Object digital fingerprinting method and system based on physical microfeatures

Abstract

This invention discloses a method and system for digital fingerprint recognition of objects based on physical microscopic features. The method addresses the problems of easily copied, transferred, or damaged anti-counterfeiting labels on high-value items, and the decreased robustness of matching due to the influence of lighting, local contamination, and wear on the texture of items during actual circulation and photography. It normalizes the ROI through LAB color space processing and a spatial transformation network; inputs the normalized ROI into explicit geometric feature pathways and implicit microscopic representation pathways, and fuses local keypoint matrices with dense vectors from a self-supervised visual model to generate a digital fingerprint code through dimensionality reduction; during registration, the fingerprint is written to an HNSW-based vector database; during recognition, Top-K candidates are retrieved, and the matching result is output through dynamic local grid voting, mask removal, and consistency verification. This invention also discloses a corresponding system, terminal device, computer-readable storage medium, and computer program product, which can significantly improve the accuracy, efficiency, and resistance to damage in object traceability and authentication.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, pattern recognition, digital image processing, anti-counterfeiting and traceability management of goods, and in particular to a method, system, terminal device and computer-readable storage medium for generating digital fingerprints using the natural physical microscopic features of an object's surface and performing high-precision closed-loop identification. Background Technology

[0002] In the circulation of high-end consumer goods, cultural relics and antiques, works of art, precision industrial components, precious metal products, documents and vouchers, and important assets, the identification of items, anti-counterfeiting authentication, and traceability management are crucial links in ensuring transaction security and asset credibility. Existing anti-counterfeiting solutions typically rely on external identifiers such as QR codes, RFID chips, fluorescent tags, anti-counterfeiting labels, or manual certificates. However, these external identifiers are not inherently integrated with the object itself and are easily transferred, copied, replaced, or damaged, making it difficult to fundamentally prove "this is the real thing."

[0003] Utilizing the naturally occurring microscopic textures, pores, scratches, fiber arrangements, metal processing patterns, glaze particles, and leather textures of an object's surface for identification has become an important direction in the field of anti-counterfeiting authentication in recent years. These features are characterized by their natural randomness, difficulty in replication, and inherent presence within the object itself, theoretically serving as a "physical fingerprint" for the object. However, in practical applications, objects are often photographed by different users under different lighting conditions, angles, devices, and backgrounds; furthermore, objects may develop localized stains, minor wear, reflections, scratches, or obstructions during circulation, leading to significant differences between multiple images of the same object.

[0004] Traditional image matching methods often rely on local keypoints or grayscale texture statistics. When images have nonlinear geometric distortions, local defects, or insufficient texture intensity, relying on a single explicit feature can easily lead to problems such as insufficient keypoint count, increased mismatches, and unstable overall similarity. On the other hand, while deep learning models have strong representation capabilities, they are easily affected by differences in lighting, background, and shooting equipment if they lack domain constraints on the microscopic texture of objects. Furthermore, the computational cost of directly comparing each of the massive registered samples is too high.

[0005] Therefore, there is an urgent need for a digital fingerprint recognition method and system that can fully utilize the microscopic physical characteristics of the object, adapt to complex shooting environments, have the ability to tolerate local damage, and support real-time retrieval of massive samples. Summary of the Invention

[0006] The purpose of this invention is to provide a digital fingerprint recognition method and system for objects based on physical microscopic features, in order to overcome the problems of existing external anti-counterfeiting marks being easily forged or transferred, traditional image matching being sensitive to illumination and geometric distortion, matching failure due to local contamination or wear, and insufficient retrieval efficiency of massive samples.

[0007] To achieve the above objectives, the present invention adopts the following technical solution.

[0008] Step S1, Image Acquisition and Spatial Normalization: Acquire local identification point images of the object to be identified and locate the region of interest (ROI); convert the image to the LAB color space, remove the L brightness channel which is greatly affected by illumination intensity, and retain the A channel, B channel and gradient direction information; estimate the affine or perspective transformation parameters through a spatial transformation network and output a normalized ROI image.

[0009] Step S2, parallel feature extraction: The standardized ROI image is simultaneously input into the explicit geometric feature pathway and the implicit micro-representation pathway. The explicit geometric feature pathway extracts a local keypoint matrix with scale and rotation invariance; the implicit micro-representation pathway uses a self-supervised visual model to extract a high-dimensional dense vector that can represent the distribution of micro-texture.

[0010] Step S3, Digital fingerprint fusion encoding: Determine the fusion weights of the two feature paths based on texture richness, number of key points and image quality score, and concatenate, normalize and reduce the dimensionality of the local key point matrix and dense vector to generate a compact and globally unique digital fingerprint encoding.

[0011] Step S4, Registration and Index Construction: In the initial identification or storage stage, the digital fingerprint code is associated with the item identifier, identification point location, registered ROI image and collected metadata and stored, and an approximate nearest neighbor vector index is constructed based on the hierarchical navigation small world graph algorithm.

[0012] Step S5, Massive Search and Candidate Screening: In the re-identification stage, the digital fingerprint code to be tested is extracted according to steps S1-S3, an approximate nearest neighbor search is performed in the vector database, the cosine similarity between the code to be tested and the code in the database is calculated, and Top-K candidate objects are screened.

[0013] Step S6, Dynamic Mesh Fault Tolerance Verification: Perform local mesh generation, region masking, local distance calculation, and voting judgment on Top-K candidate objects; when the proportion of effective matching sub-mesh and the overall similarity meet the set conditions, output the matching success result; otherwise, output the rejection result or trigger manual review.

[0014] Step S7, Lightweight Deployment Optimization: Perform activation value-aware weight quantization on the implicit micro-representation model, and combine paging attention or feature caching mechanisms to reduce GPU and memory overhead, so that the system can be deployed on mobile terminals, handheld identification devices, edge gateways or industrial cameras.

[0015] This invention also provides an object digital fingerprint recognition system, including an image acquisition and normalization module, a dual-track feature extraction module, a digital fingerprint generation module, a registration and indexing module, a vector retrieval and candidate selection module, a grid fault-tolerant verification module, and a lightweight optimized deployment module. These modules work together sequentially to form an end-to-end authentication pipeline for registration, retrieval, verification, and review. Beneficial effects

[0016] First, this invention combines LAB color space processing, gradient direction information, and spatial transformation network to reduce the impact of light intensity, shadows, and shooting posture changes on micro-texture recognition during the preprocessing stage, thereby improving image consistency in complex environments.

[0017] Second, the present invention adopts a dual-track fusion of explicit geometric features and implicit microscopic representations, enabling the system to utilize both the interpretable geometric constraints of local key points and capture microscopic texture differences that are difficult to describe stably with the naked eye, thereby improving the uniqueness and anti-counterfeiting ability of digital fingerprints.

[0018] Third, this invention transforms massive sample comparisons into vector retrieval and small-range precise verification through the HNSW approximate nearest neighbor index, significantly reducing retrieval complexity and making it suitable for large-scale warehousing, trading, and industrial production line applications.

[0019] Fourth, this invention improves fault tolerance and maintainability in real-world scenarios by using dynamic local mesh voting and mask culling mechanisms, ensuring that local contamination, scratches, wear, or occlusion do not directly lead to overall matching failure.

[0020] Fifth, this invention supports deployment on edge or handheld devices through quantization and caching optimization, and can form a closed-loop application in on-site identification, mobile inspection and industrial real-time detection. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the overall process of object digital fingerprint recognition based on physical microscopic features according to the present invention; Figure 2 This is a schematic diagram of the system module composition and data flow of the present invention; Figure 3 This is a schematic diagram of the dual-track parallel feature extraction network architecture of the present invention; Figure 4 This is a schematic diagram of the dynamic local mesh fault tolerance comparison and voting determination of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. For those skilled in the art, various modifications, substitutions, and combinations can be made without departing from the concept of the present invention, and all of these should be included within the scope of protection of the present invention.

[0023] Example 1: Method for Digital Fingerprint Registration and Recognition of Objects This embodiment provides a method for registering and recognizing digital fingerprints of objects, applicable to objects with observable surface micro-textures, such as leather products, metal parts, paper certificates, ceramics, artworks, and precious metal products. During the registration phase, the system generates a digital fingerprint for each object or each preset identification point; during the recognition phase, identity verification is completed by re-acquiring images.

[0024] In step S1, the system acquires images of local identification points using an industrial camera, mobile phone camera, microscope module, or handheld identification device. To ensure the stability of subsequent comparisons, the system can select locations on the object surface with rich texture and that are not easily occluded as identification points, and record the relative positions of the identification points with respect to the object's contour, edges, markings, or structural openings. After acquiring the images, the system uses object detection or template localization methods to obtain ROI candidate regions, and calculates a quality score based on edge sharpness, exposure, motion blur, and texture richness. Images below the quality threshold are prompted for re-acquisition.

[0025] Subsequently, the system transforms the ROI candidate regions to the LAB color space. Since the L channel primarily reflects brightness and is easily affected by ambient lighting, shadows, and reflections, this embodiment removes the L channel while retaining the A and B chromaticity channels, and overlays gradient direction information calculated by the Sobel operator, Scharr operator, or histogram of oriented gradients. This information is then input into a spatial transformation network, which outputs an affine matrix or perspective matrix to correct the ROI region to a preset scale and orientation, thereby obtaining a standardized ROI image.

[0026] In step S2, as Figure 3As shown, the standardized ROI image is simultaneously input into both the explicit geometric feature pathway and the implicit micro-representation pathway. The explicit geometric feature pathway extracts local keypoints, principal orientations, and descriptors from the standardized ROI image. SIFT, SURF, ORB, or combinations thereof can be used, and a keypoint matrix is ​​formed based on the spatial location, scale, orientation, and response intensity of the keypoints. The implicit micro-representation pathway employs a self-supervised visual model, such as the DINOv2 backbone network, to perform multi-layer feature extraction on the standardized ROI image, and pools, normalizes, and projects the intermediate layer features to obtain a dense vector of micro-texture.

[0027] To make implicit micro-representations more suitable for object anti-counterfeiting scenarios, this embodiment uses images of the same object acquired under different lighting, angles, and devices to form positive sample pairs, and uses images with similar appearances but different objects to form hard negative sample pairs. During model fine-tuning, a weighted combination of cosine nearest neighbor loss and triplet hard sample mining loss is used to make samples of the same object closer in the feature space, while maintaining sufficient spacing between samples of different objects.

[0028] In step S3, the system calculates texture richness based on the gray-level co-occurrence matrix and counts the number of effective keypoints output by the explicit geometric feature pathway. When the texture richness is high and the keypoints are evenly distributed, the weight of the explicit geometric features is increased; when the texture is fine but the keypoints are insufficient or there is obvious local reflection, the weight of the implicit microscopic representation is increased. The fused features are then normalized by L2 and reduced in dimensionality by principal component analysis or random projection to generate a fixed-length digital fingerprint code. This code can be used as a retrieval vector or hashed to form a short code for fast deduplication and cache hits.

[0029] In step S4, the registration phase writes the digital fingerprint code into the vector database and associates it with the item ID, registration time, acquisition device, identification point location, registration ROI image, quality score, and version number. The vector database uses the HNSW algorithm to construct a multi-layered graph structure index. The bottom layer stores dense connections to improve recall, while the upper layer is used for fast navigation to reduce retrieval latency. The system can asynchronously update the index after adding new samples to balance writing efficiency and retrieval stability.

[0030] In step S5, during the re-identification, the system performs the same image normalization and digital fingerprint generation process as in the registration phase on the image to be tested, obtaining the digital fingerprint code to be tested. The system performs an approximate nearest neighbor search in the vector database, calculates the cosine similarity between the code to be tested and the codes in the database, and returns Top-K candidate objects. The value of K can be set according to the database size and business recall requirements, for example, 5, 10, or 20.

[0031] In step S6, the system performs fine-tuning on each candidate object. Specifically, the image of the ROI to be tested and the candidate registered ROI images are divided into M×N local sub-grids; for each corresponding sub-grid, the number of local descriptor matches, Hamming distance, cosine distance, or a combination thereof are calculated. If a sub-grid is detected as a stain, severe reflection, scratch, occlusion, or wear area, that sub-grid is added to a mask and removed from the voting statistics. The system calculates the proportion of the remaining valid sub-grids that pass the local threshold. When this proportion is greater than a preset proportion threshold and the overall similarity exceeds the recognition threshold, a successful match is determined.

[0032] For samples where the highest similarity score does not reach the absolute confidence threshold but is significantly higher than the rejection threshold, or for samples where the proportion of effective matching sub-grids is in the ambiguity threshold range, the system triggers manual review. The review interface displays the alignment results of the image to be tested and the candidate images, the location of abnormal sub-grids, the local similarity heatmap, and the candidate ranking, assisting the reviewers in making quick judgments. The results of manual review can be written back to the system as a data source for subsequent model optimization and threshold calibration.

[0033] Example 2: System Structure Image acquisition and normalization module It is used to acquire images of local identification points of objects, perform ROI localization, LAB color space processing, gradient direction calculation, spatial transformation network correction and image quality scoring, and output a standardized ROI image. Dual-track feature extraction module It includes explicit geometric feature pathways and implicit microscopic representation pathways. The explicit geometric feature pathway extracts local keypoint matrices; the implicit microscopic representation pathway extracts dense vectors of microscopic textures through a self-supervised visual model. Digital fingerprint generation module It is used to determine the fusion weight based on texture richness, number of key points and quality score, and to normalize, reduce dimensionality, hash or version the fusion features to obtain digital fingerprint encoding. Registration and Indexing Module It is used to manage item IDs, identification points, registered ROI images, digital fingerprint codes, and collect metadata, and to build an approximate nearest neighbor index based on the HNSW algorithm. Vector retrieval and candidate filtering module Used to perform vector retrieval on the digital fingerprint encoding under test, and return Top-K candidate objects and their similarity scores. Mesh fault tolerance verification module It is used to perform local mesh generation, damage mask generation, local distance calculation, voting statistics, matching judgment, and manual review triggering. Lightweight Optimized Deployment Module It is used to perform AWQ INT4 quantization, paging attention, feature caching, and edge inference scheduling, making the system suitable for mobile terminals, handheld identification devices, and industrial edge devices. Example 3: Edge Deployment and Data Closed Loop In edge deployment scenarios, handheld identification devices can integrate a camera, a supplementary lighting module, a processor, a memory, a network communication module, and a display module. The processor can be a CPU, GPU, NPU, DSP, FPGA, or a combination thereof. In an offline environment, the device can first complete image normalization, digital fingerprint extraction, and local small-scale database retrieval; in a networked environment, it can upload the digital fingerprint encoding to the server for large-scale retrieval without uploading the complete original image, thereby reducing bandwidth consumption and improving privacy protection capabilities.

[0034] In terms of model lightweighting, the system can perform activation-aware weight quantization on the self-supervised visual model, converting the weights from floating-point format to INT4 or INT8 format, and maintaining the accuracy of micro-texture representation through calibration samples. For industrial production line scenarios requiring batch processing, the system can use paging attention or feature caching mechanisms to manage dynamic caching, reducing peak memory usage and improving concurrent throughput.

[0035] Manually reviewed results, falsely rejected samples, falsely accepted samples, and low-quality collected samples generated during system operation can be incorporated into the data loop. The system periodically performs threshold calibration, hard negative sample mining, and incremental model fine-tuning based on these samples, thereby continuously improving the robustness of recognition in complex scenarios.

[0036] Example 4: Application Scenarios In high-end leather product authentication scenarios, the system can select areas with natural pores and rich textures in the leather as authentication points and generate a digital fingerprint during registration. Subsequently, users can re-collect images of the same location using their mobile phones or in-store equipment to complete the authenticity verification.

[0037] In the context of tracing the origin of precision metal components, the system can generate digital fingerprints using machining patterns, micro-scratches, and surface particle distribution. Even if the component experiences slight wear during transportation and assembly, dynamic mesh fault-tolerant verification can still complete identification through undamaged areas.

[0038] In paper-based certificate or artwork management scenarios, the system can utilize fiber distribution, ink edges, and microscopic unevenness features of the paper surface for registration and comparison, reducing reliance on external labels or manual certificates.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, combinations, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for digital fingerprint recognition of objects based on physical microscopic features, characterized in that, Includes the following steps: S1. Acquire the original image of the local identification points of the object to be identified, and perform region of interest localization, color space processing and geometric normalization on the original image to obtain a standardized ROI image; S2. Input the standardized ROI image into the explicit geometric feature pathway and the implicit micro-representation pathway respectively to obtain local key point features and dense micro-texture representation. S3. The local key point features and the dense micro-texture representation are fused, encoded and reduced in dimension to generate a digital fingerprint code corresponding to the object to be identified. S4. During the registration phase, the digital fingerprint code is written into the vector database and an approximate nearest neighbor index is constructed; S5. Extract the digital fingerprint code of the object image to be tested during the recognition stage, and retrieve the Top-K candidate objects based on the approximate nearest neighbor index; S6. Perform dynamic local mesh fault-tolerant comparison and consistency verification on the image of the object to be tested and the registered image corresponding to the Top-K candidate object, and output the object identity matching result according to the verification result.

2. The method according to claim 1, characterized in that, S1 includes: converting the original image to the LAB color space, removing the L luminance channel, and retaining the A channel, B channel, and gradient direction information calculated from the original image; inputting the A channel, B channel, and gradient direction information into a spatial transformation network, estimating affine transformation parameters or perspective transformation parameters, and outputting a standardized ROI image that eliminates lighting differences and shooting pose differences.

3. The method according to claim 1, characterized in that, The explicit geometric feature path uses scale-invariant feature transformation, accelerated robust features, ORB features, or a combination thereof to extract local keypoints, orientations, and descriptors, forming a keypoint matrix with scale invariance and rotation invariance.

4. The method according to claim 1, characterized in that, The implicit micro-representation pathway uses a self-supervised visual model to extract dense micro-texture representations. The self-supervised visual model includes a DINOv2 backbone network, a mask image modeling network, a contrastive learning visual network, or a combination thereof, and is fine-tuned in the neighborhood through nearest neighbor constraints of the same object across illumination samples and hard negative sample constraints of different objects.

5. The method according to claim 4, characterized in that, The domain fine-tuning employs a composite loss function, which includes cosine nearest neighbor loss and triplet hard sample mining loss. The cosine nearest neighbor loss is used to reduce the feature distance of the same object under different shooting conditions, while the triplet hard sample mining loss is used to increase the feature distance between the micro-textures of different objects.

6. The method according to claim 1, characterized in that, S3 includes: calculating texture richness based on the gray-level co-occurrence matrix of the standardized ROI image, and assigning dynamic fusion weights to explicit geometric feature pathways and implicit micro-representation pathways based on the texture richness; increasing the fusion weight of implicit micro-representation pathways when the texture richness is lower than a preset threshold; and increasing the fusion weight of dense micro-texture representation when the number of geometric key points is lower than a preset threshold.

7. The method according to claim 1, characterized in that, The vector database in S4 uses a hierarchical navigation small-world graph algorithm to construct a multi-layer graph structure index, and associates each digital fingerprint code with item identifier, identification point location, acquisition time, acquisition device, registered ROI image, and quality score metadata.

8. The method according to claim 1, characterized in that, S6 includes: constructing a local grid based on the normalized ROI center point, dividing the ROI image to be tested and the candidate ROI image into multiple corresponding sub-grids; calculating the number of local descriptor matches, local Hamming distance, cosine distance or a combination thereof for each corresponding sub-grid; when the proportion of effective matching sub-grids that meet the preset distance threshold is greater than the preset proportion threshold, determining that the object to be tested and the candidate object are successfully matched.

9. The method according to claim 8, characterized in that, The dynamic local mesh fault-tolerant comparison also includes generating masks for local contamination, reflection, scratches, wear or occlusion areas, and removing the sub-mesh corresponding to the mask during voting statistics, so that the feature comparison of undamaged areas is not affected by damaged areas.

10. The method according to claim 1, characterized in that, When the highest similarity is below the absolute confidence threshold but above the rejection threshold, or when the proportion of effective matching subgrids is in the fuzzy critical range, a verification mark is generated, and a visual verification result containing information on the difference region, local grid scores, and candidate object ranking is output.

11. The method according to claim 1, characterized in that, The method also includes a lightweight deployment step on the edge: performing activation value-aware weight quantization on the self-supervised visual model in the implicit micro-representation pathway, and using a pagination attention mechanism or feature caching mechanism to manage the dynamic caching during the inference process, so that the digital fingerprint extraction, retrieval and comparison process can be executed on the edge of mobile terminals, handheld identification devices or industrial cameras.

12. A digital fingerprint recognition system for objects for implementing the method of any one of claims 1-11, characterized in that, include: The image acquisition and normalization module is used to acquire images of local identification points of an object and generate standardized ROI images. A dual-track feature extraction module is used to extract explicit geometric features and implicit microscopic representations in parallel; A digital fingerprint generation module is used to fuse the explicit geometric features and implicit microscopic representations to generate a digital fingerprint code; The registration and indexing module is used to write digital fingerprint codes into a vector database and build an approximate nearest neighbor index; the vector retrieval and candidate selection module is used to output Top-K candidate objects based on the digital fingerprint codes to be tested; and the grid fault tolerance verification module is used to perform dynamic local grid comparison, voting judgment, and manual review triggering. It also includes a lightweight optimization deployment module for optimizing the deployment of models and caches on the client side.

13. The system according to claim 12, characterized in that, The mesh fault tolerance verification module includes a damage detection unit, a mask generation unit, a local mesh matching unit, a voting statistics unit, and a visualization verification unit. The visualization verification unit is used to display the similarity of candidate objects, the proportion of effective sub-mesh, the location of abnormal sub-mesh, and the matching conclusion.

14. A terminal device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method according to any one of claims 1-11.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-11.

Images