An image uniqueness anti-counterfeiting identification method based on multi-modal feature combination

Abstract

The present application relates to a kind of based on multimodal feature combination image uniqueness anti-counterfeiting identification method, belong to image processing and anti-counterfeiting identification technical field.It includes: constructing anti-counterfeiting carrier, include by random microstructure Random texture area And the readable identification area of associated digital identity information;Registration stage, based on texture feature extraction model and morphological feature extraction model respectively, the deep texture feature of random texture area and morphological structure feature are extracted, the unique feature fingerprint is generated by the multi-modal fusion of the two, and is bound and stored with digital identity information;Verification stage, through terminal equipment acquisition image to be measured, locate random texture area and extract measured texture feature, based on metric learning algorithm, measured texture feature and the unique feature fingerprint in database are mapped to the same feature space for comparison, and output true and false determination.The present application realizes the anti-counterfeiting identification of high security and high identification accuracy by multimodal feature combination and asymmetric registration verification architecture.

Description

Technical Field

[0001] This invention belongs to the field of image processing and anti-counterfeiting recognition technology, specifically relating to an image uniqueness anti-counterfeiting recognition method based on multimodal feature combination. Background Technology

[0002] With the rapid development of the high-end goods market, anti-counterfeiting technology has become an important means of protecting brand value and consumer rights. Currently, mainstream anti-counterfeiting technologies mainly rely on anti-counterfeiting marks on labels and packaging materials, commonly including laser anti-counterfeiting labels, fluorescent ink, holographic patterns, and QR codes. While these technologies increase the difficulty of counterfeiting to some extent, they are essentially "deterministic" manufacturing, meaning that products in the same batch have identical visual characteristics or follow fixed coding rules. With the continuous advancement of digital printing, high-precision scanning, and image reproduction technologies, criminals can easily create counterfeit labels with almost indistinguishable appearances through cloning or high-fidelity replication, posing a serious challenge to traditional anti-counterfeiting methods.

[0003] In recent years, anti-counterfeiting technologies based on Physically Unclonable Functions (PUFs) have received widespread attention. Many materials in nature, such as paper, fibers, and polymer films, exhibit natural random texture characteristics during manufacturing due to the uncontrollability of processes and the inherent properties of the materials themselves. Examples include the random distribution of fibers, the diffusion of ink, and the disordered arrangement of micron-sized particles. This randomness stems from the inherent randomness of physical processes and, theoretically, cannot be precisely replicated through artificial intervention, thus possessing a natural anti-counterfeiting advantage.

[0004] However, existing anti-counterfeiting technologies based on random textures still have significant shortcomings in application. Firstly, some technologies rely on high-precision microscopes or professional industrial cameras for microscopic structure acquisition and comparison, which are expensive and complex to operate, making them difficult to popularize among ordinary consumers. Secondly, existing technologies utilize random textures in a relatively simplistic way, either extracting only single-dimensional texture features for comparison or relying solely on deep learning models to extract global features, lacking comprehensive utilization of multi-dimensional features and making them vulnerable to localized counterfeiting by specific methods. Thirdly, how to guide ordinary users to achieve fast and reliable authentication through everyday portable devices (such as smartphones), while ensuring recognition accuracy and resistance to attacks, remains a pressing technical challenge for the industry.

[0005] Therefore, there is an urgent need for an anti-counterfeiting identification method that can balance physical non-cloning security, multimodal feature fusion recognition accuracy, and user operation convenience to overcome the shortcomings of existing technologies. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides an image uniqueness anti-counterfeiting recognition method based on multimodal feature combination. The objective of this invention can be achieved through the following technical solution: A method for image uniqueness anti-counterfeiting based on multimodal feature combination includes: An anti-counterfeiting carrier is constructed, which includes at least a random texture area and a readable identifier area. The random texture area is composed of random microstructures, and the readable identifier area is used to associate the digital identity information of the anti-counterfeiting carrier. During the registration phase, a reference image of the anti-counterfeiting carrier is collected. The deep texture features of the random texture area are extracted based on the texture feature extraction model, and the morphological structure features of the random texture area are extracted based on the morphological feature extraction model. The deep texture features and morphological structure features are fused in a multimodal manner to generate a unique feature fingerprint. The unique feature fingerprint is then bound and stored with the digital identity information corresponding to the readable identifier area. During the verification phase, the test image of the carrier under test is acquired through the terminal device, the random texture area is located according to the readable identifier area, and the test texture features of the random texture area under test are extracted based on the texture feature extraction model. Based on the metric learning algorithm, the texture features to be tested are mapped to the same feature space as the unique feature fingerprints stored in the database, and then the similarity is calculated and compared. Based on the comparison results, the authenticity determination information is output.

[0007] Specifically, the random texture area is composed of micro-serrations formed by the natural diffusion of printed ink edges, or of fibers randomly distributed in the substrate, or of micron- or nano-sized particles randomly distributed in the coating. The morphology of the micro-serrations includes the number of serrations, the serration amplitude, the serration spacing, and the burr length distribution; the distribution characteristics of the fibers include the fiber orientation angle, fiber density, the number of fiber intersections, and the fiber curvature; the distribution characteristics of the particles include the centroid position of the particles, the particle spacing, the particle size distribution, and the degree of particle aggregation.

[0008] Specifically, the readable identifier area is an identifiable identifier. The readable identifier area and the random texture area are arranged adjacent to or nested in space to form a one-to-one hard association relationship, which is used to associate and index the unique feature fingerprint of the anti-counterfeiting carrier through the readable identifier area.

[0009] As a preferred technical solution of the present invention, the texture feature extraction model is a deep learning model based on a convolutional neural network. The convolutional neural network includes multiple convolutional layers, pooling layers and fully connected layers, which are used to extract the texture gradient, fiber orientation and local structural difference features of the random texture region at the microscale. The texture feature extraction model also includes a Transformer branch, which is set in parallel or serially with the convolutional neural network. The Transformer branch is used to extract the global spatial relationship features and long-range dependency features of the random texture region, and to fuse the local features extracted by the convolutional neural network with the global features extracted by the Transformer branch.

[0010] Specifically, the morphological feature extraction model is built based on traditional image processing algorithms and includes an edge detection module, a morphological analysis module, and a geometric quantization module; The edge detection module uses the Canny operator or the Sobel operator to extract the ink edge; the morphological analysis module calculates the number of jagged edges, the distribution of burr length, and the concavity and convexity of the edge; the geometric quantization module calculates the convex hull area, Fourier descriptor, and shape context features of the edge, and quantizes the above features to form a structured morphological feature vector.

[0011] Specifically, the deep texture features and morphological structure features are fused in a multimodal manner, which includes: normalizing the deep texture features and morphological structure features respectively, generating a fused feature vector by weighted fusion, splicing fusion or attention-based fusion, and using the fused feature vector as the unique feature fingerprint; the weight coefficients of the weighted fusion are adaptively adjusted or pre-calibrated according to the importance of the feature dimension.

[0012] Specifically, after acquiring the reference image of the anti-counterfeiting carrier, the method further includes a preprocessing step for the reference image: the preprocessing includes illumination consistency correction, geometric distortion correction, image denoising enhancement, and automatic cropping of the anti-counterfeiting area based on the spatial position of the readable identifier area; the geometric distortion correction includes perspective transformation correction and affine transformation correction.

[0013] Specifically, before binding and storing the unique feature fingerprint with the digital identity information corresponding to the readable identifier area, the method further includes encrypting the unique feature fingerprint: the encryption process includes, but is not limited to, generating a digital digest using a hash operation, generating encrypted ciphertext using a symmetric encryption algorithm, or generating a digital signature using an asymmetric encryption algorithm, and storing the encrypted feature fingerprint and the digital identity information in a database.

[0014] Specifically, the step of acquiring the image of the carrier under test through the terminal device includes: acquiring the image of the carrier under test containing the readable identifier area and the random texture area through the camera of the terminal device; and locating the random texture area through an image registration algorithm based on the spatial position relationship of the readable identifier area, or automatically identifying and cropping the image area of ​​the random texture area through a target detection model.

[0015] Specifically, the calculation of similarity for comparison includes: using cosine similarity, Euclidean distance or Mahalanobis distance to calculate the similarity between the texture feature to be tested and the unique fingerprint feature in the feature space; The metric learning algorithm is trained using a triplet loss function or a contrastive loss function; it outputs authenticity determination information based on a preset similarity threshold. When the similarity is greater than or equal to the threshold, it is determined to be genuine; otherwise, it is determined to be suspected counterfeit.

[0016] Specifically, when outputting authenticity determination information based on the comparison results, a comprehensive judgment is also made in conjunction with a counterfeit attack detection mechanism: the counterfeit attack detection mechanism includes focus accuracy detection, moiré pattern detection, and three-dimensional micromorphological analysis. The focus accuracy detection is used to determine whether the image under test meets the preset clarity requirements; the moiré pattern detection is used to identify periodic textures generated by flipping the screen; and the three-dimensional micromorphological analysis is used to determine the three-dimensional physical structure of random texture areas through depth information. When signs of spoofing are detected, the similarity score is lowered or a fake message is displayed directly.

[0017] Specifically, the database is a cloud-distributed database or a local server database; the comparison process between the texture feature to be tested and the unique feature fingerprint is completed on the cloud server, or the unique feature fingerprint is cached locally on the terminal device for offline comparison; when offline comparison is used, the terminal device periodically updates the feature fingerprint data in the database.

[0018] The beneficial effects of this invention are as follows: Physically unclonable and highly secure: This invention utilizes random physical characteristics such as ink serrations, fiber distribution, or particle arrangement to construct an anti-counterfeiting carrier. This random texture originates from the inherent randomness of the manufacturing process and cannot be precisely replicated through human intervention. Counterfeits will inevitably exhibit differences in precision, location, and size at the microscopic level, fundamentally eliminating the possibility of mass cloning.

[0019] Multimodal feature fusion for accurate recognition: This invention breaks through the limitations of single feature extraction. During the registration stage, it extracts local micro-texture features through a deep learning network and simultaneously extracts morphological and structural features such as edge jaggedness, burr distribution, and convex hull area through traditional image processing algorithms. The two types of features complement each other and are fused to generate a unique feature fingerprint, which effectively solves the problems of insignificant features in weak texture scenes and the ease with which single features can be targeted for counterfeiting, and significantly improves the recognition accuracy.

[0020] Asymmetric architecture balancing accuracy and efficiency: This invention employs an asymmetric architecture of "registration fusion and simplified verification." During the registration phase, dual models operate in parallel, and multimodal deep fusion ensures the uniqueness and stability of the feature fingerprint. In the verification phase, only single-modal texture features are extracted, and comparison is completed through feature space mapping. This architecture significantly reduces the computational burden on the terminal while maintaining registration accuracy, supporting real-time recognition on mobile devices such as smartphones.

[0021] Multiple anti-attack mechanisms and strong defense capabilities: This invention integrates multiple anti-attack mechanisms, including focus accuracy detection, moiré pattern detection, and three-dimensional microscopic morphology analysis, effectively distinguishing genuine physical textures from flat printed counterfeits. When signs of counterfeiting are detected, the similarity score is automatically reduced or a counterfeit warning is directly output, significantly enhancing the defense against various counterfeiting methods.

[0022] Easy to operate and widely applicable: Users only need to scan the readable identification area with their smartphones, and the system automatically guides them to locate the random texture area and completes the recognition, making the operation simple. This invention can be widely used for anti-counterfeiting and traceability in high-end goods, food and pharmaceuticals, electronic products, and documents, and has good prospects for industrial application. Attached Figure Description

[0023] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0024] Figure 1 This is a flowchart illustrating the overall method of an image uniqueness anti-counterfeiting recognition method based on multimodal feature combination according to the present invention. Figure 2 This is a flowchart of the multimodal feature extraction and fusion process during the registration phase of this invention; Figure 3 This is a flowchart of the authenticity determination and counterfeit attack detection process during the verification phase of this invention; Figure 4 This is a diagram of the overall architecture of the present invention. Detailed Implementation

[0025] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0026] Please see Figures 1-4 A method for image uniqueness anti-counterfeiting recognition based on multimodal feature combination includes: An anti-counterfeiting carrier is constructed, which includes at least a random texture area and a readable identifier area. The random texture area is composed of random microstructures, and the readable identifier area is used to associate the digital identity information of the anti-counterfeiting carrier. During the registration phase, a reference image of the anti-counterfeiting carrier is collected. The deep texture features of the random texture area are extracted based on the texture feature extraction model, and the morphological structure features of the random texture area are extracted based on the morphological feature extraction model. The deep texture features and morphological structure features are fused in a multimodal manner to generate a unique feature fingerprint. The unique feature fingerprint is then bound and stored with the digital identity information corresponding to the readable identifier area. During the verification phase, the test image of the carrier under test is acquired through the terminal device, the random texture area is located according to the readable identifier area, and the test texture features of the random texture area under test are extracted based on the texture feature extraction model. Based on the metric learning algorithm, the texture features to be tested are mapped to the same feature space as the unique feature fingerprints stored in the database, and then the similarity is calculated and compared. Based on the comparison results, the authenticity determination information is output.

[0027] Specifically, the random texture area is composed of micro-serrations formed by the natural diffusion of printed ink edges, or of fibers randomly distributed in the substrate, or of micron- or nano-sized particles randomly distributed in the coating. The morphology of the micro-serrations includes the number of serrations, the serration amplitude, the serration spacing, and the burr length distribution; the distribution characteristics of the fibers include the fiber orientation angle, fiber density, the number of fiber intersections, and the fiber curvature; the distribution characteristics of the particles include the centroid position of the particles, the particle spacing, the particle size distribution, and the degree of particle aggregation.

[0028] Specifically, the readable identifier area is an identifiable identifier; the readable identifier area and the random texture area are arranged adjacent to or nested in spatial position to form a one-to-one hard association relationship, which is used to associate and index the unique feature fingerprint of the anti-counterfeiting carrier through the readable identifier area.

[0029] Specifically, the texture feature extraction model is a deep learning model built on a convolutional neural network, which includes multiple convolutional layers, pooling layers and fully connected layers, used to extract the texture gradient, fiber orientation and local structural difference features of the random texture region at the microscale. The texture feature extraction model also includes a Transformer branch, which is set in parallel or serially with the convolutional neural network. The Transformer branch is used to extract the global spatial relationship features and long-range dependency features of the random texture region, and to fuse the local features extracted by the convolutional neural network with the global features extracted by the Transformer branch.

[0030] Specifically, the morphological feature extraction model is built based on traditional image processing algorithms and includes an edge detection module, a morphological analysis module, and a geometric quantization module; The edge detection module uses the Canny operator or the Sobel operator to extract the ink edge; the morphological analysis module calculates the number of jagged edges, the distribution of burr length, and the concavity and convexity of the edge; the geometric quantization module calculates the convex hull area, Fourier descriptor, and shape context features of the edge, and quantizes the above features to form a structured morphological feature vector.

[0031] Specifically, the deep texture features and morphological structure features are fused in a multimodal manner, which includes: normalizing the deep texture features and morphological structure features respectively, generating a fused feature vector by weighted fusion, splicing fusion or attention-based fusion, and using the fused feature vector as the unique feature fingerprint; the weight coefficients of the weighted fusion are adaptively adjusted or pre-calibrated according to the importance of the feature dimension.

[0032] Specifically, after acquiring the reference image of the anti-counterfeiting carrier, the method further includes a preprocessing step for the reference image: the preprocessing includes illumination consistency correction, geometric distortion correction, image denoising enhancement, and automatic cropping of the anti-counterfeiting area based on the spatial position of the readable identifier area; the geometric distortion correction includes perspective transformation correction and affine transformation correction.

[0033] Specifically, before binding and storing the unique feature fingerprint with the digital identity information corresponding to the readable identifier area, the method further includes encrypting the unique feature fingerprint: the encryption process includes, but is not limited to, generating a digital digest using a hash operation, generating encrypted ciphertext using a symmetric encryption algorithm, or generating a digital signature using an asymmetric encryption algorithm, and storing the encrypted feature fingerprint and the digital identity information in a database.

[0034] Specifically, the step of acquiring the image of the carrier under test through the terminal device includes: acquiring the image of the carrier under test containing the readable identifier area and the random texture area through the camera of the terminal device; and locating the random texture area through an image registration algorithm based on the spatial position relationship of the readable identifier area, or automatically identifying and cropping the image area of ​​the random texture area through a target detection model.

[0035] Specifically, the calculation of similarity for comparison includes: using cosine similarity, Euclidean distance or Mahalanobis distance to calculate the similarity between the texture feature to be tested and the unique fingerprint feature in the feature space; The metric learning algorithm is trained using a triplet loss function or a contrastive loss function; it outputs authenticity determination information based on a preset similarity threshold. When the similarity is greater than or equal to the threshold, it is determined to be genuine; otherwise, it is determined to be suspected counterfeit.

[0036] Specifically, when outputting authenticity determination information based on the comparison results, a comprehensive judgment is also made in conjunction with a counterfeit attack detection mechanism: the counterfeit attack detection mechanism includes focus accuracy detection, moiré pattern detection, and three-dimensional micromorphological analysis. The focus accuracy detection is used to determine whether the image under test meets the preset clarity requirements; the moiré pattern detection is used to identify periodic textures generated by flipping the screen; and the three-dimensional micromorphological analysis is used to determine the three-dimensional physical structure of random texture areas through depth information. When signs of spoofing are detected, the similarity score is lowered or a fake message is displayed directly.

[0037] Specifically, the database is a cloud-distributed database or a local server database; the comparison process between the texture feature to be tested and the unique feature fingerprint is completed on the cloud server, or the unique feature fingerprint is cached locally on the terminal device for offline comparison; when offline comparison is used, the terminal device periodically updates the feature fingerprint data in the database. Example 1

[0038] This embodiment uses the anti-counterfeiting application of high-end liquor as an example to provide a detailed description of the technical solution of the present invention. Those skilled in the art should understand that this embodiment is for illustrative purposes only and does not constitute a limitation on the scope of protection of the present invention. The present invention is also applicable to other product fields requiring anti-counterfeiting and traceability, such as food and pharmaceuticals, electronic products, and documents.

[0039] I. Construction of Anti-counterfeiting Carrier This embodiment involves fabricating an integrated anti-counterfeiting carrier on the anti-counterfeiting label substrate for liquor, as detailed below: 1.1 Setting the readable identifier area High-precision digital printing technology is used to print random pattern identifiers at designated locations on the label. These random patterns are generated by a computer and contain complex intersecting lines, color blocks, or irregular geometric shapes. They are unique and unpredictable, serving as spatial positioning references to associate digital identity information with index fingerprint features. The random pattern identifier area is 8mm × 8mm in size, with high pattern complexity and an extremely low probability of repetition, ensuring that identification and location can still be achieved through pattern features even with partial damage.

[0040] 1.2 Random Texture Area Settings Special penetrating ink is used for printing the random texture area of ​​the label. This ink has specific surface tension and viscosity characteristics. When the ink comes into contact with the paper substrate, the capillary action of the paper fibers causes a natural wetting and diffusion effect at the ink edge, producing a unique microscopic serrated texture. The random texture area is 10mm × 10mm in size and is placed adjacent to the QR code area. To ensure the randomness of the texture, the ambient temperature and humidity are strictly controlled during the printing process (temperature 20±2℃, humidity 45±5%), maintaining a stable and consistent interaction between the ink and paper, while ensuring that the serrated texture of each label is unique due to local differences in the paper fibers.

[0041] 1.3 Establishing Hard Associations The QR code and the random texture area are spatially adjacent, forming a one-to-one hard association. During the registration phase, the system binds the digital identity information associated with the QR code to the unique feature fingerprint extracted from the random texture area, ensuring that each QR code uniquely corresponds to a physically unclonable random texture feature, preventing malicious reuse after the feature fingerprint and identity information are separated.

[0042] II. Registration Phase: Image Acquisition, Preprocessing, and Multimodal Feature Extraction 2.1 Reference Image Acquisition Industrial high-resolution scanners are used to capture original high-definition images of anti-counterfeiting labels. The image resolution is no less than 4800×3600 pixels, the color depth is 24-bit true color, and the image format is lossless (such as TIFF or PNG). During acquisition, ensure that the anti-counterfeiting labels are laid flat, the lighting is uniform, and the QR code and random texture areas are fully covered to ensure accurate location and feature extraction in subsequent processing.

[0043] 2.2 Image Preprocessing The following preprocessing steps are performed on the reference image in sequence: Illumination consistency correction: An adaptive histogram equalization algorithm is used to eliminate brightness deviations caused by uneven ambient light, making the image brightness distribution more uniform and avoiding the impact of illumination differences on feature extraction.

[0044] Geometric distortion correction: By using perspective transformation and affine transformation, the image distortion caused by the shooting angle deviation is corrected, and the label image is corrected to a standard front view to ensure spatial consistency for subsequent feature comparison.

[0045] Image denoising and enhancement: A bilateral filtering algorithm is used to preserve microscopic jagged details while suppressing background noise, and the Laplacian operator is used to enhance edge contrast, making texture details clearer.

[0046] Automatic anti-counterfeiting area cropping: Based on the positioning pattern of the QR code (such as the position detection graphic), the coordinates of the QR code area are identified, and a random texture area is accurately cropped according to the preset spatial offset to ensure the spatial consistency between the cropped area and the area collected in the subsequent verification stage.

[0047] 2.3 Depth Texture Feature Extraction A hybrid model combining CNN (Convolutional Neural Network) and Transformer is used to extract deep texture features from random texture regions. The CNN branch employs an improved ResNet-50 architecture, containing multiple convolutional, pooling, and fully connected layers to extract texture gradients, jagged edge orientations, fiber orientations, and local structural differences in random texture regions at a microscale. The network is pre-trained on the ImageNet dataset and fine-tuned on an anti-spoofing texture dataset to enhance its ability to recognize specific texture features.

[0048] Transformer Branch: Employing the Vision Transformer architecture, set up in parallel with the CNN branch, it is used to extract global spatial relationship features and long-range dependency features of random texture regions. The Transformer captures global correlations between pixels through a self-attention mechanism, compensating for the limitation of the receptive field of CNNs.

[0049] Feature fusion: The local feature vectors output by the CNN branch are concatenated and fused with the global feature vectors output by the Transformer branch to form a 512-dimensional deep texture feature vector.

[0050] 2.4 Extraction of Morphological and Structural Features A morphological feature extraction model is constructed based on traditional image processing algorithms. This model consists of three core modules: Edge detection module: The Canny operator is used to extract the jagged outline of the ink. By setting high and low thresholds (high threshold 150, low threshold 50), continuous and complete edge information is extracted while suppressing noise interference.

[0051] Morphological analysis module: Based on edge detection, it calculates quantitative indicators such as the number of serrations, serration amplitude, serration spacing, and burr length distribution. The number of serrations counts the number of serrations per unit length; the serration amplitude calculates the average vertical distance from the tooth tip to the tooth valley; the serration spacing calculates the horizontal distance between adjacent tooth tips; and the burr length distribution analyzes the length distribution characteristics of edge burrs.

[0052] The geometric quantization module calculates the convex hull area of ​​the edge (reflecting the overall distribution range of the jagged region), the Fourier descriptor (characterizing the shape features of the edge), and the shape context features (describing the spatial distribution relationship of edge points). After quantization, these features form a structured morphological feature vector with 256 dimensions.

[0053] 2.5 Multimodal Fusion and Unique Feature Fingerprint Generation Feature normalization: The depth texture feature vector (512-dimensional) and the morphological structure feature vector (256-dimensional) are normalized using the L2 norm to eliminate differences in feature dimensions and ensure that the contributions of each feature dimension are balanced during fusion.

[0054] Feature fusion: A weighted fusion method based on an attention mechanism is adopted. The two types of features are adaptively weighted using a learnable attention weight vector to generate a fused feature vector (512 dimensions). The attention weights are automatically adjusted according to the importance of the feature dimensions, making the fused features more discriminative.

[0055] Feature fingerprint generation: The fused feature vector is hashed using SHA-256 to generate a 256-bit digital digest, forming an immutable and unique feature fingerprint. The hash operation ensures that even if the feature vector changes slightly, the generated hash value will produce a significant difference, enhancing the sensitivity of the feature fingerprint.

[0056] Encrypted Storage: The fingerprint is bound to the QR code identity information, encrypted using an asymmetric encryption algorithm (RSA-2048), and then stored in a cloud-based distributed database. The encryption process uses a public key, while the private key is held only by the authorized verification server, preventing the fingerprint from being tampered with or illegally obtained during transmission and storage.

[0057] III. Verification Phase: User Terminal Data Collection and Authenticity Verification 3.1 Image Acquisition Users open the anti-counterfeiting verification mini-program on their smartphones, click the "Scan Code Verification" button, and the mini-program uses the phone's camera to scan the label's QR code. Upon successful QR code recognition, the mini-program automatically triggers macro shooting mode, guiding the user to point the phone's camera at a random textured area to capture an image containing that area. During the capture process, the system monitors image clarity in real time; if blurry or shaky images are detected, the system automatically prompts the user to adjust the shooting distance or stabilize the phone.

[0058] 3.2 Random Texture Region Location Based on the spatial relationship of the QR code, random texture areas are automatically located using a scheme that combines image registration and object detection: Coarse localization: Based on the positioning pattern recognition of the QR code, the coordinates of the QR code area are identified, and the candidate areas of the random texture area are coarsely located according to the spatial offset (X-direction offset, Y-direction offset) preset in the registration stage.

[0059] Precise localization: Within the candidate region, an object detection model (such as YOLOv5) is used to precisely locate the boundary of the random texture region. The image to be tested is then registered with the standard template in the registration stage using an image registration algorithm (based on SIFT feature point matching) to ensure that the cropped random texture region has the same spatial scale and rotation angle as the registration stage.

[0060] 3.3 Feature Extraction A lightweight CNN model (MobileNetV3) is run on a mobile device to extract the texture features of the test image in real time. This model is obtained through knowledge distillation, transferring learning from the large model (CNN+Transformer) used in the registration phase, significantly reducing computational complexity while maintaining high recognition accuracy. The model's inference time on the mobile device is less than 200 milliseconds, meeting real-time requirements. The extracted feature vector has a dimension of 512, consistent with the dimension of the deep texture feature vector in the registration phase, facilitating subsequent similarity calculations.

[0061] 3.4 Metric Learning Alignment Feature space mapping: The texture features to be tested are mapped to the unique feature fingerprint stored in the cloud to the same feature space. Since the texture features to be tested are depth feature vectors, while the unique feature fingerprints are hash values, this embodiment achieves the mapping by establishing a feature index table: the fused feature vectors from the registration stage are pre-stored in the database, and during verification, the corresponding fused feature vectors are directly retrieved and compared with the texture features to be tested.

[0062] Similarity Calculation: Cosine similarity is used to calculate the distance between the texture feature to be tested and the fused feature vector in the feature space. The cosine similarity value ranges from [-1, 1], with values ​​closer to 1 indicating greater feature similarity. The calculation formula is: Similarity = (A·B) / (|A|×|B|), where A is the texture feature to be tested and B is the fused feature vector.

[0063] Metric learning training: The model is trained using a triplet loss function. During training, triplets are constructed consisting of anchor samples (genuine products), positive samples (another genuine product with the same label), and negative samples (counterfeit products). The distance between the anchor and the positive sample in the feature space is made smaller than the distance between the anchor and the negative sample, thereby enhancing the discriminative ability of the feature space.

[0064] 3.5 Detection of Impersonation Attacks Simultaneously execute a triple spoofing attack detection mechanism: Focus accuracy detection: The Laplacian variance algorithm is used to evaluate the sharpness of the image under test. When the variance value is lower than a preset threshold, the image is considered to be insufficiently sharp. This detection mechanism is used to identify image blur caused by copying or low-quality shooting, avoiding blurry images from affecting the accuracy of recognition.

[0065] Moiré pattern detection: This involves performing frequency domain analysis on the image under test to detect the presence of periodic high-frequency interference. Moiré patterns typically occur during screen copying due to interference between screen pixels and camera sensor pixels. The presence of abnormally high-frequency peaks in the frequency domain indicates the presence of moiré patterns, suggesting potential screen copying fraud.

[0066] Three-dimensional microscopic morphological analysis: Utilizing multi-frame image information from the smartphone's autofocus process, the depth-of-field distribution of random texture areas is calculated, and the three-dimensional physical structure of the texture is analyzed. Genuine physical textures (such as ink jaggedness or fiber protrusions) possess three-dimensional characteristics, with irregular variations in depth-of-field distribution; while flat-printed counterfeits have a uniform depth-of-field distribution and lack three-dimensional variation. By analyzing the variance and gradient changes in the depth-of-field distribution, the difference between genuine textures and flat-printed products can be effectively distinguished.

[0067] 3.6 Authenticity Determination and Result Output Decision logic: When the similarity is greater than or equal to the preset threshold (set to 0.85 in this embodiment) and there are no abnormalities in the triple attack detection, it is determined to be genuine.

[0068] When the similarity is less than the threshold or any attack detection mechanism is triggered, it is judged as a suspected forgery.

[0069] Results feedback: Authentic Product: Returns to the "Authentic Product" conclusion, displays product traceability information (production batch, production date, quality inspection report, etc.), and enlarges the microscopic comparison image of the random texture area to intuitively display anti-counterfeiting features.

[0070] Suspected forgery: Returns a "Suspected forgery" conclusion, provides a counterfeit type hint (such as "insufficient clarity", "may be a screen copy", "abnormal texture 3D structure"), and lowers the credibility score.

[0071] Inspection Records: Each verification result is automatically uploaded to the cloud server to form an inspection log, which records information such as verification time, verification equipment, and verification results, supporting traceability management throughout the product lifecycle.

[0072] IV. Technical Effects of this Embodiment Physically unclonable: The immersion-style micro-serrated texture formed by penetrating ink originates from the random distribution of paper fibers and cannot be replicated by high-definition printing, copying, or digital printing. Each label has a unique and unclonable physical fingerprint.

[0073] High recognition accuracy: By multimodal fusion of deep learning texture features and morphological structure features, combined with feature space optimization trained by metric learning, the ability to distinguish between real and fake samples is significantly improved.

[0074] Excellent user experience: Ordinary users can complete macro photography and authenticity verification using only a smartphone, without the need for professional equipment. The operation process is simple and the feedback results are intuitive, significantly reducing the barrier to entry for anti-counterfeiting verification.

[0075] Strong anti-attack capability: It integrates multiple anti-attack mechanisms such as focus accuracy detection, moiré pattern detection, and three-dimensional micro-morphology analysis, which can effectively identify common counterfeiting methods such as high-definition printing, copying, film imitation, and screen copying, greatly improving the security and reliability of the system. Example 2

[0076] This embodiment takes anti-counterfeiting of pharmaceutical packaging as an example to demonstrate a random texture carrier form, fusion method, and deployment scheme that differs from the main embodiment.

[0077] I. Construction of Anti-counterfeiting Carrier Readable Identification Area: A random pattern is laser-engraved on the surface of the medicine packaging box. This random pattern is generated by computer and contains complex intersecting lines, color blocks, or irregular geometric shapes. It is unique and unpredictable, serving as a spatial positioning reference for associating digital identity information with index fingerprint features. The random pattern identification area is 8mm × 8mm in size, with high pattern complexity and an extremely low probability of repetition, ensuring that identification and positioning can still be achieved through pattern features even in cases of partial damage.

[0078] Random texture area: A transparent coating containing randomly distributed micron-sized fibers with a length of 50-200 microns is applied to a designated area of ​​the packaging box to form a non-replicable fiber network texture with a size of 8mm×8mm.

[0079] Hard association: Random pattern identifiers and random texture areas are arranged adjacently, forming a one-to-one binding relationship.

[0080] II. Registration Phase Image acquisition: Reference images are acquired using an industrial line scan camera with a resolution of no less than 2400×2400 pixels.

[0081] Preprocessing: Same as the main embodiment, performing illumination correction, geometric correction, noise reduction and enhancement, and region clipping.

[0082] Deep texture feature extraction: ResNet-50 is used to extract features such as fiber orientation, fiber density, and intersection distribution, and outputs a 512-dimensional vector.

[0083] Morphological and structural feature extraction: Based on the Sobel operator, fiber edges are extracted, and features such as fiber orientation angle distribution, fiber density, number of intersections, fiber curvature, and convex hull area are calculated, outputting a 256-dimensional vector.

[0084] Multimodal fusion: The splicing fusion method is adopted to directly splice the two types of features to generate a 768-dimensional fusion feature vector, and generate a unique feature fingerprint through SHA-256 hashing.

[0085] Encrypted storage: The data is stored in the local server database after being encrypted using AES symmetric encryption.

[0086] III. Verification Phase Image acquisition: Quality inspectors use handheld terminals to scan traceable digital data and automatically take macro photos.

[0087] Localization and feature extraction: Random texture regions are located through image registration, and the handheld terminal runs MobileNetV2 to extract the texture features (512 dimensions) to be tested.

[0088] Metric alignment: The texture features to be tested and the fused feature vector are mapped to the same 256-dimensional space by PCA, and the similarity is calculated by Euclidean distance.

[0089] Authenticity determination: When the Euclidean distance is ≤0.75, the product is determined to be genuine, and traceability information is output; otherwise, it is determined to be suspected of being counterfeit. The verification results are stored on the local server.

[0090] IV. Technical Effects This embodiment is applicable to scenarios such as pharmaceutical and food packaging that require internal quality traceability management. It uses fiber-distributed random textures to enrich the carrier form, and the splicing and fusion method has high computational efficiency. Local server deployment is suitable for internal quality inspection within enterprises. Example 3

[0091] This embodiment uses electronic product warranty and anti-counterfeiting as an example to demonstrate a particle-distributed random texture and offline verification scheme.

[0092] I. Construction of Anti-counterfeiting Carrier Readable identification area: A 6mm x 6mm anti-counterfeiting QR code is laser-engraved on the body of the electronic product.

[0093] Random texture area: Randomly distributed nano-sized silica particles (100-500 nanometers) are doped into the coating of the warranty label to form a random particle distribution structure with a size of 5mm×5mm, which is nested with the QR code.

[0094] Hard association: The QR code and the random texture area are bound to each other in a one-to-one correspondence, and the feature fingerprint is pre-stored locally on the terminal, supporting offline verification.

[0095] II. Registration Phase Image acquisition: Images are acquired using a high-magnification microscope camera with a resolution of no less than 4000×3000 pixels and 20x magnification.

[0096] Preprocessing: Same as in the main embodiment.

[0097] Deep texture feature extraction: Lightweight MobileNetV3 is used to extract granular texture features, and a 256-dimensional vector is output.

[0098] Morphological and structural feature extraction: Adaptive threshold segmentation is used to extract particle contours, and features such as particle number, centroid coordinates, spacing distribution, particle size distribution, and Euler number are calculated to output a 128-dimensional vector.

[0099] Multimodal fusion: Weighted fusion (texture weight 0.6, morphology weight 0.4) is used to generate a 256-dimensional fusion feature vector, which is then quantized and compressed into a 128-dimensional compact feature fingerprint.

[0100] Pre-stored to terminal: After encryption, the feature fingerprint is pre-stored to the local database of the smartphone, supporting offline verification and regular synchronization updates.

[0101] III. Verification Phase (Offline Mode) Image capture: Users scan a QR code with their smartphones in an offline environment, and the system automatically takes macro photos.

[0102] Localization and feature extraction: Random texture regions were cropped using the YOLOv5-Lite object detection model, and MobileNetV3 was run on the mobile device to extract the texture features (256 dimensions).

[0103] Metric alignment: The texture features to be tested and the local compact feature fingerprint are mapped to the same 128-dimensional space by a pre-trained projection matrix, and cosine similarity is used for calculation.

[0104] Counterfeit attack detection: Simultaneously perform focus accuracy detection and 3D microscopic morphology analysis (based on multi-frame image depth information) to distinguish between real grain textures and flat printed materials.

[0105] Authenticity Verification: If the similarity is ≥0.80 and there are no signs of tampering, the product is considered genuine and warranty information will be displayed; otherwise, a suspected counterfeit message will be displayed, and a second online verification is recommended. Verification records are cached locally and will be uploaded after the network is restored.

[0106] IV. Technical Effects This embodiment is applicable to scenarios such as electronic product warranty and offline ticket verification. The nanoparticle distribution is suitable for miniaturized anti-counterfeiting, the lightweight model supports real-time inference on the terminal, and the offline verification function is suitable for environments without network access.

[0107] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An image uniqueness anti-counterfeiting identification method based on multi-modal feature combination, characterized in that, include: An anti-counterfeiting carrier is constructed, which includes at least a random texture area and a readable identifier area. The random texture area is composed of random microstructures, and the readable identifier area is used to associate the digital identity information of the anti-counterfeiting carrier. During the registration phase, a reference image of the anti-counterfeiting carrier is collected. The deep texture features of the random texture area are extracted based on the texture feature extraction model, and the morphological structure features of the random texture area are extracted based on the morphological feature extraction model. The deep texture features and morphological structure features are fused in a multimodal manner to generate a unique feature fingerprint. The unique feature fingerprint is then bound and stored with the digital identity information corresponding to the readable identifier area. During the verification phase, the test image of the carrier under test is acquired through the terminal device, the random texture area is located according to the readable identifier area, and the test texture features of the random texture area under test are extracted based on the texture feature extraction model. Based on the metric learning algorithm, the texture features to be tested are mapped to the same feature space as the unique feature fingerprints stored in the database, and then the similarity is calculated and compared. Based on the comparison results, the authenticity determination information is output.

2. The method according to claim 1, characterized in that, The random texture area is composed of micro-serrations formed by the natural diffusion of printed ink edges, or of fibers randomly distributed in the substrate, or of micron- or nano-sized particles randomly distributed in the coating. The morphology of the micro-serrations includes the number of serrations, the serration amplitude, the serration spacing, and the burr length distribution; the distribution characteristics of the fibers include the fiber orientation angle, fiber density, the number of fiber intersections, and the fiber curvature; the distribution characteristics of the particles include the centroid position of the particles, the particle spacing, the particle size distribution, and the degree of particle aggregation.

3. The method according to claim 1, characterized in that, The readable identifier area is an identifiable identifier. The readable identifier area and the random texture area are arranged adjacent to or nested in space to form a one-to-one hard association relationship, which is used to associate and index the unique feature fingerprint of the anti-counterfeiting carrier through the readable identifier area.

4. The method according to claim 1, characterized in that, The texture feature extraction model is a deep learning model built on a convolutional neural network. The convolutional neural network includes multiple convolutional layers, pooling layers, and fully connected layers, which are used to extract the texture gradient, fiber orientation, and local structural difference features of the random texture region at the microscale. The texture feature extraction model also includes a Transformer branch, which is set in parallel or serially with the convolutional neural network. The Transformer branch is used to extract global spatial relationship features and long-range dependency features of random texture regions, and to fuse the local features extracted by the convolutional neural network with the global features extracted by the Transformer branch.

5. The method according to claim 1, characterized in that, The morphological feature extraction model is built based on traditional image processing algorithms, including an edge detection module, a morphological analysis module, and a geometric quantization module; The edge detection module uses the Canny operator or the Sobel operator to extract the ink edge; the morphological analysis module calculates the number of jagged edges, the distribution of burr length, and the concavity and convexity of the edge; the geometric quantization module calculates the convex hull area, Fourier descriptor, and shape context features of the edge, and quantizes the above features to form a structured morphological feature vector.

6. The method according to claim 1, characterized in that, The deep texture features and morphological structure features are fused in a multimodal manner, specifically including: normalizing the deep texture features and morphological structure features respectively, generating a fused feature vector by weighted fusion, splicing fusion or attention-based fusion, and using the fused feature vector as a unique feature fingerprint; the weight coefficients of the weighted fusion are adaptively adjusted or pre-calibrated according to the importance of the feature dimension.

7. The method according to claim 1, characterized in that, After acquiring the reference image of the anti-counterfeiting carrier, the method further includes a preprocessing step for the reference image: the preprocessing includes illumination consistency correction, geometric distortion correction, image denoising enhancement, and automatic cropping of the anti-counterfeiting area based on the spatial position of the readable identifier area; the geometric distortion correction includes perspective transformation correction and affine transformation correction.

8. The method according to claim 1, characterized in that, Before binding and storing the unique feature fingerprint with the digital identity information corresponding to the readable identifier area, the method further includes encrypting the unique feature fingerprint: the encryption process includes, but is not limited to, generating a digital digest using a hash operation, generating encrypted ciphertext using a symmetric encryption algorithm, or generating a digital signature using an asymmetric encryption algorithm, and storing the encrypted feature fingerprint and digital identity information together in the database.

9. The method according to claim 1, characterized in that, The acquisition of the test image of the test carrier through the terminal device specifically includes: acquiring the test image containing a readable identifier area and a random texture area through the camera of the terminal device; and locating the random texture area through an image registration algorithm based on the spatial relationship of the readable identifier area, or automatically identifying and cropping the image area of ​​the random texture area through a target detection model.

10. The method according to claim 1, characterized in that, The calculation of similarity comparison specifically includes: using cosine similarity, Euclidean distance or Mahalanobis distance to calculate the similarity between the texture feature to be tested and the unique fingerprint feature in the feature space; The metric learning algorithm is trained using a triplet loss function or a contrastive loss function; it outputs authenticity determination information based on a preset similarity threshold. When the similarity is greater than or equal to the similarity threshold, it is determined to be genuine; otherwise, it is determined to be suspected counterfeit.

11. The method according to claim 10, characterized in that, When outputting authenticity determination information based on the comparison results, a comprehensive judgment is also made in conjunction with the counterfeit attack detection mechanism: the counterfeit attack detection mechanism includes focus accuracy detection, moiré pattern detection, and three-dimensional micromorphological analysis. The focus accuracy detection is used to determine whether the image under test meets the preset clarity requirements; the moiré pattern detection is used to identify periodic textures generated by flipping the screen; and the three-dimensional micromorphological analysis is used to determine the three-dimensional physical structure of random texture areas through depth information. When signs of spoofing are detected, the similarity score is lowered or a fake message is displayed directly.

12. The method according to claim 1, characterized in that, The database is either a cloud-based distributed database or a local server database; the comparison process between the texture features to be tested and the unique feature fingerprints is completed on the cloud server, or the unique feature fingerprints are cached locally on the terminal device for offline comparison; when offline comparison is used, the terminal device periodically updates the feature fingerprint data in the database.

Images