-0.4, and the weight of line features is 0.3-0.

5. The fusion encoding adopts a cascaded splicing and dimensionality reduction method, with a dimensionality reduction target of 128-512 dimensions.

7. The method for visually identifying the authenticity of cards based on floral anti-counterfeiting features according to claim 1, characterized in that, In step S4, the process of calculating the Fourier descriptor similarity of the floral pattern includes: resampling the contour curve of the floral pattern skeleton at equal intervals, with 256-1024 sampling points; performing a discrete Fourier transform on the resampled contour curve to extract the first 32-128 order Fourier descriptor coefficients; and calculating the normalized Euclidean distance between the Fourier descriptor coefficients of the card to be identified and the genuine reference as the Fourier descriptor similarity.

8. The method for visually identifying the authenticity of cards based on floral anti-counterfeiting features according to claim 1, characterized in that, It also includes step S5, forgery detection and comprehensive judgment based on generative adversarial network: the image of the floral pattern to be analyzed is input into the pre-trained generative adversarial network forgery detection model, and the generative adversarial network forgery detection model outputs a forgery probability score by learning the generation distribution of genuine floral patterns; The depth comparison score and the forgery probability score are weighted and fused to generate a comprehensive identification score. The authenticity of the card is determined by comparing the comprehensive identification score with a preset authenticity threshold. If the comprehensive identification score is lower than the preset authenticity threshold, the card is identified as a forgery. The generative adversarial network forgery detection model includes a generator and a discriminator. The generator adopts a U-Net structure, and the discriminator adopts a PatchGAN structure. The encoder of the generator contains 4-6 downsampled convolutional layers, each with a kernel size of 4×4 and a stride of 2. The receptive field size of the discriminator is 70×70 pixels.

9. The method for visually identifying the authenticity of cards based on floral anti-counterfeiting features according to claim 8, characterized in that, In step S5, the comprehensive identification score is fed back to step S1 to dynamically adjust the exposure parameters and focusing distance of the macro imaging unit in order to optimize the acquisition quality of the floral image in subsequent identification tasks. The Fourier descriptor bias information in the depth alignment score is fed back to step S2 to adjust the scale parameter of curvature flow evolution to improve the skeleton extraction accuracy.

10. A card authenticity visual identification system based on floral anti-counterfeiting features, used to implement the card authenticity visual identification method based on floral anti-counterfeiting features as described in claim 9, characterized in that, include: The floral pattern macro imaging and preprocessing module is used to acquire images of the floral pattern anti-counterfeiting area on the surface of the card to be identified using a high-resolution macro imaging unit to obtain the original floral pattern image. The module then performs white balance correction, illumination uniformity compensation, multi-scale sharpening enhancement, region of interest cropping, and sub-pixel level registration on the original floral pattern image to obtain the floral pattern image to be analyzed. The floral skeleton extraction module is used to perform multi-scale Gaussian difference filtering and skeleton refinement based on curvature flow evolution on the floral image to be analyzed, extract the floral line skeleton and perform topological connectivity repair and branch pruning optimization to obtain floral skeleton structure data. The hierarchical coding and fingerprint construction module is used to extract layout features, structural features and line features at macro, meso and micro scales based on the skeleton structure data of the floral pattern, and to fuse and encode the features at the three scales according to hierarchical weights to generate a multi-dimensional feature fingerprint of the floral pattern. The deep comparison module is used to retrieve the reference feature fingerprint of the same model of genuine floral pattern from the genuine feature database, calculate the Fourier descriptor similarity and color space distribution KL divergence value between the multidimensional feature fingerprint of floral pattern and the reference feature fingerprint of genuine floral pattern, and generate a deep comparison score. The forgery detection and comprehensive judgment module is used to input the floral image to be analyzed into a pre-trained generative adversarial network forgery detection model to obtain a forgery probability score, weight and fuse the depth comparison score and the forgery probability score to generate a comprehensive identification score, and output the card authenticity judgment conclusion based on the comparison result of the comprehensive identification score and the preset authenticity judgment threshold.

Images