A machine vision-based printed label anti-counterfeiting identification method and system

Abstract

The present application relates to the technical field of machine vision detection, and in particular to a printing label anti-counterfeiting identification method and system based on machine vision, which utilizes multispectral imaging to collect data to generate an initial spectral image set; adopts refraction correction modeling to peel off the interference signal of the transparent coating layer; through iterative filtering, residual noise is filtered out to generate a pure signal; a layer separation and spatial decoupling algorithm is used to analyze the multi-layer structure and extract a feature vector; the distribution of hidden anti-counterfeiting codes and fiber patterns is analyzed in combination with coding rules; a pattern matching and threshold determination logic is used to generate a label authenticity determination result; and a continuous identification optimization report is generated based on dynamic feedback control. The present application can effectively eliminate complex coating layer interference, realize accurate analysis of the deep anti-counterfeiting features of multi-layer labels, and significantly improve the robustness and accuracy of anti-counterfeiting identification in a high-speed production environment.

Description

Technical Field

[0001] This invention belongs to the field of machine vision inspection technology, specifically a machine vision-based method and system for anti-counterfeiting identification of printed labels. Background Technology

[0002] The core objective of the printed label anti-counterfeiting technology field is to accurately assess and analyze the authenticity of product labels by developing advanced identification methods and equipment, thereby enabling targeted protective measures to safeguard market order and corporate interests. Printed label anti-counterfeiting technology is part of brand protection and the security of commodity circulation. Its main task is to quantify and analyze the anti-counterfeiting features of labels, providing scientific support for authenticity determination for inspectors and decision-makers.

[0003] Machine vision-based anti-counterfeiting label recognition is a complex technology aimed at accurately identifying multi-layered composite labels. The concept of anti-counterfeiting recognition includes identifying internal structure, separating interference signals, and extracting hidden features to minimize the overall false positive rate. This allows companies to more effectively reduce the impact of counterfeit products on brand reputation, improve the market environment, and drive continuous iteration of anti-counterfeiting technology. Traditional methods have several shortcomings in practice. Previous methods often neglected the comprehensive analysis of the label's internal structure and multi-dimensional depth information, leading to inaccuracies in judging the authenticity of multi-layered or special material labels. Traditional methods lack the application of multispectral imaging and refractive correction technology in dealing with interference from transparent coatings or special effect layers, limiting the clear separation of anti-counterfeiting information at different depths. Furthermore, traditional methods often lack scientific layer separation models to support the accurate extraction of hidden anti-counterfeiting codes and special fiber patterns. In the interaction analysis between high-speed production line inspection and real-time data feedback, traditional methods fail to fully integrate dynamic comparison with genuine and counterfeit template databases, resulting in insufficient evaluation of label recognition efficiency in high-speed operating environments. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned shortcomings in the prior art by providing a machine vision-based method and system for anti-counterfeiting identification of printed labels.

[0005] The objective of this invention is achieved through the following technical solution: A machine vision-based anti-counterfeiting identification method for printed labels includes the following steps: S1. Based on multispectral imaging equipment, spectral data of multi-layer anti-counterfeiting labels are acquired. Multi-wavelength scanning and photoelectric conversion methods are used to perform comprehensive analysis of reflection and transmission signals, and preliminary processing of spectral response is performed to generate an initial spectral image dataset. S2. Based on the initial spectral image dataset, spectral analysis algorithms and refractive correction modeling are used to remove interference signals from the transparent coating and to identify depth spatial information, generating a spectral feature map after interference correction. S3. Based on the spectral feature map after interference correction, an iterative filtering model is used to filter out residual noise in a loop, and the intensity of light interference is analyzed to generate a pure set of depth layer signals. S4. Based on a clean set of deep layer signals, a layer separation algorithm and a spatial decoupling method are used to perform layer-by-layer analysis of the multi-layer structure, extract internal structural features, and generate feature vectors for each separated layer. S5. Combining the anti-counterfeiting coding rules and the feature vectors of each layer after separation, a feature extraction model is used to analyze the distribution of hidden anti-counterfeiting codes and special fiber patterns, and to evaluate the influence of image patterns, generating multi-dimensional feature extraction analysis results. S6. Based on the results of multidimensional feature extraction and analysis, a pattern matching algorithm is used to compare the preset real and fake template databases and evaluate the feature consistency to generate a label real and fake matching degree evaluation. S7. Based on the label authenticity matching degree evaluation, a threshold judgment logic and multi-criteria evaluation are used to confirm the authenticity of the product or mark it as counterfeit, and generate the label authenticity judgment result. S8. Based on the label authenticity determination results, dynamic feedback control and real-time integration with the production line are adopted to continuously optimize the identification sequence and analyze misjudgments, and generate a continuous identification optimization report.

[0006] A machine vision-based anti-counterfeiting identification system for printed labels includes an initial spectral image dataset generation module, an interference-corrected spectral feature map generation module, a clean deep layer signal set generation module, a separated layer feature vector generation module, a multi-dimensional feature extraction and analysis result generation module, a label authenticity matching degree evaluation generation module, a label authenticity determination result generation module, and a continuous identification optimization report generation module.

[0007] The beneficial effects of this invention are as follows: By applying multispectral imaging and multi-wavelength scanning, this invention achieves more accurate capture of label reflection and transmission signals, improving the completeness of initial spectral data. This method utilizes spectral analysis algorithms and refractive correction modeling to effectively remove interference from transparent coatings, facilitating the identification of deeply hidden anti-counterfeiting signals. Furthermore, the application of iterative filtering models ensures more thorough removal of residual noise, providing a high signal-to-noise ratio foundation for layer separation algorithms to analyze internal features. By combining blind source separation and spatial decoupling methods, this method can deeply analyze the physical composition of multi-layer structures and accurately extract hidden anti-counterfeiting codes and fiber patterns. Utilizing dynamic feedback control and real-time integration with the production line, it provides comprehensive technical support for the accurate identification of printed labels in high-speed operating environments, thereby promoting the optimization of anti-counterfeiting identification sequences and a continuous reduction in the false positive rate. Attached Figure Description

[0008] The invention will be further illustrated with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the invention. For those skilled in the art, other drawings can be obtained based on the following drawings without any creative effort.

[0009] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0010] The present invention will be further described in conjunction with the following embodiments.

[0011] Depend on Figure 1As can be seen, the anti-counterfeiting identification method and system for printed labels described in this embodiment includes the following steps: S1. Acquiring spectral data of multi-layer anti-counterfeiting labels using a multispectral imaging device, employing multi-wavelength scanning and photoelectric conversion methods to perform comprehensive analysis of reflection and transmission signals, and performing preliminary processing of the spectral response to generate an initial spectral image dataset; S2. Based on the initial spectral image dataset, using spectral analysis algorithms and refractive correction modeling, stripping interference signals from the transparent coating, and identifying depth spatial information to generate an interference-corrected spectral feature map; S3. Based on the interference-corrected spectral feature map, using an iterative filtering model to cyclically filter out residual noise, and analyzing the intensity of light interference to generate a clean set of depth layer signals; S4. Based on the clean set of depth layer signals, using a layer separation algorithm and spatial decoupling method to further... S5. Perform layer-by-layer analysis of the multi-layer structure and extract internal structural features to generate feature vectors for each separated layer; S6. Combine the anti-counterfeiting coding rules and the feature vectors of each separated layer, use a feature extraction model to analyze the distribution of hidden anti-counterfeiting codes and special fiber patterns, and evaluate the influence of image patterns to generate multi-dimensional feature extraction analysis results; S7. Based on the multi-dimensional feature extraction analysis results, use a pattern matching algorithm to compare with a preset genuine and counterfeit template database, and evaluate the consistency of features to generate a label authenticity matching degree evaluation; S8. Based on the label authenticity matching degree evaluation, use threshold judgment logic and multi-criteria evaluation to confirm genuine products or mark counterfeits, and generate label authenticity judgment results; S9. Based on the label authenticity judgment results, use dynamic feedback control and real-time integration of the production line to continuously optimize the recognition sequence and analyze misjudgments, and generate a continuous recognition optimization report.

[0012] Specifically, the anti-counterfeiting identification method and system for printed labels described in this embodiment includes an initial spectral image dataset comprising light intensity distribution, reflectance matrix, and transmittance spectral lines at different wavelengths; an interference-corrected spectral feature map comprising refractive index correction components, depth correlation signals, and corrected spectral envelopes; a pure depth layer signal set comprising denoised layered waveforms and signal-to-noise ratio enhancement sequences; separated layer feature vectors comprising spatial texture descriptors, color consistency vectors, and depth structure indices; multidimensional feature extraction and analysis results comprising hidden coding sequences, spatial distribution of special fibers, and microtext outlines; label authenticity matching evaluation comprising Euclidean distance score, cosine similarity index, and statistical confidence interval; label authenticity determination results comprising a genuine product verification report, counterfeit warning labels, and abnormal feature classifications; and a continuous identification optimization report comprising a misjudgment rate trend curve, production line operation feedback, and real-time identification sequence adjustment.

[0013] Specifically, in the initial spectral image acquisition step, a multispectral imaging device is used to scan the multi-layered labels across the entire spectrum. Narrowband filtering technology is employed, setting 12 key wavelength sampling points within the 400nm to 1000nm range. The captured light signals are converted into 16-bit digital signals by a CMOS image sensor. Next, non-uniformity correction is applied, assigning correction factors ranging from 0.85 to 1.15 based on the response sensitivity of each pixel in the sensor to eliminate artifacts caused by uneven light sources. The analysis results reveal the reflectance intensity and internal transmission patterns of the label surface, providing fundamental data support for subsequent deep-layer identification. This process ultimately generates a detailed dataset containing light intensity distribution maps and transmission spectral lines.

[0014] In the interference signal stripping step, a multilayer medium refractive model is used for optical path tracing based on the initial spectral dataset. For common label surface coatings such as transparent clear varnish or PET, a light deflection differential equation is established to identify the effects of refractive index... The system addresses virtual image displacement caused by differences in polarization. By calculating the extinction ratio of P-polarized and S-polarized light, polarization filtering is used to suppress specular reflection interference. During this process, the system performs depth spatial identification on coatings ranging from 50 to 200 micrometers in thickness, stripping away overlapping surface signals to extract the features of the masked bottom anti-counterfeiting layer. The generated interference-corrected spectrum significantly improves the spatial resolution of feature points, providing clear physical boundaries for multilayer structure analysis.

[0015] In the residual noise filtering step, an adaptive iterative filtering model is used to refine the spectral feature map. The system first calculates the variance distribution of local image regions and predicts frequency bands with low signal-to-noise ratio (SNR). Then, wavelet transform is used to decompose the signal into five scales, filtering out noise components that are random fluctuations. The iterative Wiener filtering algorithm cyclically updates the gain coefficients at this stage until the mean square error of the reconstructed signal drops below 0.01. Thresholding is then used to separate the weak anti-spoofing signal from the cluttered background interference. The generated clean depth layer signal set highlights the hidden signal peaks, ensuring the input quality for subsequent layer separation algorithms.

[0016] In the multi-layer structure analysis step, the Blind Source Separation (BSS) algorithm is used to decouple the mixed signals. The total detected light intensity is used as the observation variable, and the energy attenuation of light energy penetrating each dielectric layer is calculated by constructing an interface transmission matrix. The interface is used as the transformation matrix, and light energy as the flow direction to analyze the shortest propagation distance and scattering loss between each layer. Backscattering analysis is used to extract the physical parameter configuration of each layer (such as the substrate layer, ink layer, and coating layer), and the image is spatially decomposed based on spectral gradient differences. The generated feature vector captures the geometric moment features of the label's internal microstructure, effectively distinguishing the label's physical layers and adhesion state.

[0017] In the feature extraction and evaluation step, a deep analysis is performed on the parsed vectors of each layer, based on pre-defined anti-counterfeiting coding rules. Morphological gradient extraction technology is used to locate QR codes and random fiber information hidden in the middle or bottom layers. By simulating the scaling of structural elements on the 3D feature map, microtext with linewidths at the 10-micrometer level is detected. The evaluation model also quantifies the impact of layer separation quality on the reliability of code reading and analyzes the interference weight of background patterns on feature extraction. The generated comprehensive distribution map details the topological structure of the hidden anti-counterfeiting elements, providing a high-dimensional chain of evidence for subsequent authenticity determination.

[0018] In the authenticity matching evaluation step, a deep feature comparison method is used to dynamically compare the extracted results with a cloud-based standard library. A standard feature vector set is established, and fuzzy comprehensive evaluation logic is applied to handle reasonable tolerances generated during the printing process. The Euclidean distance and cosine similarity between the test vector and the template are calculated, and a kernel function is used to fit the probability distribution of multiple batches of genuine products. The system predicts the membership degree of the samples within the statistical space to assess whether they are within the normal production fluctuation range. The matching evaluation report generated in this step includes not only the score but also the confidence interval of the judgment conclusion, providing a quantitative basis for classification decisions.

[0019] In the decision-making process for determining authenticity, a multi-indicator voting mechanism is used for final confirmation. The logistic classification regression model normalizes the matching scores to obtain... arrive The probability distribution between them. The Support Vector Machine (SVM) algorithm searches for the optimal classification hyperplane at this stage, classifying samples as genuine, suspected counterfeit, or anomaly. The system balances multiple evaluation dimensions such as spectral bias rate, morphological similarity, and coding integrity, capturing the final conclusion through the majority principle. The feedback correction algorithm continuously calibrates boundary parameters based on the motion trajectory of the judgment history, ensuring the accuracy of judgments under different printing processes.

[0020] In the continuous recognition optimization step, closed-loop control is implemented based on the decision conclusions. A Long Short-Term Memory (LSTM) network is used to analyze the temporal correlation of the detection sequences on the production line, identifying recognition accuracy deviations caused by ambient lighting or conveyor belt speed fluctuations. A Bayesian optimization method automatically adjusts the exposure time of the imaging equipment (dynamically adjusted from 10ms to 20ms) and the sampling frequency to adapt to the high-speed operating environment. A genetic algorithm is used to optimize the cycle time of the detection production line, finding the optimal balance between recognition efficiency and accuracy at a production speed of 120 meters per minute. The final optimization report guides the system's continuous iteration, effectively reducing the false positive rate trend in large-scale detection processes.

[0021] The anti-counterfeiting identification method and system for printed labels described in this embodiment, based on a multispectral imaging device, acquires spectral data from multi-layer anti-counterfeiting labels, employs multi-wavelength scanning and photoelectric conversion methods, performs comprehensive analysis of reflected and transmitted signals, and performs preliminary processing of the spectral response to generate an initial spectral image dataset. The specific steps are as follows: S101: Based on the multispectral imaging device, using narrowband filtering technology, light signals are captured from multiple wavelength points of the anti-counterfeiting label within the range of 400nm to 1000nm. Then, the light energy is converted into digital signals by a photosensitive element, and non-uniformity correction is applied. Differential correction factors are assigned based on the sensitivity differences of each pixel in the sensor, and multi-channel data is integrated to generate a multispectral raw grayscale image; S102: Based on the multispectral raw grayscale image, a radiometric calibration algorithm is used to... By referencing standard whiteboard and blackboard data, the influence of ambient light and system bias are identified. Interpolation is used to supplement missing spectral frequency information, revealing the continuous variation trend of the spectral signal and generating a spectral reflectance distribution map. S103: Based on the spectral reflectance distribution map, an image enhancement algorithm is used to extract the edge and texture features of the labels through histogram equalization, and feature contrast enhancement is performed. Then, a smoothing filtering algorithm is used to initially suppress pixel noise according to a preset kernel function, generating a preliminary processed spectral image. S104: Based on the preliminary processed spectral image, a spectral dimension compression method is used. Principal component analysis is used to reduce the dimensionality of the spectral channels, retaining key spectral feature components. Data standardization is used to unify the dimensions, assess the data integrity, and generate an initial spectral image dataset.

[0022] In sub-step S101, data is synchronously acquired using a multispectral device. First, a narrowband filter wheel rapidly switches to illuminate the label surface with discrete wavelengths. The CMOS sensor converts the photon flow into a voltage signal, which is then output as 12-bit raw data via an ADC (Analog-to-Digital Converter). Non-uniformity correction involves calculating the dark current compensation and gain coefficient for each pixel to ensure a consistent grayscale response under uniform illumination. For example, edge pixels have a gain set to 1.12 due to lens distortion, while center pixels have a gain set to 1.02. The resulting raw grayscale image retains the original energy distribution across the spectral dimensions, providing a high-precision foundation for subsequent quantitative analysis. In sub-step S102, the core of generating the spectral reflectance distribution map lies in eliminating ambient light interference. The calibration process calculates the spectral radiance shift of the current environment by comparing the brightness values ​​of a white board (99% reflectance) and a black board (0.1% reflectance) in real time. Since the sensor can only acquire data at a limited number of wavelengths, spline interpolation is used to fit the missing frequency bands, generating a continuous spectral curve. This continuity reveals the excitation phenomenon of anti-counterfeiting ink at specific wavelengths, improving the sensitivity of the identification system to special materials. In sub-step S103, local contrast enhancement technology is used to highlight microscopic anti-counterfeiting features. Histogram equalization processing alters the gray-level frequency distribution of the image, making the special fiber textures hidden in low-brightness areas clearly visible. A smoothing filtering algorithm is employed... A Gaussian kernel function is used to initially smooth the thermal noise generated during the acquisition process. This step enhances the edge gradients of the spectral image, enabling subsequent layer separation algorithms to more accurately capture the boundaries between physical layers. In sub-step S104, Principal Component Analysis (PCA) is used to reduce the dimensionality of the 12 spectral channels. Due to the high correlation between data in adjacent bands, PCA can extract the top four principal components with a contribution rate exceeding 95%, significantly reducing the computational burden. Data standardization eliminates dimensional differences, making the acquired data from different exposure times comparable. The final generated initial spectral image dataset significantly improves the processing efficiency of the algorithm in high-speed detection environments while maintaining feature information. Assuming a counterfeit detection node, the data captured by the multispectral device includes light intensities in bands such as 450nm, 550nm, and 650nm. Correction factors are allocated based on laboratory calibration data, where edge pixels... In radiometric calibration, the measured brightness of the white board was 240, and that of the black board was 5, thus establishing a linear mapping relationship. PCA dimensionality reduction compressed the 12-dimensional data into a 3-dimensional feature vector, preserving the core fluorescence excitation phenomenon. The final analysis results showed that a hidden code exists at the bottom of the tag in the 650nm band, and the system predicted a recognition success rate of 98.5%. These analyses constitute a high-precision digital foundation for tag features.

[0023] The anti-counterfeiting identification method and system for printed labels described in this embodiment, based on the initial spectral image dataset, employs spectral analysis algorithms and refractive correction modeling to strip away interference signals from the transparent coating and identify depth spatial information, generating a spectral feature map after interference correction. The specific steps are as follows: S201, Based on the initial spectral image dataset, using a multilayer medium refractive model, geometrically modeling the propagation path of light at the transparent coating interface, analyzing the deflection characteristics of the light beam in media with different refractive indices, revealing the interference law caused by interface reflection, and generating a refractive interference path analysis; S202, Based on the refractive interference path analysis, using a physical optics transfer model, by establishing and analyzing a description of light intensity attenuation... Differential equations are used to simulate the energy distribution of light as it penetrates a multilayer structure, including absorption and scattering modes, to generate a light energy depth transfer analysis; S203, based on the light energy depth transfer analysis, a polarization filtering method is applied to calculate the extinction ratio of differentiated polarization states in multispectral data, analyze the rotational effect of the transparent layer on polarized light, identify surface specular reflection interference, and generate a polarization signal extraction analysis; S204, based on the polarization signal extraction analysis, interference cancellation regression is used, and through statistical modeling combined with a refraction physics model, the overlapping signal of the stripped transparent coating is quantified, the true reflection characteristics at the bottom of the multilayer structure are extracted, and a spectral feature map after interference correction is generated; the Fresnel reflection and refraction correction formula is used to analyze the rotational effect of the transparent layer on polarized light. ;in, These are the corrected reflectance observations. It is the refractive index of the transparent coating. It is the angle of incidence of light. It refers to the coating thickness. It is the wavelength of the incident light. It is the phase difference remainder term.

[0024] In substep S201, the multilayer dielectric refraction model is primarily used to analyze the complex reflection patterns of light within the multilayer composite label. The system sets the refractive index of the transparent gloss layer to be... The refractive index of the substrate layer is By tracing the optical path geometrically, the positional shift of the light spot due to two interface reflections is calculated. This process establishes a mapping table reflecting the refraction interference, enabling the algorithm to understand which aspects of the detected light intensity belong to surface noise and which belong to deep reflections, providing a geometric correction basis for subsequent signal reconstruction. In substep S202, the physical optics transfer model describes the flux of light energy in the vertical direction by decoupling the simplified form of Maxwell's equations. This is achieved by establishing an energy attenuation differential equation. Simulate the absorption coefficient of light in each layer of medium and scattering coefficient This not only reveals the amount of light loss when penetrating the surface coating, but also infers the thickness consistency of the label through the energy distribution pattern. The generated depth transmission analysis map provides energy-level parameter support for the quantitative removal of interference signals. In sub-step S203, the change in polarization state is used to distinguish between diffuse reflection and specular reflection. Specular reflection light usually maintains the incident polarization state, while diffuse reflection light after penetrating deep layers undergoes depolarization due to multiple collisions. The system obtains the spectral intensity at 0 degrees and 90 degrees by rotating the polarizer and calculates the polarizability. The polarization phase difference is corrected according to the Fresnel formula. This method effectively identifies strong surface glare and performs exceptionally well with highly glossy labels, preserving the underlying feature signals hidden by strong light. In sub-step S204, interference cancellation regression combines a physical model with statistical regression. Using the refraction path and energy attenuation calculated in the preceding steps as prior conditions, pixel-by-pixel subtraction is applied to the acquired composite spectral map. The algorithm performs convolution kernel cancellation on the original image based on the interference template generated by the physical model, stripping away the optical contribution of the surface coating. The resulting corrected feature map reveals the true texture of the bottom layer covered by the transparent layer, with a sharpness improvement of approximately 65% ​​compared to the original image, significantly enhancing the readability of the deep anti-counterfeiting code. Consider identifying a drug regulatory label coated with a glossy film. In S201, the light is incident at a 45-degree angle, and the refraction model predicts a displacement bias of 0.15 pixels. In S203, the measured specular extinction ratio is 12dB, and the phase difference correction is calculated using the Fresnel correction formula. In stage S204, regression cancellation successfully removed the white halo caused by the reflection of the glossy film, improving the clarity of the randomly distributed anti-counterfeiting fibers in the underlying layer. This series of operations ensures robust identification under complex packaging lighting conditions.

[0025] The anti-counterfeiting identification method and system for printed labels described in this embodiment, based on the interference-corrected spectral feature map, employs an iterative filtering model to cyclically filter out residual noise and analyzes the intensity of light interference to generate a clean deep-layer signal set. The specific steps are as follows: S301, Based on the interference-corrected spectral feature map, an adaptive filtering model is used to predict the distribution of noise power in different image regions by calculating the local variance in the spatial domain of the spectral signal, analyzing the spatial correlation between the signal and noise, and generating a preliminary denoised spectral signal; S302, Based on the preliminary denoised spectral signal, wavelet transform is used to further refine the signal... Multi-scale decomposition processing is performed to highlight useful high-frequency features and suppress random fluctuation noise, generating a multi-scale subdivided spectral stream; S303, based on the multi-scale subdivided spectral stream, the iterative Wiener filtering algorithm is used to analyze the statistical characteristics of residual noise and iteratively update the filter gain coefficient. Referring to the signal-to-noise ratio objective function, a refined analysis of the deep layer signal is generated; S304, based on the refined analysis of the deep layer signal, through threshold cutting processing, the data is divided into effective feature areas and background interference areas according to the signal intensity distribution characteristics, hidden signal peaks are identified, and the anti-spoofing feature layer is highlighted by quantifying the salience of the signal, generating a pure deep layer signal set.

[0026] In sub-step S301, the adaptive filtering model adjusts the filtering intensity based on the local detail complexity of the image. In textured anti-counterfeiting areas, the system reduces the smoothing intensity to preserve edge information; in flat background areas, it increases the smoothing coefficient. A noise estimation matrix is ​​dynamically established by calculating the local variance values ​​within the neighborhood of each pixel. This spatial correlation analysis method effectively suppresses particle noise generated by the photosensitive element while avoiding the blurring of anti-counterfeiting features caused by traditional fixed-parameter filtering. In sub-step S302, wavelet transform decomposes the image into approximate and detail components. Under the three-level decomposition structure, high-frequency coefficients at each scale are extracted; these coefficients typically correspond to microtext or intricate textures on the label. By performing nonlinear thresholding on the detail components, high-frequency components belonging to random noise are eliminated. The generated subdivided spectral stream exhibits extremely high contrast at different spatial scales, providing a clean signal source for the reconstruction of subsequent layers. In sub-step S303, Wiener filtering uses the minimum mean square error criterion to find the best estimate of the original image in the frequency domain. The system iterates multiple times, re-estimating the noise autocorrelation function based on the residuals of the previous output. In each iteration, the gain coefficient is adaptively attenuated or enhanced based on the current signal-to-noise ratio (SNR) until the reconstruction error meets a preset threshold. This deep iterative process allows for a significant improvement in SNR even for weak signals acquired in low-light environments. In sub-step S304, energy thresholding transforms the continuous signal stream into discrete feature regions. The system statistically analyzes the global light intensity distribution, establishes a dynamic threshold curve, and identifies points with an intensity exceeding three times the standard deviation of the environmental baseline as potential anti-counterfeiting elements. The quantitative evaluation results highlight the significance of the hidden anti-counterfeiting layer. The generated deep layer signal set not only contains image coordinates but also reflectivity peaks reflecting the material's physical properties, providing a data carrier for the independent component analysis of the multi-layer structure. Imagine a high-speed detection environment where the spectral feature map is affected by stripe noise generated by electromagnetic interference from a drive motor. In step S301, local variance analysis accurately pinpoints the period of noise occurrence. In S302, wavelet decomposition separates the stripe frequency from the anti-counterfeiting code texture. In stage S303, after five iterations of Wiener processing, the SNR is improved from 15dB to 38dB. Finally, threshold segmentation of S304 successfully extracted signals from special fluorescent particles located deep within the tag. These analyses collectively ensured the system's sensitivity in extreme industrial environments.

[0027] The anti-counterfeiting identification method and system for printed labels described in this embodiment, based on the pure deep-layer signal set, employs a layer separation algorithm and spatial decoupling method to perform layer-by-layer analysis of the multi-layer structure, extract internal structural features, and generate feature vectors for each separated layer. The specific steps are as follows: S401: Based on the pure deep-layer signal set, a blind source separation algorithm is used. The mixed spectral signal of the multi-layer composite structure is set as the observation variable, and the physical thickness and spectral contribution of each layer are used as constraints. An optimization problem is solved to analyze the independent components of each layer, reducing inter-layer interference and generating an independent layer spectral component design; S402: Based on the independent layer spectral component design, inverse scattering analysis is used, including constructing an inter-layer transmission model, where the interface is used as the transformation matrix, and the light... The flow direction can be used to calculate the shortest propagation attenuation between each layer. Structural features are extracted based on energy distribution to generate a layered physical parameter configuration. S403: Based on the layered physical parameter configuration, a layered structure decoupling model is applied. Texture features and boundary weights of each layer are updated based on spectral similarity. The multi-layered image is spatially split and adjusted, matching the physical layers within the labels to generate a spatially decoupled layer image. S404: Based on the spatially decoupled layer image, independent layer spectral component design, and layered physical parameter configuration, a comprehensive feature description strategy is formulated. Geometric moments and spectral gradients of each layer are extracted to capture microstructural changes and generate the separated feature vectors for each layer. The mixed spectral signal of the multi-layered composite structure is set as the observation variable, and the Boug-Lambert law transformation formula is used. ;in, The detected light intensity represents the depth. The energy value at that location, The absorption coefficient represents the energy attenuation per unit thickness of a material. It is the first Transmission coefficients at the layer interface.

[0028] In substep S401, the Blind Source Separation (BSS) algorithm aims to decouple the original spectral features of each printed layer from a single observed spectrum. The system treats the superimposed light intensity received by the sensor as a linear or nonlinear combination of contributions from different physical layers. By constraining the statistical independence of independent components (e.g., maximizing non-Gaussianity), the algorithm can extract the independent components of the blue background layer, the black information layer, and the hidden fluorescent layer. The Boug-Lambert law transformation formula is used here as a physical constraint to correct the exponential attenuation of the light intensity in deeper layers, ensuring that the resolved bottom components are not distorted due to energy loss. In substep S402, backscattering analysis quantitatively describes the collision probability of photons passing through ink particles by establishing an interlayer transfer matrix. The system uses the interface as the transformation matrix and estimates the layer thickness using the ratio of reflected energy to incident energy. and density The layered parameter configuration includes not only thickness information but also the scattering cross-section of each layer of material. This method of deriving the physical structure from the optical response enables the system to identify whether the label has been re-attached or tampered with, improving the ability to identify physical counterfeiting. In sub-step S403, the layered structure decoupling model decouples the multispectral image from three-dimensional space into multiple two-dimensional physical layer slices. The system assigns each pixel to a specific layer label based on the similarity of the spectral features of the pixels. Boundary weights are dynamically adjusted according to the intensity of the inter-layer spectral gradient. Through this spatial decoupling, the originally overlapping text and background textures are separated into different virtual layers, completely solving the problem of difficulty in reading anti-counterfeiting codes due to color masking. In sub-step S404, high-dimensional feature vectors are extracted for each decoupled layer image. Geometric moments (including central moments and invariant moments) are calculated to capture the shape stability of the anti-counterfeiting pattern. Spectral gradients are extracted to describe the diffusion characteristics of the ink and the edge sharpness of the anti-counterfeiting fibers. These features together constitute a vector space that can represent the multidimensional physical properties of the label. The final separated feature vector set exhibits extremely high discriminative power, enabling accurate identification even with slight wear on the label surface through the integrity of deep features. Consider analyzing a three-layer anti-counterfeiting label containing hidden random fibers. In S401, the Boug-Lambert formula predicts a 40% energy loss in the fiber layer due to its coverage by the paper substrate. In S402, the fiber layer depth is calculated to be 85 micrometers. S403 isolates the fiber layer as a separate black-and-white image. S404 extracts a 64-dimensional feature vector describing the fiber curvature and distribution density. This hierarchical analytical depth is unattainable by traditional visual methods, providing physical support for the scientific determination of label authenticity.

[0029] The anti-counterfeiting identification method and system for printed labels described in this embodiment, combining anti-counterfeiting coding rules and the separated feature vectors of each layer, employs a feature extraction model to analyze the distribution of hidden anti-counterfeiting codes and special fiber patterns, and evaluates the influence of image patterns to generate multi-dimensional feature extraction analysis results. The specific steps are as follows: S501, Based on the anti-counterfeiting coding rules and the separated feature vectors of each layer, morphological gradient extraction is used to locate the coding region. The code point positions are locked by simulating the translation and scaling of structural elements on the feature map, enabling the detection of hidden code blocks within multiple depths. The potential impact of layer separation quality on code reading is analyzed, and an initial anti-counterfeiting coding location map is generated; S502, Based on the initial anti-counterfeiting coding location map, combined with fiber morphology data, a refined skeleton model is used to analyze the topology of the anti-counterfeiting fibers. The structure is designed to simulate the spatial trend of fiber bundles by defining local rules for fiber connections, assess the direct impact of fiber distribution on recognition accuracy, and generate a fiber pattern morphology analysis map; S503, based on the fiber pattern morphology analysis map, data fusion technology is used to combine encoded information and fiber patterns, and by integrating and analyzing geometric features and grayscale statistical data, the interaction between hidden information and background texture is analyzed, and its comprehensive impact on recognition robustness is assessed, and a comprehensive distribution map of anti-counterfeiting features is generated; S504, based on the comprehensive distribution map of anti-counterfeiting features, a feature saliency evaluation model is applied to conduct a comprehensive quality assessment of the extracted anti-counterfeiting elements, and by quantitatively analyzing the feature clarity after interference suppression, a pattern recognition scheme is provided, a comprehensive evaluation of the effectiveness of anti-counterfeiting features is conducted, and multi-dimensional feature extraction analysis results are generated.

[0030] In substep S501, morphological gradient operators are used to capture drastic changes in the edges of the encoded region. For QR codes hidden in the microstructure, the system searches for blocks with specific frequency characteristics through dilation and erosion operations. The localization process not only identifies the coordinates in the XY plane but also determines the vertical distribution depth of the encoding through layer indexing. The evaluation stage quantifies the impact of optical defocus caused by interlayer adhesives on the sharpness of the encoded edges; if the edge contrast is below 0.3, an adaptive sharpening algorithm is triggered. In substep S502, the refined skeleton model simplifies the complex random fiber image into a connected graph with a single pixel width. The system defines the distribution rules of fiber intersections, endpoints, and lengths. By calculating the topological features of the fibers (such as Euler number and curvature), a unique description of the fiber pattern is established. This analysis is crucial for anti-counterfeiting fibers, which have a naturally random characteristic, as it effectively eliminates false feature points caused by printing noise, and the generated morphological analysis map provides a topological basis for subsequent probability matching. In substep S503, data fusion technology enables multiple verifications of the image, code, and texture. The parsed QR code text information is spatially aligned with the corresponding fiber spatial coordinates. The complexity of the background texture below the encoded area is analyzed to assess whether there is any illegal intrusion of the encoding by the background color. This comprehensive analysis considers the interrelationships between anti-counterfeiting elements; for example, a specific code must be located in the upper left of a specific fiber bundle. This fusion of relative positional relationships greatly increases the difficulty of counterfeiting. In the S504 sub-step, the feature saliency evaluation model scores the extraction results. The difference between the anti-counterfeiting signal and background fluctuations is measured by calculating the signal-to-noise ratio (SCR). If the saliency of a feature point is higher than 0.85, it is marked as a core matching point. This step ultimately generates a multi-dimensional analysis result containing hidden codes, fiber distribution, and text outlines. It not only provides a visual image but also includes a list of physical properties of the feature points, providing detailed data for achieving a genuine-counterfeit comparison with an extremely low false alarm rate.

[0031] The anti-counterfeiting identification method and system for printed labels described in this embodiment, based on the multi-dimensional feature extraction and analysis results, employs a pattern matching algorithm to compare with a preset genuine and counterfeit template database and evaluates feature consistency to generate a label authenticity matching degree evaluation. The specific steps are as follows: S601, Based on the multi-dimensional feature extraction and analysis results, a deep feature comparison method is used for analysis. By establishing a matching model of the feature vector to be tested, a standard template library, and their dynamic association, the influence of the differential tolerance range on the judgment result is simulated. Referring to Euclidean distance and Hamming distance factors, a dynamic feature matching score result is generated; S602, Based on the dynamic feature matching score result, a fuzzy comprehensive evaluation is used to determine the authenticity matching degree. S603. Based on the spectral consistency index and geometric structure matching parameters, evaluate and compare the similarity between the label to be tested and the standard part in multi-dimensional space, and generate an optimized evaluation scheme for label anti-counterfeiting consistency; S604. Based on the optimized evaluation scheme for label anti-counterfeiting consistency, use probability density estimation, simulate the data distribution of multiple batches of genuine products through kernel function fitting technology, predict the membership probability of the sample to be tested in the statistical space, evaluate the confidence of the judgment result, and generate a matching probability distribution prediction result; S605. Based on the matching probability distribution prediction result, use a decision tree hierarchical model to classify and quantify the matching degree, and generate a label authenticity matching degree evaluation by referring to historical comparison thresholds and weight allocation ratios.

[0032] In substep S601, deep feature comparison achieves quantitative alignment of multimodal features. The system establishes a multidimensional search tree, retrieving the closest template from a million-level standard library within milliseconds. Euclidean distance is used to measure the geometric similarity of spectral vectors, while Hamming distance is specifically used for fast verification of binary anti-counterfeiting codes. The dynamic association model considers reasonable deviations in the production process (such as 3% scaling due to printing pressure), thus introducing adaptive tolerance in the comparison to prevent false negatives of qualified products. In substep S602, the fuzzy comprehensive evaluation method resolves the contradiction between veto power and comprehensive trade-offs. The system defines three evaluation sets: spectral consistency, geometric integrity, and positional accuracy. The weight allocation ratio is set according to the security level of different labels, with the core feature having a weight as high as 0.6. Through the operation of the membership function, the position of the label to be tested in the multidimensional evaluation space is calculated, providing a quantitative probabilistic basis for the qualitative judgment of authenticity. In substep S603, the probability density estimation model can dynamically adapt to changes in production batches. Because slight color differences may exist between different batches of raw materials, the system uses kernel density estimation (KDE) to fit the distribution surface of the latest 10,000 genuine samples. A sample is considered safe if it falls within the 95% confidence interval. This statistical modeling method gives the identification system the ability to learn on its own, enabling it to distinguish between process deviations and counterfeiting. In sub-step S604, the decision tree grading model transforms continuous probability values ​​into graded conclusions. Based on the matching probability distribution, four levels are set: Genuine Grade A, Genuine Grade B (slight process deviation), Suspected Counterfeit, and Definitely Counterfeit. The node logic of the decision tree integrates historical comparison thresholds and real-time risk weights. The final matching evaluation report records the matching performance of each individual feature in detail, generating a clear interpretive path for the final judgment result.

[0033] The anti-counterfeiting identification method and system for printed labels described in this embodiment, which integrates the label authenticity matching degree evaluation and preliminary authenticity classification results, adopts threshold judgment logic and multi-criteria evaluation to confirm authenticity or mark counterfeit products, and generates label authenticity judgment results, specifically includes the following steps: S701, based on the label authenticity matching degree evaluation and preliminary authenticity classification results, logistic classification regression is used to normalize the recognition score by simulating the judgment boundary, loss function, and regularization term to obtain an optimized judgment score configuration; S702, based on the optimized judgment score configuration, a support vector machine algorithm is used. This algorithm finds the hyperplane that maximizes the margin and optimally divides the judgment threshold to improve classification. To improve accuracy and reduce the impact of mislabeling, an optimized threshold determination scheme is generated; S703, based on the optimized threshold determination scheme, a multi-index voting mechanism is adopted. Among multiple determination criteria, including spectral deviation rate, morphological similarity, and coding integrity, the final determination conclusion is captured by simulating the majority principle and weight correction, balancing the deviations of each index, and generating a comprehensive determination logic strategy; S704, based on the comprehensive determination logic strategy, a feedback correction algorithm is adopted. By simulating the motion trajectory of the determination history, the determination parameters of the representative classification scheme are iteratively updated, so that the determination model moves closer to the optimal decision position, the determination conclusion is continuously calibrated, the determination boundary is refined, and the label authenticity determination result is generated.

[0034] In the S701 substep, Logistic Regression maps the complex matching score to... The loss function optimization employs a regularization term to prevent overfitting on specific printing batches. The normalized score generated in this process reflects the original probability that the label possesses genuine characteristics. Substep S702 further introduces a Support Vector Machine (SVM) to find the optimal classification hyperplane in high-dimensional space. The system utilizes a nonlinear kernel function to process ambiguous samples on the decision margin, ensuring maximum classification margin and effectively reducing the false negative rate for high-quality counterfeit products. In substep S703, a multi-index voting mechanism demonstrates robust decision-making. When spectral analysis and geometric extraction contradict each other (e.g., localized wear on the label leading to impaired geometric features), the system makes a decision based on preset weight correction rules. For example, if the anti-counterfeiting code is completely correct and the spectral consistency is higher than 0.95, minor morphological scratches are ignored. This majority principle combined with weight preference logic enables the system to exhibit extremely high fault tolerance under complex real-world conditions. In substep S704, the feedback correction algorithm self-evolves based on real-time feedback from the detection pipeline. If the system detects a fixed offset in a certain dimension for multiple consecutive samples (possibly due to ambient light aging), it automatically adjusts the position of the decision hyperplane. This simulation of historical trajectories and parameter iteration ensures that the judgment model consistently operates at its optimal point. The final judgment not only provides a "true / false" conclusion but also details the rationale behind the judgment, offering solid evidence for companies' brand protection decisions.

[0035] The anti-counterfeiting identification method and system for printed labels described in this embodiment, based on the label authenticity determination result, employs dynamic feedback control and real-time integration of the production line to continuously optimize the identification sequence and analyze misjudgments, generating a continuous identification optimization report. The specific steps are as follows: S801, Based on the label authenticity determination result, a trend analysis method is used to investigate the fluctuations of misjudgment data during the identification process. Subsequently, a long short-term memory network model is used to analyze the temporal correlation of identification errors, thereby predicting identification performance and generating a current status analysis result for identification accuracy; S802, Based on the current status analysis result for identification accuracy, a sensitivity analysis model is used to analyze external variables such as production line operating speed and ambient light, and analyze... The correlation strength between the identification system and the judgment error is used to conduct an operational robustness assessment and generate an environmental impact assessment of the identification system; S803, based on the environmental impact assessment of the identification system, a Bayesian optimization method is used to analyze the robustness of the differentiated identification strategy in a high-speed environment through posterior probability analysis, select the optimal sampling frequency and exposure time, and generate an identification parameter optimization analysis; S804, based on the identification parameter optimization analysis, combined with closed-loop control theory and real-time scheduling model, the behavior of the detection process is simulated through dynamic system theory, the pipeline processing cycle is optimized using a genetic algorithm, and the optimal balance point between identification efficiency and accuracy is captured using linear programming to formulate a detection scheme that matches the high-speed production environment and generate a continuous identification optimization report.

[0036] In substep S801, trend analysis focuses on identifying system performance drift. A Long Short-Term Memory (LSTM) network remembers detection trends from hours ago, detecting potential problems such as a gradual decrease in contrast over time, and providing early warnings of sensor overheating or lens dust accumulation. The generated accuracy analysis results provide data support for preventative maintenance. Substep S802's sensitivity model quantifies the weight of external environmental interference, such as the change in false positive rate sensitivity as the operating speed increases from 80 m / min to 150 m / min. In substep S803, Bayesian optimization searches for the optimal solution in a vast parameter space. It doesn't need to exhaustively list all possible exposure and gain combinations, but rather uses probabilistic reasoning to quickly pinpoint the settings that best balance noise and dynamic range under the current lighting conditions. Substep S804 then sends these optimized parameters to the execution unit in real time. A genetic algorithm optimizes the allocation of computational resources on the processing server, ensuring stable detection rate through a scheduling model when detection tasks surge. This closed-loop control transforms the anti-counterfeiting detection system from a static tool into a dynamic, self-optimizing intelligent node. The generated optimization report summarizes the path to improve the system's operating efficiency and accuracy, enabling long-term, efficient operation of printing anti-counterfeiting detection.

[0037] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A machine vision-based anti-counterfeiting identification method for printed labels, characterized in that: Includes the following steps: S1. Based on multispectral imaging equipment, spectral data of multi-layer anti-counterfeiting labels are acquired. Multi-wavelength scanning and photoelectric conversion methods are used to perform comprehensive analysis of reflection and transmission signals, and preliminary processing of spectral response is performed to generate an initial spectral image dataset. S2. Based on the initial spectral image dataset, spectral analysis algorithms and refractive correction modeling are used to remove interference signals from the transparent coating and to identify depth spatial information, generating a spectral feature map after interference correction. S3. Based on the spectral feature map after interference correction, an iterative filtering model is used to filter out residual noise in a loop, and the intensity of light interference is analyzed to generate a pure set of depth layer signals. S4. Based on a clean set of deep layer signals, a layer separation algorithm and a spatial decoupling method are used to perform layer-by-layer analysis of the multi-layer structure, extract internal structural features, and generate feature vectors for each separated layer. S5. Combining the anti-counterfeiting coding rules and the feature vectors of each layer after separation, a feature extraction model is used to analyze the distribution of hidden anti-counterfeiting codes and special fiber patterns, and to evaluate the influence of image patterns, generating multi-dimensional feature extraction analysis results. S6. Based on the results of multidimensional feature extraction and analysis, a pattern matching algorithm is used to compare the preset real and fake template databases and evaluate the feature consistency to generate a label real and fake matching degree evaluation. S7. Based on the label authenticity matching degree evaluation, a threshold judgment logic and multi-criteria evaluation are used to confirm the authenticity of the product or mark it as counterfeit, and generate the label authenticity judgment result. S8. Based on the label authenticity determination results, dynamic feedback control and real-time integration with the production line are adopted to continuously optimize the identification sequence and analyze misjudgments, and generate a continuous identification optimization report.

2. The method according to claim 1, wherein the method is characterized in that: The process involves acquiring spectral data from multi-layer anti-counterfeiting labels using a multi-spectral imaging device, employing multi-wavelength scanning and photoelectric conversion methods, comprehensively analyzing reflection and transmission signals, and performing preliminary processing of the spectral response to generate an initial spectral image dataset. S101. Based on a multispectral imaging device, a narrowband filtering technology is used to capture light signals from multiple wavelength points of the anti-counterfeiting label within a preset range. Then, the light energy is converted into digital signals through a photosensitive element. Non-uniformity correction is applied, and differentiated correction factors are allocated according to the sensitivity differences of each pixel of the sensor. Multi-channel data is integrated to generate a multispectral original grayscale image. S102. Based on the original multispectral grayscale image, a radiometric calibration algorithm is used to identify the influence of ambient light and system bias by referring to standard whiteboard and blackboard data. The missing spectral frequency band information is supplemented by interpolation, revealing the continuous change trend of the spectral signal and generating a spectral reflectance distribution map. S103. Based on the spectral reflectance distribution map, an image enhancement algorithm is used to extract the edge and texture features of the label through histogram equalization and perform feature contrast enhancement. Then, a smoothing filtering algorithm is used to suppress pixel noise for the first time according to a preset kernel function to generate a preliminary processed spectral image. S104. Based on the pre-processed spectral image, the spectral dimension compression method is used to reduce the dimensionality of the spectral channels through principal component analysis, retaining key spectral feature components. The data standardization method is used to unify the dimensions, the integrity of the data is evaluated, and an initial spectral image dataset is generated.

3. The method according to claim 1, characterized in that: Based on the initial spectral image dataset, spectral analysis algorithms and refractive correction modeling are used to remove interference signals from the transparent coating and identify depth spatial information to generate an interference-corrected spectral feature map. Specifically: S201. Based on the initial spectral image dataset, a multilayer medium refraction model is adopted. By geometrically modeling the propagation path of light at the interface of the transparent coating, the deflection characteristics of the light beam in the medium with different refractive indices are analyzed, the interference law caused by interface reflection is revealed, and a refraction interference path analysis is generated. S202. Based on the aforementioned refraction interference path analysis, using the physical optical transmission model, by establishing and analyzing the differential equation describing the light intensity attenuation, the energy distribution of light when penetrating a multi-layer structure is simulated, and a light energy depth transmission analysis is generated. S203. Based on the aforementioned light energy depth transfer analysis, a polarization filtering method is applied. By calculating the extinction ratio of the differential polarization states in the multispectral data, the rotational effect of the transparent layer on the polarized light is analyzed, surface specular reflection interference is identified, and polarization signal extraction and analysis are generated. S204. Based on the polarization signal extraction and analysis, interference cancellation regression is adopted. Through statistical modeling and combined with the refraction physical model, the overlapping signal of the stripped transparent coating is quantified, the true reflection characteristics of the bottom of the multilayer structure are extracted, and the interference-corrected spectral feature map is generated.

4. The anti-counterfeiting identification method for printed labels based on machine vision according to claim 1, characterized in that: The method involves using an iterative filtering model to cyclically filter out residual noise based on the interference-corrected spectral feature map, and analyzing the intensity of light interference to generate a clean set of depth layer signals. S301. Based on the spectral feature map after interference correction, an adaptive filtering model is adopted. By calculating the local variance in the spatial domain of the spectral signal, the distribution of noise power in different image regions is predicted, the spatial correlation between signal and noise is analyzed, and a preliminary denoised spectral signal is generated. S302. Based on the preliminary denoised spectral signal, the wavelet transform method is used to perform multi-scale decomposition processing on the signal, highlighting useful high-frequency features and suppressing random fluctuation noise, and generating a multi-scale subdivided spectral stream. S303. Based on the multi-scale subdivided spectral flow, the iterative Wiener filtering algorithm is used to analyze the statistical characteristics of residual noise, and the filter gain coefficient is updated iteratively. With reference to the signal-to-noise ratio objective function, a refined analysis of the deep layer signal is generated. S304. Based on the refined analysis of the deep layer signal, the data is divided into effective feature areas and background interference areas according to the signal intensity distribution characteristics through threshold cutting processing, hidden signal peaks are identified, and the anti-counterfeiting feature layer is highlighted by quantifying the salience of the signal, thereby generating a pure deep layer signal set.

5. The method according to claim 1, characterized in that: The pure deep layer signal set is used to perform layer-by-layer analysis of the multi-layer structure using a layer separation algorithm and spatial decoupling method, and to extract internal structural features to generate the feature vectors of each separated layer as follows: S401. Based on the pure deep layer signal set, a blind source separation algorithm is adopted. The mixed spectral signal of the multi-layer composite structure is set as the observation variable. The physical thickness and spectral contribution of each layer are used as constraints. The independent components of each layer are analyzed by solving the optimization problem, reducing inter-layer interference and generating independent layer spectral component design. S402. Based on the design of the independent layer spectral composition, back scattering analysis is used, including constructing an interlayer transport model, where the interface is used as the transformation matrix and light energy is used as the flow direction, calculating the shortest propagation attenuation between each layer, extracting structural features according to the energy distribution, and generating layered physical parameter configurations. S403. Based on the layered physical parameter configuration, apply the layered structure decoupling model, update the texture features and boundary weights of each layer according to spectral similarity, perform spatial dimension splitting and adjustment on the multi-layer image, match the physical layering inside the label, and generate a spatially decoupled layer image. S404. Based on the spatial decoupling layer image, the independent layer spectral composition design and the layered physical parameter configuration, formulate a comprehensive feature description strategy, extract the geometric moments and spectral gradients of each layer image, capture microstructural changes, and generate the separated feature vectors of each layer.

6. The method according to claim 1, wherein the method is characterized by: The method combines anti-counterfeiting coding rules and the separated feature vectors of each layer, employs a feature extraction model, analyzes the distribution of hidden anti-counterfeiting codes and special fiber patterns, and evaluates the influence of image patterns to generate multi-dimensional feature extraction analysis results. S501. Based on the anti-counterfeiting coding rules and the separated feature vectors of each layer, morphological gradient extraction is used to locate the coding region. The code point position is locked by simulating the translation and scaling of the structural elements on the feature map, thereby realizing the detection of hidden code blocks in multiple layers of depth. The potential impact of layer separation quality on code reading is analyzed, and an initial anti-counterfeiting coding location map is generated. S502. Based on the initial positioning map of the anti-counterfeiting code, combined with fiber morphology data, a refined skeleton model is used to analyze the topological structure of the anti-counterfeiting fibers. By defining local rules for fiber connections, the spatial trend of fiber bundles is simulated, the direct impact of fiber distribution on recognition accuracy is evaluated, and a fiber pattern morphology analysis map is generated. S503. Based on the fiber pattern morphology analysis diagram, multi-feature fusion technology is used to combine the encoded information and fiber pattern. By integrating and analyzing geometric features and grayscale statistics, the interaction between hidden information and background texture is analyzed, and its comprehensive impact on recognition robustness is evaluated, and a comprehensive distribution map of anti-counterfeiting features is generated. S504. Based on the comprehensive distribution map of anti-counterfeiting features, a feature saliency evaluation model is applied to conduct a comprehensive quality assessment of the extracted anti-counterfeiting elements. By quantitatively analyzing the feature clarity after interference suppression, a pattern recognition scheme is provided to conduct a comprehensive evaluation of the effectiveness of anti-counterfeiting features and generate multi-dimensional feature extraction and analysis results.

7. The method according to claim 1, characterized in that: The process of generating a label authenticity matching degree evaluation by comparing the results of multi-dimensional feature extraction and analysis with a pattern matching algorithm, comparing the results with a pre-set database of true and false templates, and evaluating feature consistency is as follows: S601. Based on the multidimensional feature extraction and analysis results, a deep feature comparison method is used for analysis. By establishing a matching model of the feature vector to be tested, a standard template library and its dynamic association, the influence of the differential tolerance range on the judgment result is simulated. With reference to Euclidean distance and Hamming distance factors, a dynamic feature matching score result is generated. S602. Based on the feature matching dynamic scoring results, fuzzy comprehensive evaluation is adopted, and the similarity between the label to be tested and the standard part in multi-dimensional space is evaluated and compared by combining the spectral consistency index and the geometric structure matching parameters, and a label anti-counterfeiting consistency optimization evaluation scheme is generated. S603. Based on the aforementioned label anti-counterfeiting consistency optimization evaluation scheme, probability density estimation is adopted, and the data distribution of multiple batches of genuine products is simulated through kernel function fitting technology to predict the membership probability of the test sample in the statistical space, evaluate the confidence of the judgment result, and generate the matching probability distribution prediction result. S604. Based on the predicted results of the matching probability distribution, a decision tree hierarchical model is adopted to classify and quantify the matching degree. Referring to the historical comparison threshold and weight allocation ratio, a label authenticity matching degree evaluation is generated.

8. The method according to claim 1, characterized in that: The label authenticity matching evaluation, which employs threshold judgment logic and multi-criteria assessment, confirms authenticity or marks counterfeit products, and generates label authenticity judgment results as follows: S701. Based on the label authenticity matching degree evaluation and preliminary authenticity classification results, logistic classification regression is adopted. By simulating the judgment boundary, loss function and regularization term, the identification score is normalized to obtain the optimized judgment score configuration. S702. Based on the optimized judgment score configuration, the support vector machine algorithm is adopted. This algorithm finds the hyperplane that maximizes the margin and performs optimal division of the judgment threshold to improve the classification accuracy and reduce the impact of mislabeling, thereby generating an optimized threshold judgment scheme. S703. Based on the optimized threshold determination scheme, a multi-index voting mechanism is adopted. Among multiple determination criteria, including spectral deviation rate, morphological similarity and coding integrity, the final determination conclusion is captured by simulating the majority principle and weight correction, balancing the deviation of each index, and generating a comprehensive determination logic strategy. S704. Based on the comprehensive judgment logic strategy, a feedback correction algorithm is adopted. By simulating the motion trajectory of the judgment history, the judgment parameters of the representative classification scheme are iteratively updated, so that the judgment model moves closer to the optimal decision position, the judgment conclusion is continuously calibrated, the judgment boundary is refined, and the label authenticity judgment result is generated.

9. The method according to claim 1, characterized in that: The process of continuously optimizing the identification sequence and analyzing misjudgments based on the label authenticity determination results, using dynamic feedback control and real-time integration with the production line, and generating a continuous identification optimization report, specifically involves: S801. Based on the label authenticity determination result, a trend analysis method is used to explore the fluctuation of misjudged data in the identification process. Then, a long short-term memory network model is used to analyze the temporal correlation of identification errors, thereby predicting the identification performance and generating the current status analysis result of identification accuracy. S802. Based on the analysis results of the current recognition accuracy, a sensitivity analysis model is used to analyze the external variables of multiple aspects such as production line operating speed and ambient light, and to analyze the correlation strength between them and the judgment error. In this way, an operational robustness assessment is conducted, and an environmental impact assessment of the recognition system is generated. S803. Based on the environmental impact assessment of the identification system, the Bayesian optimization method is used to analyze the robustness of the differentiated identification strategy in a high-speed environment through posterior probability analysis, select the optimal sampling frequency and exposure time, and generate identification parameter optimization analysis. S804. Based on the optimization analysis of the identification parameters, combined with closed-loop control theory and real-time scheduling model, the behavior of the detection process is simulated through dynamic system theory, the pipeline processing cycle is optimized using genetic algorithm, and the optimal balance point between identification efficiency and accuracy is captured by linear programming. A detection scheme matching the high-speed production environment is formulated, and a continuous identification optimization report is generated.

10. A machine vision based anti-counterfeiting identification system for printed labels, characterized in that: It includes modules for generating initial spectral image datasets, generating spectral feature maps after interference correction, generating clean deep layer signal sets, generating feature vectors for each separated layer, generating multidimensional feature extraction and analysis results, generating label authenticity matching degree evaluation, generating label authenticity determination results, and generating continuous recognition optimization reports.

Images