A Baijiu authenticity auxiliary identification method, device, equipment and storage medium

Abstract

The present application relates to the technical field of wine bottle identity verification, and particularly relates to a baijiu authenticity auxiliary identification method, device, equipment and storage medium; standard visual angle image data sets are stored in a preset identity database to obtain a wine bottle identity database; a matching image data set is obtained from the wine bottle identity database according to a preset OCR identification method and a to-be-tested wine bottle image data set; similarity analysis is performed on the to-be-tested wine bottle image data set and the matching image data set according to a preset voting election method and a preset weighted summation formula to obtain a baijiu authenticity auxiliary identification result; through acquisition of multi-dimensional image data sets of wine bottles and preprocessing, combination of a standardized database and OCR technology to build a comparison basis, and innovation of a voting weighted identification mechanism, the baijiu authenticity identification accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of liquor bottle identification technology, and in particular to a method, device, equipment and storage medium for auxiliary identification of the authenticity of liquor. Background Technology

[0002] Currently, the identification of genuine and counterfeit liquor bottles entering the market mainly relies on the unique serial number assigned at the time of manufacture, combined with physical and digital anti-counterfeiting measures. However, facing highly sophisticated counterfeit products from a professional counterfeit industry chain, existing main anti-counterfeiting technologies have significant shortcomings: counterfeiters can not only replicate the appearance of the bottle but also forge the serial number, making it difficult to accurately distinguish genuine products using only serial number lookups and conventional image comparisons, often resulting in failed identification. Furthermore, existing technologies lack targeted auxiliary verification mechanisms to effectively supplement and verify the main anti-counterfeiting results. Using only a single identification result as the basis for judgment can easily lead to misjudgments due to the precise imitation of high-quality counterfeits. Therefore, there is an urgent need for an auxiliary identification scheme based on the unique serial number, which focuses on core differences that are difficult for counterfeits to replicate, to supplement the verification of liquor bottles entering the market, fill the identification gaps in existing technologies, and improve the overall anti-counterfeiting system's ability to distinguish counterfeits. Summary of the Invention

[0003] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a method, device, equipment and storage medium for auxiliary identification of the authenticity of liquor.

[0004] The first aspect of this invention provides a method for assisting in the identification of genuine and counterfeit liquor, comprising: acquiring a global perspective image dataset of the liquor bottle body, a multi-angle image dataset of the bottle body, a multi-angle image dataset of the bottle cap, and a local perspective image dataset; preprocessing the global perspective image dataset of the liquor bottle body, the multi-angle image dataset of the bottle body, the multi-angle image dataset of the bottle cap, and the local perspective image dataset to obtain a standard perspective image dataset; storing the standard perspective image dataset in a preset identity database to obtain a liquor bottle identity database; acquiring a liquor bottle image dataset to be tested, and retrieving a matching image dataset from the liquor bottle identity database according to a preset OCR recognition method and the liquor bottle image dataset to be tested; and performing a similarity analysis on the liquor bottle image dataset to be tested and the matching image dataset according to a preset voting method and a preset weighted summation formula to obtain an auxiliary determination result of the authenticity of the liquor.

[0005] Furthermore, the step of retrieving a matching image dataset from the bottle identity database based on a preset OCR recognition method and a sample bottle image dataset includes: preprocessing the sample bottle image dataset to obtain a standard sample image dataset; recognizing the standard sample image dataset using the OCR recognition method to obtain the factory serial number; searching the bottle identity database based on the factory serial number to obtain the search result; and retrieving the matching image dataset from the bottle identity database based on the factory serial number when the search result is successful.

[0006] Further, the step of performing similarity analysis on the test wine bottle image dataset and the matching image dataset according to a preset voting method and a preset weighted summation formula to obtain the auxiliary judgment result of the authenticity of the liquor includes: performing two-dimensional feature extraction on the matching image dataset to obtain a first prior feature; constructing a wine bottle digital identity model according to a preset unsupervised learning model and a preset deep convolutional neural network model; performing feature extraction on the matching image dataset according to the wine bottle digital identity model and the first prior feature to obtain a first deep network extracted feature and a first three-dimensional reconstructed feature; performing two-dimensional feature extraction on the standard test image dataset to obtain a second prior feature; and performing feature extraction on the standard test image dataset according to the wine bottle digital identity model and the second prior feature to obtain a second deep network extracted feature. The system extracts features and second / third-dimensional reconstructed features; it performs similarity analysis on the first prior feature, second prior feature, first deep network extracted feature, second deep network extracted feature, second / third-dimensional reconstructed feature, and first / third-dimensional reconstructed feature to obtain prior feature similarity sequences, deep network feature similarity sequences, and reconstructed feature similarity sequences; it performs fast hashing operations on the matched image dataset and the standard test image dataset to obtain average hash similarity sequences, difference hash similarity sequences, and perceptual hash similarity sequences; it analyzes the average hash similarity sequences, difference hash similarity sequences, perceptual hash similarity sequences, prior feature similarity sequences, deep network feature similarity sequences, and reconstructed feature similarity sequences according to the voting method and weighted summation formula to obtain auxiliary judgment results for the authenticity of baijiu.

[0007] Further, the step of performing fast hashing operations on the matched image dataset and the standard test image dataset to obtain an average hash similarity sequence, a difference hash similarity sequence, and a perceptual hash similarity sequence includes: performing fast hashing operations on the matched image dataset to obtain a first average hash sequence, a first difference hash sequence, and a first perceptual hash sequence; performing fast hashing operations on the standard test image dataset to obtain a second average hash sequence, a second difference hash sequence, and a second perceptual hash sequence; performing similarity analysis on the first average hash sequence and the second average hash sequence to obtain an average hash similarity sequence; performing similarity calculation on the first difference hash sequence and the second difference hash sequence to obtain a difference hash similarity sequence; and performing similarity analysis on the first perceptual hash sequence and the second perceptual hash sequence to obtain a perceptual hash similarity sequence.

[0008] Further, the step of analyzing the average hash similarity sequence, differential hash similarity sequence, perceptual hash similarity sequence, prior feature similarity sequence, deep network feature similarity sequence, and reconstructed feature similarity sequence according to the voting method and weighted summation formula to obtain the auxiliary judgment result of the authenticity of liquor includes: performing similarity analysis on the standard test image dataset and the matching image dataset according to a preset structural similarity analysis method to obtain a structural similarity sequence; performing similarity analysis on the standard test image dataset and the matching image dataset according to a preset signal-to-noise ratio analysis method to obtain a peak signal-to-noise ratio similarity sequence; and performing similarity analysis on the standard test image dataset and the matching image dataset according to a preset matcher to obtain a peak signal-to-noise ratio similarity sequence. The matching similarity is obtained; the average hash similarity sequence, difference hash similarity sequence, perceptual hash similarity sequence, matching similarity, structural similarity sequence, peak signal-to-noise ratio similarity sequence, prior feature similarity sequence, deep network feature similarity sequence, and reconstructed feature similarity sequence are calculated according to the weighted summation formula to obtain the total feature matching similarity value; the average hash similarity sequence, difference hash similarity sequence, perceptual hash similarity sequence, matching similarity, structural similarity sequence, peak signal-to-noise ratio similarity sequence, prior feature similarity sequence, deep network feature similarity sequence, and reconstructed feature similarity sequence are analyzed according to the voting method and the total feature matching similarity value to obtain the auxiliary judgment result of the authenticity of liquor.

[0009] Further, the step of performing similarity analysis on the standard test image dataset and the matched image dataset according to a preset matcher to obtain the matching similarity includes: performing feature analysis on the matched image dataset according to a preset SIFT algorithm to obtain a first feature point descriptor; performing feature analysis on the standard test image dataset according to the SIFT algorithm to obtain a second feature point descriptor; performing a comparison analysis on the first feature point descriptor and the second feature point descriptor according to the matcher to obtain a total number of matching pairs and a comparison analysis result; obtaining the number of correct matching pairs from the total number of matching pairs based on the comparison analysis result; and calculating the ratio between the number of correct matching pairs and the total number of matching pairs to obtain the matching similarity.

[0010] Further, the step of performing similarity analysis on the first prior feature, the second prior feature, the first deep network extracted feature, the second deep network extracted feature, the second 3D reconstructed feature, and the first 3D reconstructed feature to obtain a prior feature similarity sequence, a deep network feature similarity sequence, and a reconstructed feature similarity sequence includes: performing similarity analysis on the first prior feature and the second prior feature to obtain a prior feature similarity sequence; performing similarity analysis on the first deep network extracted feature and the second deep network extracted feature to obtain a deep network feature similarity sequence; and performing similarity analysis on the first 3D reconstructed feature and the second 3D reconstructed feature to obtain a reconstructed feature similarity sequence.

[0011] Furthermore, a device for assisting in the identification of genuine and counterfeit liquor includes: a first data acquisition module for acquiring a global view image dataset of the liquor bottle body, a multi-angle image dataset of the bottle body, a multi-angle image dataset of the bottle cap, and a local view image dataset; a preprocessing module for preprocessing the global view image dataset of the liquor bottle body, the multi-angle image dataset of the bottle body, the multi-angle image dataset of the bottle cap, and the local view image dataset to obtain a standard view image dataset; a data storage module for storing the standard view image dataset into a preset identity database to obtain a liquor bottle identity database; a data retrieval module for acquiring a liquor bottle image dataset to be tested and retrieving a matching image dataset from the liquor bottle identity database according to a preset OCR recognition method and the liquor bottle image dataset to be tested; and a similarity analysis module for performing similarity analysis on the liquor bottle image dataset to be tested and the matching image dataset according to a preset voting method and a preset weighted summation formula to obtain the auxiliary identification result of genuine and counterfeit liquor.

[0012] A second aspect of the present invention provides a device for assisting in the identification of genuine and counterfeit liquor, the device comprising: a memory and at least one processor, the memory storing instructions; the at least one processor calling the instructions in the memory to cause the computer device to execute the various steps of the method for assisting in the identification of genuine and counterfeit liquor as described above.

[0013] A third aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the above-described auxiliary method for distinguishing genuine and counterfeit liquor.

[0014] In the technical solution of this invention, by collecting a global view image dataset of the bottle body, a multi-angle image dataset of the bottle body, a multi-angle image dataset of the bottle cap, and a local view image dataset, the core features such as the appearance structure and detailed texture of the bottle are comprehensively captured. After image preprocessing, the accuracy of feature extraction is improved. A standardized identity database provides an authoritative and unified standard template for verification, ensuring data consistency and reducing comparison errors. OCR technology accurately extracts key information, realizes efficient retrieval of matching images, builds a reliable comparison foundation, avoids the drawback of easy copying of digital anti-counterfeiting labels, and innovatively adopts a voting election and weighted summation judgment mechanism. It not only highlights the core identification information through weight optimization and avoids the bias of a single algorithm, but also ensures the stability of the judgment through result conflict coordination, improves the anti-interference ability and the accuracy of auxiliary identification of the authenticity of liquor. The overall solution serves as an important supplementary verification basis for the main anti-counterfeiting result. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0016] Figure 1 This is a first flowchart of a method for assisting in the identification of genuine and counterfeit liquor provided in an embodiment of the present invention;

[0017] Figure 2 This is a second flowchart of a method for assisting in the identification of genuine and counterfeit liquor provided in an embodiment of the present invention;

[0018] Figure 3 This is a third flowchart of an auxiliary method for determining the authenticity of liquor provided in an embodiment of the present invention;

[0019] Figure 4 This is a fourth flowchart of an auxiliary method for determining the authenticity of liquor provided in an embodiment of the present invention;

[0020] Figure 5 The fifth flowchart of an auxiliary method for determining the authenticity of liquor provided in an embodiment of the present invention;

[0021] Figure 6 The sixth flowchart of an auxiliary method for determining the authenticity of liquor provided in an embodiment of the present invention;

[0022] Figure 7 The seventh flowchart of an auxiliary method for determining the authenticity of liquor provided in an embodiment of the present invention;

[0023] Figure 8 This is a schematic diagram of a device for assisting in the identification of genuine and counterfeit liquor provided in an embodiment of the present invention;

[0024] Figure 9 This is a schematic diagram of a device for assisting in the identification of genuine and counterfeit liquor, provided in an embodiment of the present invention. Detailed Implementation

[0025] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the method for assisting in the identification of genuine and counterfeit liquor in this invention includes:

[0027] 101. Obtain the global view image dataset of the bottle body, the multi-angle image dataset of the bottle body, the multi-angle image dataset of the bottle cap, and the local view image dataset;

[0028] In this embodiment, the system consists of multiple industrial area scan cameras, various lenses, multiple light sources, a turntable, a power supply, a controller, and a display screen. To extract multi-dimensional image data, multiple light sources are used to illuminate the product from different angles, combined with the assistance of the turntable, to acquire multi-dimensional RGB feature images of genuine, well-packaged liquor bottles. These images include a global view dataset of the bottle body, a multi-angle dataset of the bottle body, a multi-angle dataset of the bottle cap, and a local view image dataset. Simultaneously, multiple identity information of the genuine liquor is acquired, and all the above image datasets and identity information are entered into the database to ultimately form a liquor bottle identity database.

[0029] 102. Preprocess the global view image dataset of the bottle body, the multi-angle image dataset of the bottle body, the multi-angle image dataset of the bottle cap, and the local view image dataset to obtain the standard view image dataset.

[0030] In this embodiment, the image preprocessing operation includes performing affine transformation on all image data. The purpose of affine transformation is to level all image data, extract ROI, and denoise all image data to adjust the wine bottle to a frontal pose and ensure that the relative positions of some features are consistent. Since the wine bottle may be rotated, scaled, or translated in the image, it is necessary to normalize the wine bottle to a uniform angle and scale by using the wine bottle key point matching method to eliminate the influence of pose differences on subsequent feature extraction. The standard view image data of each genuine wine bottle is associated with a unique factory serial number.

[0031] 103. Store the standard viewpoint image dataset into a preset identity database to obtain the wine bottle identity database;

[0032] In this embodiment, the standard perspective image dataset is stored in the identity database to construct a standardized wine bottle identity database, which serves as the core data support for assisting in the verification of the authenticity of wine bottles. This provides authoritative and unified standard image templates, ensuring the efficient implementation of auxiliary steps such as OCR retrieval and multi-dimensional similarity comparison, and provides accurate supplementary verification basis for the main anti-counterfeiting results, thereby improving the overall reliability of verification and adapting to the auxiliary verification needs in various scenarios.

[0033] 104. Obtain the image dataset of the wine bottle to be tested, and retrieve the matching image dataset from the wine bottle identity database according to the preset OCR recognition method and the image dataset of the wine bottle to be tested;

[0034] In this embodiment, the standard view image data of each genuine wine bottle is associated with a unique factory serial number. The image data of the wine bottle to be tested is re-collected after it enters the market. The core of this verification is to confirm whether the factory serial number corresponding to the image data to be tested is consistent with the serial number associated with the standard view image data in the database. After accurately extracting the serial number from the image dataset of the wine bottle under test using OCR recognition technology, a matching image dataset can be efficiently retrieved from the wine bottle identity database. This matching image dataset is a specific genuine wine bottle image data corresponding to the image data of the wine bottle under test, precisely selected from the standard perspective image data stored in the wine bottle identity database based on the extracted serial number. If retrieval fails, it indicates that the serial number of the wine bottle image data under test does not match the serial number of the genuine wine bottle stored in the wine bottle identity database, which can be preliminarily determined to be counterfeit wine, providing a rapid screening basis for auxiliary verification. Based on the serial number retrieval and data retrieval process, a comparison foundation between the sample under test and the standard sample is quickly established, which not only improves the efficiency of the auxiliary verification process, but also provides authoritative data support for subsequent multi-dimensional comparisons, serving as a supplementary verification basis for the main anti-counterfeiting results, effectively enhancing the accuracy and credibility of the overall verification, and adapting to various batch detection and precise auxiliary verification scenarios.

[0035] 105. Based on the preset voting method and the preset weighted summation formula, perform similarity analysis on the test wine bottle image dataset and the matching image dataset to obtain auxiliary judgment results on the authenticity of the liquor.

[0036] In this embodiment, a similarity analysis is performed on the image dataset of the wine bottle to be tested and the matching image dataset by combining the voting method and the weighted summation formula. This efficiently outputs the auxiliary judgment result of the authenticity of the liquor. The weighted summation highlights the core identification information by optimizing the index weights and avoids single-dimensional bias. The voting mechanism coordinates the conflict between different judgment results and ensures the stability of the conclusion. It also enhances the anti-interference ability and effectively avoids the omission and misjudgment of counterfeit products, providing a reliable basis for the auxiliary identification of the authenticity of wine bottles and adapting to various actual detection scenarios.

[0037] In this embodiment, by collecting a global view image dataset of the bottle body, a multi-angle image dataset of the bottle body, a multi-angle image dataset of the bottle cap, and a local view image dataset, the core features such as the appearance structure and detailed texture of the bottle are comprehensively captured. After image preprocessing, the accuracy of feature extraction is improved. A standardized identity database provides an authoritative and unified standard template for verification, ensuring data consistency and reducing comparison errors. OCR technology accurately extracts key information, enabling efficient retrieval of matching images, building a reliable comparison foundation, and avoiding the drawbacks of easily copied digital anti-counterfeiting labels. An innovative voting and weighted summation judgment mechanism is adopted, which not only highlights the core identification information through weight optimization and avoids the bias of a single algorithm, but also ensures the stability of the judgment through result conflict coordination, improving the anti-interference ability and the accuracy of auxiliary identification of the authenticity of liquor. The overall scheme serves as an important supplementary verification basis for the main anti-counterfeiting result.

[0038] Please see Figure 2 In a second embodiment of the method for assisting in the identification of genuine and counterfeit liquor according to the present invention, step 104 includes:

[0039] 201. Preprocess the dataset of wine bottle images to be tested to obtain a standard dataset of images to be tested;

[0040] In this embodiment, the image preprocessing logic of the test bottle image dataset is consistent with the preprocessing logic of the standard view image dataset. The standardized data provides a consistent and reliable foundation for subsequent OCR serial number recognition, multi-dimensional similarity comparison and other steps, reduces errors caused by data differences, improves the overall scheme's identification accuracy, stability and process smoothness, and adapts to various actual detection scenarios.

[0041] 202. Use the OCR recognition method to identify the standard image dataset to obtain the factory serial number;

[0042] In this embodiment, the core principle of the OCR (Optical Character Recognition) method is to convert the text in an image into editable text data. First, a CNN model (such as CTPN or DBNet) is used to detect text in the image data and locate the text regions in the image. Then, the detected text lines are normalized and fed into the recognition model, which is a Convolutional Recurrent Neural Network (CRNN) model. In addition, commonly used open-source recognition models include Tesseract, EasyOCR, and PaddleOCR. This model combines CNN to extract features and RNN to learn the contextual relationships of sequences. Finally, it decodes the normalized text lines through CTC or Attention mechanisms, directly converting them into text strings, thereby achieving high-precision recognition of text in various layouts and complex scenes.

[0043] 203. Search the bottle identification database based on the manufacturing serial number to obtain the search results;

[0044] 204. If the search result is successful, the matching image dataset is retrieved from the bottle identification database based on the factory serial number.

[0045] In this embodiment, if the search result is a search failure, it indicates that the wine bottle to be detected is counterfeit. If the search is successful, all image data corresponding to the factory serial number are retrieved to obtain a matching image dataset. A search failure can directly determine that the wine bottle to be detected is counterfeit, quickly eliminating counterfeit wines with counterfeit serial numbers and significantly shortening the invalid judgment process. A search success accurately retrieves all matching image datasets corresponding to the serial number, providing standardized and authoritative genuine product image templates for subsequent multi-dimensional similarity comparisons. This approach balances judgment efficiency and accuracy, simplifies the process while providing reliable data support for subsequent identification, effectively improving the overall identification efficiency, accuracy, and authority of the solution, and adapting to the needs of batch detection and precise verification.

[0046] In this embodiment, an efficient and reliable auxiliary system for determining the authenticity of liquor is constructed to comprehensively improve identification efficiency. The OCR recognition method integrates mechanisms such as CNN text localization and RNN sequence learning to achieve high-precision extraction of the factory serial number in complex scenarios, avoiding misjudgments caused by recognition errors. The database retrieval mechanism with the factory serial number as the core can quickly determine counterfeit products when the retrieval fails, and accurately retrieve all matching images of the corresponding genuine products when the retrieval is successful, providing an authoritative standard template for subsequent auxiliary comparison. It takes into account both batch detection efficiency and accurate verification requirements, improving identification efficiency and accuracy. As a supplementary verification basis for the main anti-counterfeiting results, it provides reliable auxiliary support for the verification of the authenticity of liquor in multiple scenarios.

[0047] Please see Figure 3 In the third embodiment of the method for assisting in the identification of genuine and counterfeit liquor in this invention, step 105 includes:

[0048] 301. Perform two-dimensional feature extraction on the matching image dataset to obtain the first prior features;

[0049] In this embodiment, the first prior feature extraction refers to calculating edge features, shape features, grayscale features, texture features, and color features of the matching image dataset. Edge features are regions with abrupt changes in grayscale in the image, and extraction methods include Sobel, Canny, and Laplacian, used for contour analysis of the wine bottle image. Shape features include area, perimeter, and moments, used for matching and classifying the shape of the wine bottle, and blob analysis is used to calculate the area, perimeter, and moments of the corresponding region in the wine bottle image. Grayscale features are based on the grayscale value statistics of image pixels, including grayscale mean, variance, histogram, and entropy, used for brightness and contrast correlation analysis of the wine bottle image. Texture features include LBP, GLCM, and Gabor filtering, used to analyze the texture information of the wine bottle. Color features are calculated by calculating the color distribution, including color histogram and color moments, used to analyze the color distribution information of the wine bottle.

[0050] 302. A digital identity model for a wine bottle is constructed based on a pre-defined unsupervised learning model and a pre-defined deep convolutional neural network model.

[0051] In this embodiment, an autoencoder (unsupervised learning model) with an encoder-decoder architecture is used. The encoder consists of 3 convolutional layers (3×3 kernel size, stride 1) and 2 max pooling layers (2×2 pooling kernel). The decoder consists of 3 deconvolutional layers and 1 output layer. The input consists of 500,000 unlabeled genuine wine bottle images (covering different brands, models, shooting angles, and lighting conditions), without manual labeling of authenticity. The goal is to minimize image reconstruction error (using mean squared error (MSE) as the loss function). The autoencoder learns the general visual features (such as edges, textures, basic shapes, global contours, etc.) of the 500,000 unlabeled genuine wine bottle images. After training, the "encoder part" of the autoencoder is truncated (discarded). The encoder, which is capable of separating the foreground (the main body of the wine bottle) from the background and extracting basic visual features from complex images, reduces the difficulty and sample requirements for subsequent identification feature learning. The encoder of the autoencoder is used as the front-end feature extraction module of a deep convolutional neural network model (DCNN, such as ResNet152). The two are fixedly connected through fully connected layers, forming an integrated network architecture of the autoencoder's encoder and the DCNN main body (i.e., the wine bottle digital identity model). The DCNN (deep convolutional neural network) main structure is an integrated feature extraction architecture formed by a ResNet152 backbone network and a fixed connection between the front-end autoencoder's encoder. The overall structure is divided into three parts from top to bottom: The encoder portion of the front-end feature guidance layer autoencoder (3 convolutional layers and 2 max-pooling layers) learns general visual features of the wine bottle and fuses them with prior features, providing prior guidance for subsequent deep feature extraction. The ResNet152 backbone network layer adopts the standard ResNet152 network structure to perform deep convolutional calculations on the fused features, outputting feature maps containing high-dimensional semantic and spatial structural information. The dual-branch output layer simultaneously draws two parallel branches from the ResNet152 backbone network to extract deep network features and 3D reconstruction features. The 3D reconstruction branch network is not an independent network, but a parallel output branch of the DCNN main structure. The relationship between the two is as follows: the input of the 3D reconstruction branch network... The high-dimensional feature map is directly taken from the third-to-last convolutional layer of the DCNN main body (ResNet152). This feature map contains both spatial structure information and texture information, providing a foundation for 3D reconstruction. After completing deep convolution, the DCNN main body outputs two sets of features: one is fed into the deep feature branch, which passes through global average pooling (GAP), fully connected layers, ReLU activation, and Dropout to output the first feature extracted by the deep network; the other is fed into the 3D reconstruction branch, which passes through two convolutional layers, deconvolutional layers, point cloud generation layers, and normal vector calculation layers to output the first 3D reconstruction feature. The DCNN main body is responsible for unified feature learning and spatial information preservation, while the 3D reconstruction branch is responsible for recovering the 3D structure of the bottle from the features output by the main body.

[0052] 303. Based on the digital identity model of the wine bottle and the first prior features, feature extraction is performed on the matching image dataset to obtain the features extracted by the first deep network and the first three-dimensional reconstruction features;

[0053] In this embodiment, the standard viewpoint image dataset, after preprocessing (denoising and normalization), is input into the encoder of the autoencoder, which outputs a 64-dimensional general feature vector. The first prior features (edge ​​features, grayscale features, color moments, and other hand-designed features) are converted into a 64-dimensional general feature vector through an independent fully connected layer. This 64-dimensional feature vector is then fused element-wise with the general feature vector output by the encoder to obtain a 128-dimensional fused feature vector. The 128-dimensional fused feature vector is then processed by the main body of the DCNN (intermediate convolutional layers of ResNet152) to obtain a 2048-dimensional high-dimensional feature map (package). The feature vector, which contains general visual features and identification guidance features, is compressed into a 1024-dimensional low-dimensional feature vector through a global average pooling (GAP) layer. Then, it is processed by a fully connected layer (using ReLU activation function) and a dropout layer (dropout rate=0.2) to output the final feature extracted by the first deep network. This 128-dimensional fused feature vector condenses the abstract semantic information of the bottle (label pattern proportions, printing font details, bottle marking layout, etc.). Guided by the first prior feature, it can still accurately distinguish the foreground and background information of the bottle in low-contrast scenes.

[0054] A high-dimensional feature map (preserving image spatial structure information, with the same size as the input image) is extracted from the third-to-last convolutional layer of the DCNN main body and input into the 3D reconstruction branch network. In the 3D reconstruction branch network structure, it passes through 2 convolutional layers (enhancing spatial texture features), 1 deconvolutional layer (restoring spatial resolution), 1 point cloud generation layer, and 1 normal vector calculation layer, containing a set of "first 3D reconstruction features" consisting of point cloud coordinate features and normal vector features. The point cloud generation layer, based on the spatial feature map, uses a "pixel coordinates and 3D world coordinates" conversion algorithm (formula: X=u×Z / f, Y=v×Z / f, where u is the first pixel coordinate). (where v is the second pixel coordinate, Z is the depth value, and f is the camera focal length), generating a set of 3D coordinates (X, Y, Z) for all pixels on the surface of the bottle; the normal vector calculation layer uses the PCA algorithm to perform principal component analysis on the neighborhood point cloud of each 3D coordinate point to solve for the normal vector (nx, ny, nz) of that point, where nx is the component of the normal vector in the X-axis direction of the 3D Cartesian coordinate system, ny is the component of the normal vector in the Y-axis direction of the 3D Cartesian coordinate system, and nz is the component of the normal vector in the Z-axis direction of the 3D Cartesian coordinate system; the final output is a set of "first 3D reconstruction features" containing point cloud coordinate features and normal vector features;

[0055] 304. Perform two-dimensional feature extraction on the standard image dataset to obtain the second prior features;

[0056] In this embodiment, the extraction logic of the first prior feature is continued, and the extraction of basic interpretable features at the two-dimensional level of the image is used to obtain the second prior feature, which includes edge features (extracted by Sobel / Canny algorithm, reflecting the bottle outline and label border), shape features (area, perimeter, rectangle, etc., obtained through blob analysis, representing the overall shape of the bottle), grayscale features (grayscale mean, variance, histogram, reflecting the distribution of brightness and contrast), texture features (LBP / GLCM / Gabor filtering, capturing label printing texture and bottle cap texture), and color features (color histogram, color rectangle, reflecting the color distribution of the bottle / label). The second prior feature, as the underlying basic feature, not only provides prior knowledge guidance for subsequent deep feature extraction, but also can be directly compared with the first prior feature to quickly verify the consistency between the test image and the standard image at the basic visual level.

[0057] 305. Based on the digital identity model of the wine bottle and the second prior features, feature extraction is performed on the standard image dataset to be tested to obtain the features extracted by the second deep network and the second three-dimensional reconstruction features;

[0058] In this embodiment, the second prior feature serves as guiding information input to the model, helping the model quickly distinguish between the foreground (the main body of the wine bottle) and the background, enhancing the feature response in low-contrast scenes, and preventing the model from being interfered with by irrelevant information. The model learns the abstract semantic features of the image (second deep network features) through deep convolutional layers. These features are low-dimensional vectors that condense key discriminative information such as bottle markings, pattern details, and printing processes. At the same time, the model generates second three-dimensional reconstruction features through three-dimensional reconstruction technology, including point cloud coordinate features (representing the spatial distribution of the wine bottle surface) and normal vector features (reflecting the surface texture and process details), making up for the spatial dimension information that two-dimensional features cannot capture.

[0059] 306. Perform similarity analysis on the first prior feature, the second prior feature, the first deep network extracted feature, the second deep network extracted feature, the second 3D reconstructed feature, and the first 3D reconstructed feature to obtain the prior feature similarity sequence, the deep network feature similarity sequence, and the reconstructed feature similarity sequence;

[0060] In this embodiment, by performing similarity analysis on the first prior feature, the second prior feature, the first deep network extracted feature, the second deep network extracted feature, the second 3D reconstruction feature, and the first 3D reconstruction feature, a full-dimensional feature matching analysis system is constructed. This system accurately captures the core differences that are difficult for high-quality counterfeit products to replicate, such as the printing of wine bottle labels, texture details, and spatial craftsmanship. This improves the accuracy and anti-interference of the identification of genuine and counterfeit products, and provides reliable technical support for the efficient anti-counterfeiting of liquor.

[0061] 307. Perform fast hashing operations on the matching image dataset and the standard test image dataset to obtain the average hash similarity sequence, the difference hash similarity sequence, and the perceptual hash similarity sequence;

[0062] In this embodiment, average hashing efficiently concentrates core visual features and resists light and noise interference; difference hashing is sensitive to subtle differences such as bottle texture and label printing; perception hashing focuses on image structure information, resists size scaling and slight deformation, and can quickly capture feature differences that are difficult for counterfeit products to replicate, providing a comprehensive and objective basis for assisting in the determination of the authenticity of liquor, and improving identification efficiency, accuracy and scene adaptability.

[0063] 308. Based on the voting method and weighted summation formula, the average hash similarity sequence, differential hash similarity sequence, perceptual hash similarity sequence, prior feature similarity sequence, deep network feature similarity sequence, and reconstructed feature similarity sequence are analyzed to obtain auxiliary judgment results for the authenticity of liquor.

[0064] In this embodiment, a weighted summation formula is used to perform similarity analysis on average hash similarity, differential hash similarity, perceptual hash similarity, prior feature similarity sequence, deep network feature similarity sequence, and reconstructed feature similarity sequence to generate an accurate total feature matching similarity value. A voting mechanism is used to coordinate conflicts between multiple methods, and the threshold of the objecting method is adjusted in reverse to improve consistency. This voting mechanism sets independent thresholds for each similarity calculation method, and preliminary authenticity verification is completed through threshold comparison. When results conflict, a majority vote is used to finalize the conclusion, and the threshold of the objecting method is adjusted in reverse. This effectively avoids bias in single-method judgments, coordinates multi-dimensional judgment conflicts, improves overall consistency and stability, reduces the risk of missed or incorrect judgments, and enhances the accuracy and anti-interference capability of auxiliary identification of genuine and counterfeit liquor, adapting to the needs of industrialized batch testing and market supervision.

[0065] In this embodiment, the prior features encompass two-dimensional basic features such as edges, shape, and grayscale. Utilizing specialized algorithms, the macroscopic shape and basic surface attributes of the wine bottle are comprehensively captured, providing rich underlying information. The wine bottle digital identity model, built upon unsupervised learning (autoencoders) and deep convolutional neural networks (such as DCNN), accelerates training convergence and improves generalization ability under the guidance of the first prior features. It also enhances feature response in low-contrast scenes, accurately distinguishes foreground from background, and simultaneously extracts low-dimensional deep network features, condensing abstract semantic differences such as label proportions and font details. The three-dimensional reconstruction features capture spatial features such as point cloud coordinates and normal vectors. The process information covers all dimensions of identification points. The overall solution achieves comprehensive, accurate, and efficient feature extraction, providing a highly recognizable standard template for subsequent similarity comparison, improving the accuracy and anti-interference ability of authenticity-assisted judgment, and adapting to various practical detection scenarios. It generates average hash similarity sequences, difference hash similarity sequences, and perceptual hash similarity sequences through three types of hash operations. Average hash is resistant to illumination interference, difference hash is sensitive to subtle differences, and perceptual hash is protected against deformation effects, quickly completing the initial screening and providing objective evidence. Simultaneously, it extracts first prior features, features extracted by the first deep network, features from the first 3D reconstruction, and features from the second deep network. The system extracts features from a network and reconstructs second and third-dimensional features. Prior features solidify the two-dimensional foundation, deep network features condense abstract semantics, and three-dimensional reconstruction features supplement spatial process details. Through feature comparison, it accurately captures the core differences that are difficult to replicate in high-quality counterfeits. Finally, relying on a weighted summation formula and voting method, it analyzes and obtains auxiliary judgment results for the authenticity of liquor, improving the accuracy, anti-interference, and consistency of liquor auxiliary identification, and adapting to the diverse needs of industrial batch testing and market supervision. Average hashing efficiently condenses core visual features and resists light and noise interference; difference hashing is sensitive to subtle differences such as bottle texture and label printing; and perceptual hashing focuses on image structure information. This system resists size scaling and slight deformation, operates quickly, and can capture flaws that are difficult for counterfeits to replicate, providing an objective basis for judgment. It extracts features from all dimensions to generate prior features (which can explain two-dimensional basic information), deep network features (abstract semantic core differences), and three-dimensional reconstruction features (including spatial process details such as point cloud coordinates and normal vectors), covering key identification points of the wine bottle and improving feature recognition. It accelerates the initial screening through hashing and reduces invalid calculations, while supporting accurate comparison with high-quality features, improving identification efficiency, accuracy, and anti-interference, effectively distinguishing subtle differences between high-quality and counterfeit products, and adapting to various practical detection scenarios.

[0066] Please see Figure 4 In the fourth embodiment of the method for assisting in the identification of genuine and counterfeit liquor in this invention, step 307 includes:

[0067] 401. Perform a fast hash operation on the matching image dataset to obtain a first average hash sequence, a first difference hash sequence, and a first perceptual hash sequence;

[0068] 402. Perform fast hashing on the standard image dataset to be tested to obtain the second average hash sequence, the second difference hash sequence, and the second perceptual hash sequence;

[0069] 403. Perform similarity analysis on the first average hash sequence and the second average hash sequence to obtain the average hash similarity sequence;

[0070] In this embodiment, the steps for analyzing average hash similarity are as follows: First, all images in the matching image dataset and all standard test images in the standard test image dataset are reduced to an 8×8 size. Then, the average gray value of each of these two types of images is calculated. Subsequently, for each pixel of each type of image, its gray value is compared with the average gray value of the corresponding image (greater than or equal to the average gray value is recorded as 1, less than is recorded as 0), thereby generating a 64-bit hash value (i.e., the first average hash). Finally, the first average hash sequence of the matching image dataset (obtained by integrating all the first average hashes calculated from the matching image dataset) and the second average hash sequence of the standard test image dataset (obtained by integrating all the second average hashes calculated from the standard test image dataset) are obtained. Commonly used image reduction methods include nearest neighbor interpolation and bilinear interpolation. The average hash algorithm is used to analyze image similarity. Through 8×8 size scaling, average gray value calculation, and 64-bit hash value generation, the core features of the wine bottle image are concisely and efficiently condensed. It simplifies the image data while preserving the image through the comparison logic between pixels and average gray values. Retaining key visual information effectively resists interference from slight shooting noise and lighting changes. Similarity comparison between the first and second average hash sequences can quickly capture differences in the bottle's basic features, providing an efficient and low-cost basis for assisting in the determination of the authenticity of liquor, improving identification efficiency and scene adaptability. For example, when using bilinear interpolation to reduce the size of an image, the operation logic is as follows: establish a coordinate mapping relationship between the target image and the original image (calculate the theoretical position of the target pixel in the original image based on the reduction ratio, usually non-integer coordinates); round the theoretical mapped coordinates (e.g., rounding to the nearest integer), locate the nearest single original pixel, and directly assign the value of that original pixel to the target pixel, completing the reduction process; calculate the Hamming distance between the hash values ​​corresponding to the positions in the first and second average hash sequences, and normalize the Hamming distance to a similarity value (normalization formula: similarity = 1 − Hamming distance / 64); the Hamming distance between two binary hash sequences of the same length refers to the number of bits that are different at the same position in these two binary sequences.

[0071] 404. Calculate the similarity between the first and second differential hash sequences to obtain a differential hash similarity sequence;

[0072] In this embodiment, the steps for analyzing differential hash similarity are as follows: First, all images in the matching image dataset and all standard test images in the standard test image dataset are reduced to a size of 8×9 (8 rows and 9 columns). Then, for each row of pixels in each type of image, the grayscale values ​​of adjacent pixels in that row are compared sequentially (a pixel with a grayscale value greater than that of the pixel with a grayscale value greater than that of the pixel with a grayscale value greater than that of the pixel with a grayscale value greater than that of the pixel with a grayscale value less ... The second difference hash is integrated, and the difference hash similarity is obtained through comparative analysis. A 64-bit hash value (i.e., the second difference hash) is generated through steps such as 8×9 size scaling and grayscale comparison of adjacent pixels. It focuses on the grayscale change pattern of image pixels. It is sensitive to subtle differences in the wine bottle image and can accurately capture the flaws of counterfeit products in label printing, bottle texture, etc. It is also resistant to slight shooting interference. The similarity comparison between the first difference hash data and the second difference hash data provides an efficient basis for assisting in the determination of the authenticity of liquor, improving the accuracy of identification and scene adaptability. For the hash values ​​corresponding to the positions in the first difference hash sequence and the second difference hash sequence, the Hamming distance is calculated respectively, and the Hamming distance is normalized to the similarity value (normalization formula: similarity = 1 − Hamming distance / 64). The Hamming distance between two binary hash sequences of the same length refers to the number of bits that are different at the same position in the two binary sequences.

[0073] 405. Perform similarity analysis on the first perceptual hash sequence and the second perceptual hash sequence to obtain the perceptual hash similarity sequence;

[0074] In this embodiment, the implementation steps of perceptual hash similarity analysis are as follows: First, all images in the matching image dataset and all standard test images in the standard test image dataset are reduced to a size of 32×32. Then, Discrete Cosine Transform (DCT) is performed on each type of reduced image, and the DCT coefficients of the 8×8 region in the upper left corner of the transformed image are extracted. Next, the median of all DCT coefficients within this 8×8 region is calculated. Then, for each DCT coefficient within the region, its relationship with the median is compared (greater than or equal to the median is recorded as 1, otherwise as 0), thereby generating a 64-bit hash value (i.e., the first perceptual hash). Finally, the first perceptual hash sequence of the matching image dataset (obtained by integrating all first perceptual hashes calculated from the matching image dataset) and the second perceptual hash sequence of the standard test image dataset (obtained by integrating all second perceptual hashes calculated from the standard test image dataset) are obtained. Comparative analysis yields a perceptual hash similarity sequence (integrated from all compared perceptual hash similarities). A 64-bit hash value (i.e., the second perceptual hash) is generated through steps such as size standardization and DCT transformation. This process focuses on key structural information of the image and effectively resists interference from size scaling and slight deformation. The generated first and second perceptual hash sequences accurately condense the features of the wine bottle image, quickly capturing subtle differences in bottle details through similarity comparison. This provides an efficient and reliable basis for assisting in the determination of the authenticity of liquor, improving identification accuracy and adaptability. Hamming distance is calculated for the hash values ​​corresponding to positions in the first and second perceptual hash sequences, and the Hamming distance is normalized to a similarity value (normalization formula: similarity = 1 − Hamming distance / 64). The Hamming distance between two binary hash sequences of the same length refers to the number of bits that differ at the same position in these two binary sequences.

[0075] In this embodiment, the average hash, through 8×8 scaling and comparison with the average grayscale, concisely condenses the core visual features and resists interference from lighting noise; the difference hash, through 8×9 scaling, captures the grayscale changes of adjacent pixels and is sensitive to subtle differences such as label printing and bottle texture; the perceptual hash, through 32×32 scaling and DCT transformation, focuses on the key structure of the image and resists the effects of scaling and slight deformation; all three algorithms generate 64-bit hash values, balancing feature extraction efficiency and accuracy. By comparing the similarity of pairwise hash data, it can quickly capture details and flaws that are difficult for counterfeiters to replicate. The comparison process is efficient and low-cost, adaptable to various shooting scene interferences, and provides multi-dimensional objective evidence for the auxiliary identification of genuine and counterfeit liquor, improving the efficiency, accuracy and scene adaptability of the judgment.

[0076] Please see Figure 5 The fifth embodiment of the method for assisting in the identification of genuine and counterfeit liquor in this invention includes step 308:

[0077] 501. Perform similarity analysis on the standard test image dataset and the matching image dataset according to the preset structural similarity analysis method to obtain the structural similarity sequence;

[0078] In this embodiment, the signal-to-noise ratio analysis method uses the structural similarity calculation formula as the core analysis tool. The expression of the structural similarity calculation formula is as follows:

[0079]

[0080] In the formula, To match the structural similarity between an image in the image dataset (denoted as image a) and an image in the standard test image dataset (denoted as image b), , Let be the local mean of image a and image b. , The local standard deviations of the two images are given. For local covariance, The coefficient of the first stability, The coefficient is the second most stable factor to prevent the denominator from being zero; usually... , , The maximum value of each pixel is used to integrate all structural similarity scores, ultimately forming a structural similarity sequence. The structural similarity calculation formula (SSIM) is used to analyze image similarity. By introducing C1 and C2 stability coefficients to avoid calculation anomalies, and combining the pixel maximum value normalization parameter, the structural details of bottle labels, patterns and other differences are accurately captured. The results are consistent with the logic of human visual perception. The integrated structural similarity sequence provides a reliable basis for assisting in the determination of the authenticity of liquor, effectively improving the accuracy and rationality of identification.

[0081] 502. Perform similarity analysis on the standard test image dataset and the matched image dataset according to the preset signal-to-noise ratio analysis method to obtain the peak signal-to-noise ratio similarity sequence;

[0082] In this embodiment, the signal-to-noise ratio (SNR) analysis method uses the peak signal-to-noise ratio (PSNR) similarity calculation formula as its core analysis tool. The expression for the PNR similarity calculation formula is as follows:

[0083]

[0084] In the formula, For peak signal-to-noise ratio similarity. The maximum pixel value of the image. Mean square error,

[0085]

[0086] In the formula, m and n are the width and height of the image, I is the image in the matching image dataset, D is the image in the standard test image dataset, and i and j are the two-dimensional coordinate indices of the image pixels, used to locate the position of a specific pixel in the image: i represents the pixel number in the width direction of the image, with a value range from 0 to m-1; j represents the pixel number in the height direction of the image, with a value range from 0 to m-1. Up to n-1, all peak signal-to-noise ratio (PSNR) similarities are integrated to form a final PSNR similarity sequence. By analyzing the PSNR similarity between images in the matching image dataset and images in the standard test image dataset, the pixel-level error between the matching image data and the standard test image is first accurately quantified using MSE. Combined with parameters such as image width and height, and maximum pixel value, the error is transformed into an intuitive signal-to-noise ratio index using the PSNR formula. This index can capture subtle differences that are difficult to replicate in counterfeit products, such as printing ink marks and texture details. The integrated PSNR similarity sequence provides an objective quantitative basis for assisting in the determination of the authenticity of liquor, effectively improving the accuracy and reliability of identification and adapting to actual detection needs. Finally, the PSNR similarity sequence is obtained by integrating all the PSNR similarities.

[0087] 503. Perform similarity analysis on the standard test image dataset and the matching image dataset according to the preset matcher to obtain the matching similarity;

[0088] In this embodiment, the FLANN matcher is combined to efficiently complete the feature comparison between the standard test image dataset and the matching image dataset, generate matching similarity, accurately capture the microscopic differences of the wine bottle, effectively resist the interference of collection, improve the accuracy of auxiliary identification of the authenticity of liquor, adapt to industrial batch detection, and provide a reliable basis for comprehensive judgment.

[0089] 504. Calculate the average hash similarity sequence, differential hash similarity sequence, perceptual hash similarity sequence, matching similarity, structural similarity sequence, peak signal-to-noise ratio similarity sequence, prior feature similarity sequence, deep network feature similarity sequence, and reconstructed feature similarity sequence according to the weighted summation formula to obtain the total feature matching similarity value.

[0090] In this embodiment, in order to obtain the optimal weight combination, such that the predicted value... As close as possible to the true value The core model used to solve for the optimal weight combination is a weighted multi-feature fusion linear regression model (belonging to the category of supervised regression modeling). This weighted multi-feature fusion linear regression model is specifically optimized and trained for the weights of three types of similarity features: hash operation, multi-dimensional algorithm analysis, and feature hierarchical comparison. The specific process is as follows:

[0091] During the model training phase, the core objective is to find the optimal weight combination for each similarity method. This allows us to use the weighted summation formula. The total feature matching similarity value is calculated, where m represents the number of m methods to calculate the similarity (in this embodiment, there are 9 specific similarity calculation methods, and the calculated similarities include average hash similarity sequence, difference hash similarity sequence, perceptual hash similarity sequence, matching similarity, structural similarity sequence, signal-to-noise ratio similarity sequence, prior feature similarity sequence, deep network feature similarity sequence, and reconstructed feature similarity sequence), i represents the sample number of the i-th similarity of the m-th method (i=1,2,...,z), and z is the last sample sequence number; the similarity value calculated by the k-th method is... , Let k be the weight of the k-th method. For the i-th sample, the predicted value and the actual value are... The similarity values ​​calculated using the above methods are weighted to obtain the final total feature matching similarity value, which is then used to calculate the predicted value before training the model. (The predicted value for the i-th sample) should be as close as possible to the true value. To accurately quantify the deviation between predicted and actual values, the mean squared error (MSE) is used as the loss function for this linear regression model, expressed as follows:

[0092]

[0093] By minimizing the loss function The weights can be solved. The weights are non-negative, normalized, and sum to 1. The weights of each similarity method are optimized through the mean squared error loss function, and the weighted average is used to obtain the accurate total feature matching similarity value. Combined with the voting mechanism, when multiple methods determine conflict, the majority result is taken as the final conclusion. At the same time, the threshold of the objection method is adjusted in reverse to improve subsequent consistency. It integrates three types of similarity information: hash, algorithm, and feature, avoids the limitations of a single method, improves the accuracy and reliability of auxiliary determination of the authenticity of liquor, and adapts to actual detection needs.

[0094] 505. Based on the voting method and the total feature matching similarity value, the average hash similarity sequence, differential hash similarity sequence, perceptual hash similarity sequence, matching similarity, structural similarity sequence, peak signal-to-noise ratio similarity sequence, prior feature similarity sequence, deep network feature similarity sequence, and reconstructed feature similarity sequence are analyzed to obtain auxiliary judgment results for the authenticity of liquor.

[0095] In this embodiment, the mean of all average hash similarities in the average hash similarity sequence is calculated to obtain a first average similarity value; the mean of all differential hash similarities in the differential hash similarity sequence is calculated to obtain a second average similarity value; the mean of all perceptual hash similarities in the perceptual hash similarity sequence is calculated to obtain a third average similarity value; the mean of all structural similarities in the structural similarity sequence is calculated to obtain a fourth average similarity value; the mean of all peak signal-to-noise ratio (PSNR) similarities in the peak signal-to-noise ratio (PSNR) similarity sequence is calculated to obtain a fifth average similarity value; and the mean of all prior feature similarities in the prior feature similarity sequence is calculated to obtain a sixth average similarity value; the deep network feature similarity... The average similarity of all deep network features in the similarity sequence is calculated to obtain the seventh average similarity value; the average similarity of all reconstructed features in the reconstructed feature similarity sequence is calculated to obtain the eighth average similarity value; by calculating the arithmetic mean of the eight similarity sequences, the numerical fluctuations caused by single images, single viewpoints or local features can be effectively smoothed, abnormal interference can be eliminated, and the judgment criteria of each algorithm can be made more stable and representative. The mean can objectively reflect the overall matching level of each type of similarity in the global range, and the sequence features are transformed into a unified and comparable scalar, providing a stable and fair input for subsequent voting judgment. This method can improve the anti-interference ability of the system in complex shooting environments and enhance the reliability and consistency of the true and false judgment results.

[0096] First, an independent threshold is set for each similarity calculation method. For example, the first method is the average hash similarity method, with a hash similarity threshold 'a'; the second method is the difference hash similarity method, with a threshold 'b'; and the third method is the perceptual hash similarity method, with a threshold 'c'. The remaining methods follow this rule and each has its corresponding threshold. For the same wine bottle to be detected, the mean similarity sequence of each method is calculated for the average hash similarity method, difference hash similarity method, perceptual hash similarity method, structural similarity method, signal-to-noise ratio similarity method, prior feature similarity method, deep network feature similarity method, and reconstructed feature similarity method, resulting in the first, second, third, fourth, fifth, sixth, seventh, and eighth average similarity values. Then, the average similarity value of each method is compared with its corresponding threshold. Only when the average similarity value matches the threshold set by the method is the result of a comparison. If the similarity value is greater than its corresponding set threshold (e.g., the first average similarity value is greater than the hash similarity threshold 'a' set by the average hash similarity method), then the average similarity value can be adopted. This is done to avoid the calculated similarity value deviating too much from the total feature matching similarity value, resulting in inaccurate judgment results. Then, the total feature matching similarity value is compared with the average similarity value that can be adopted by each method after the above screening. If a certain average similarity value is greater than the total feature matching similarity value (e.g., the first average similarity value is greater than the total feature matching similarity value), then the judgment result corresponding to that method is "true"; otherwise, it is "false". For matching similarity, it is directly compared with the total feature matching similarity value. If the matching similarity is greater than the total feature matching similarity value, then the judgment result corresponding to that method is "true"; otherwise, it is "false". If the judgment results of multiple methods are inconsistent, a voting mechanism is used to select the result with the majority consensus as the final auxiliary judgment conclusion for the authenticity of the liquor.

[0097] In this embodiment, by integrating three categories and nine similarity indicators—hash operations, multi-dimensional algorithm analysis (SIFT to resist pose interference, SSIM to fit human eye perception, PSNR to quantify pixel error), and feature-level comparison (prior, deep network, 3D reconstruction features)—it comprehensively covers the macroscopic features, mesoscopic structure, microscopic pixels, and spatial process details of the bottle, accurately capturing subtle differences that are difficult for counterfeiters to replicate. A mean squared error loss function is used to optimize the weights of each method, generating a weighted and accurate total feature matching similarity value, avoiding misjudgments caused by the limitations of a single method. Coupled with a voting mechanism, in the event of conflict among multiple methods, the majority consensus result is used as the final conclusion. The overall solution combines comprehensiveness, accuracy, and anti-interference, effectively resisting interference from shooting posture, lighting changes, and other scene interferences, reducing the risk of missed or misjudged high-quality counterfeits, providing reliable technical support for the auxiliary identification of genuine and counterfeit liquor, and adapting to various practical application scenarios such as industrial testing and market supervision.

[0098] Please see Figure 6 In the sixth embodiment of the method for assisting in the identification of genuine and counterfeit liquor in this invention, step 503 includes:

[0099] 601. Perform feature analysis on the matching image dataset according to the preset SIFT algorithm to obtain the first feature point descriptor;

[0100] In this embodiment, a multi-scale space is constructed using the SIFT algorithm, and stable key points with identification value are detected and accurately located. A principal direction is assigned to them to ensure rotation invariance. Then, the gradient features of the neighborhood of the key points are extracted to generate a 128-dimensional standardized first feature point descriptor. This process fully captures the macroscopic structure and microscopic details of the genuine wine bottle, and also has scale and illumination robustness, providing an authoritative and stable genuine benchmark for subsequent comparison with the second feature point descriptor of the standard test image dataset.

[0101] 602. Perform feature analysis on the standard image dataset to be tested using the SIFT algorithm to obtain the second feature point descriptor;

[0102] In this embodiment, a multi-scale image space is constructed using the SIFT algorithm, taking a standard image dataset as input. Local extrema with discriminative value are detected and screened, and key points are accurately located after removing low-contrast pseudo-feature points. A principal direction is assigned to the key points to ensure rotation invariance, and then the gradient features of the key point neighborhood are extracted to generate a 128-dimensional standardized second feature point descriptor. This process captures the macroscopic structure and microscopic details of the standard image dataset, and has scale and illumination robustness, laying the foundation for subsequent comparison with the first feature point descriptor of the matching image dataset.

[0103] 603. Based on the matcher, perform a comparison analysis on the first feature point descriptor and the second feature point descriptor to obtain the total number of matching pairs and the comparison analysis results;

[0104] In this embodiment, the FLANN (Fast Nearest Neighbor Search Library) matcher is used to compare the similarity between the first feature point descriptor and the second feature point descriptor, and to calculate the number of matching pairs between the two feature points (including the number of correct matching pairs and the number of incorrect matching pairs).

[0105] 604. Based on the comparative analysis results, obtain the number of correct matching pairs from the total number of matching pairs;

[0106] In this embodiment, the comparison analysis result is a set of matching quantification information output by the FLANN matcher, including the similarity distance value of each pair of feature points, the coordinate mapping relationship of the first feature point descriptor and the second feature point descriptor, and the preliminary classification label of matching quality based on the K nearest neighbor distance ratio. This provides a quantitative basis for screening the number of correct matching pairs. Based on the K nearest neighbor distance ratio in the comparison analysis result, an empirical threshold (such as ratio < 0.7) is set to screen out high-confidence matching pairs with a ratio lower than the threshold, and incorrect matching pairs that do not meet the threshold are removed. Finally, the number of high-confidence matching pairs is counted, which is the number of correct matching pairs.

[0107] 605. Calculate the ratio of the number of correct matches to the total number of matches to obtain the matching similarity.

[0108] In this embodiment, the SIFT algorithm is used to extract 128-dimensional feature point descriptors from the matching image dataset and the standard test image dataset, respectively. This fully captures the macroscopic structure and microscopic details of the wine bottle, and is robust to scale, rotation, and illumination, effectively resisting problems such as acquisition angle deviation and light and shadow interference. It provides an authoritative and stable genuine benchmark and accurate reference for the test features for feature comparison. Combined with the FLANN matcher, it realizes fast similarity comparison of high-dimensional descriptors, accurately distinguishes the number of correct and incorrect matching pairs, and then generates a standardized matching similarity through ratio calculation. The quantitative results can be seamlessly integrated into a multi-dimensional comprehensive judgment framework. This solution improves the accuracy of identifying counterfeit wine bottles, reduces the risk of missed or false judgments, and is suitable for industrial batch detection scenarios, balancing judgment efficiency and reliability, providing strong technical support for the auxiliary identification of genuine and counterfeit liquor.

[0109] Please see Figure 7 The seventh embodiment of the method for assisting in the identification of genuine and counterfeit liquor in this invention includes step 306:

[0110] 701. Perform similarity analysis on the first prior feature and the second prior feature to obtain the prior feature similarity sequence;

[0111] In this embodiment, the consistency of the test bottle and the standard bottle in basic visual attributes is verified by calculating the distance between two types of two-dimensional basic features (such as Euclidean distance and cosine similarity), thereby analyzing and obtaining the prior feature similarity sequence.

[0112] 702. Perform similarity analysis on the features extracted by the first deep network and the features extracted by the second deep network to obtain the deep network feature similarity sequence;

[0113] In this embodiment, based on the distance metric of deep feature vectors, differences at the abstract semantic level (such as subtle deformations in counterfeit label patterns and differences in printed fonts) are captured, thereby analyzing and obtaining a deep network feature similarity sequence.

[0114] 703. Perform similarity analysis on the first and second 3D reconstructed features to obtain a reconstructed feature similarity sequence;

[0115] In this embodiment, by comparing the consistency of the spatial structure and surface details of the first and second three-dimensional reconstruction features through point cloud registration and normal vector angle calculation, the three-dimensional process features that are difficult to replicate in counterfeit products (such as the spatial shape of bottle body relief and bottle cap thread) are identified.

[0116] In this embodiment, a multi-dimensional feature matching system is constructed through hierarchical extraction and targeted comparison of second prior features, second deep network extracted features, and second 3D reconstructed features. This system provides precise support for assisting in the determination of the authenticity of liquor. The second prior features cover multiple types of basic 2D features such as edges and shapes, providing prior guidance for subsequent extraction and quickly verifying basic visual consistency. The deep network features condense abstract semantic information, accurately capturing subtle differences such as label deformation and font variations. The 3D reconstructed features supplement spatial dimension information, identifying difficult-to-replicate process details such as relief and threads through point cloud and normal vector comparison. The three types of features are compared in pairs, covering the full range of basic vision, abstract semantics, and spatial process, effectively avoiding interference from irrelevant information, enhancing adaptability to low-contrast scenes, improving the comprehensiveness and accuracy of feature matching, strengthening the ability to identify high-quality counterfeits, providing a reliable basis for assisting in the determination of authenticity, and improving the overall accuracy and anti-interference of liquor bottle identification.

[0117] The foregoing has described a method for assisting in the identification of genuine and counterfeit liquor in an embodiment of the present invention. The following describes a device for assisting in the identification of genuine and counterfeit liquor in an embodiment of the present invention. Please refer to [link / reference]. Figure 8 One embodiment of the auxiliary device for distinguishing genuine and counterfeit liquor in this invention includes:

[0118] The first data acquisition module 1 is used to acquire a global view image dataset of the bottle body, a multi-angle image dataset of the bottle body, a multi-angle image dataset of the bottle cap, and a local view image dataset.

[0119] Preprocessing module 2 is used to preprocess the global view image dataset of the bottle body, the multi-angle image dataset of the bottle body, the multi-angle image dataset of the bottle cap, and the local view image dataset to obtain the standard view image dataset.

[0120] Data storage module 3 is used to store standard viewpoint image datasets into a preset identity database to obtain the wine bottle identity database;

[0121] Data retrieval module 4 is used to acquire the image dataset of the wine bottle to be tested, and retrieve the matching image dataset from the wine bottle identity database according to the preset OCR recognition method and the image dataset of the wine bottle to be tested.

[0122] The similarity analysis module 5 is used to perform similarity analysis on the test wine bottle image dataset and the matching image dataset according to the preset voting method and the preset weighted summation formula, so as to obtain the auxiliary judgment result of the authenticity of the liquor.

[0123] In this embodiment, by collecting a global view image dataset of the bottle body, a multi-angle image dataset of the bottle body, a multi-angle image dataset of the bottle cap, and a local view image dataset, the core features such as the appearance structure and detailed texture of the bottle are comprehensively captured. After image preprocessing, the feature extraction accuracy is improved. A standardized identity database provides an authoritative and unified standard template for verification, ensuring data consistency and reducing comparison errors. OCR technology accurately extracts key information, enabling efficient retrieval of matching images and building a reliable comparison foundation. This avoids the drawbacks of easily copied digital anti-counterfeiting labels. An innovative voting and weighted summation judgment mechanism is adopted, which not only highlights core identification information through weight optimization and avoids the bias of a single algorithm, but also ensures the stability of the judgment through result conflict coordination, improving anti-interference ability and the accuracy of auxiliary identification of genuine and counterfeit liquor. The overall solution specifically addresses the pain points of easy counterfeiting of physical anti-counterfeiting, easy failure of digital anti-counterfeiting, and low accuracy of traditional image recognition. It can effectively curb complex counterfeiting behaviors such as high-quality counterfeits and recycling, while taking into account the efficiency of batch detection and the needs of accurate verification.

[0124] Figure 9 This is a schematic diagram of the structure of a liquor authenticity verification device 900 provided in an embodiment of the present invention. This liquor authenticity verification device 900 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 910 (e.g., one or more processors) and a memory 920, and one or more storage media 930 (e.g., one or more mass storage devices) storing application programs 933 or data 932. The memory 920 and storage media 930 can be temporary or persistent storage. The program stored in the storage media 930 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the liquor authenticity verification device 900. Furthermore, the processor 910 may be configured to communicate with the storage media 930 and execute a series of instruction operations in the storage media 930 on the liquor authenticity verification device 900 to implement the steps of the liquor authenticity verification method provided in the above-described method embodiments.

[0125] A device 900 for assisting in the identification of counterfeit liquor may further include one or more power supplies 940, one or more wired or wireless network interfaces 950, one or more input / output interfaces 960, and / or one or more operating systems 931, such as Windows Server, MacOSX, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 9 The structure of the liquor authenticity identification device shown does not constitute a limitation on the liquor authenticity identification device. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0126] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0127] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0128] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for auxiliary identification of genuine and counterfeit liquor, characterized in that, include: Acquire a global view image dataset of the bottle body, a multi-angle image dataset of the bottle body, a multi-angle image dataset of the bottle cap, and a local view image dataset. Preprocess the global view image dataset of the bottle body, the multi-angle image dataset of the bottle body, the multi-angle image dataset of the bottle cap, and the local view image dataset to obtain the standard view image dataset. The standard viewpoint image dataset is stored in a preset identity database to obtain the wine bottle identity database; Obtain the dataset of images of the wine bottles to be tested, and preprocess the dataset to obtain a standard dataset of images to be tested. The matching image dataset is retrieved from the bottle identity database based on the preset OCR recognition method and the image dataset of the bottle to be tested. Based on a preset voting method and a preset weighted summation formula, a similarity analysis is performed on the dataset of wine bottle images to be tested and the dataset of matching images to obtain auxiliary results for determining the authenticity of the liquor. Specifically, this includes: Two-dimensional feature extraction is performed on the matching image dataset to obtain the first prior features; A digital identity model for wine bottles is constructed based on a pre-defined unsupervised learning model and a pre-defined deep convolutional neural network model. Based on the digital identity model of the wine bottle and the first prior features, feature extraction is performed on the matching image dataset to obtain the first deep network extracted features and the first 3D reconstruction features; Two-dimensional feature extraction is performed on the standard dataset of images to be tested to obtain the second prior features; Based on the digital identity model of the wine bottle and the second prior features, feature extraction is performed on the standard image dataset to be tested, so as to obtain the features extracted by the second deep network and the second three-dimensional reconstruction features. Similarity analysis is performed on the first prior feature, the second prior feature, the first deep network extracted feature, the second deep network extracted feature, the second 3D reconstruction feature, and the first 3D reconstruction feature to obtain the prior feature similarity sequence, the deep network feature similarity sequence, and the reconstruction feature similarity sequence; Fast hashing is performed on the matched image dataset and the standard test image dataset to obtain the average hash similarity sequence, the difference hash similarity sequence, and the perceptual hash similarity sequence, specifically including: A fast hash operation is performed on the matching image dataset to obtain a first average hash sequence, a first difference hash sequence, and a first perceptual hash sequence; A fast hash operation is performed on the standard image dataset to obtain the second average hash sequence, the second differential hash sequence, and the second perceptual hash sequence; A similarity analysis is performed on the first average hash sequence and the second average hash sequence to obtain the average hash similarity sequence; The similarity between the first and second differential hash sequences is calculated to obtain the differential hash similarity sequence; A similarity analysis is performed on the first and second perceptual hash sequences to obtain a perceptual hash similarity sequence; Based on the voting method and weighted summation formula, the average hash similarity sequence, differential hash similarity sequence, perceptual hash similarity sequence, prior feature similarity sequence, deep network feature similarity sequence, and reconstructed feature similarity sequence are analyzed to obtain auxiliary judgment results for the authenticity of liquor.

2. The method for auxiliary identification of genuine and counterfeit liquor as described in claim 1, characterized in that, The step of retrieving the matching image dataset from the wine bottle identity database according to the preset OCR recognition method and the wine bottle image dataset to be tested includes: The dataset of wine bottle images to be tested is preprocessed to obtain a standard dataset of images to be tested. The standard image dataset to be tested is identified using the OCR recognition method to obtain the factory serial number; The wine bottle identification database is searched based on the manufacturing serial number to obtain the search results; If the search result is successful, the matching image dataset is retrieved from the bottle identification database based on the factory serial number.

3. The method for auxiliary identification of genuine and counterfeit liquor as described in claim 1, characterized in that, The analysis of the average hash similarity sequence, difference hash similarity sequence, perceptual hash similarity sequence, prior feature similarity sequence, deep network feature similarity sequence, and reconstructed feature similarity sequence based on the voting method and weighted summation formula to obtain auxiliary judgment results for the authenticity of liquor includes: Based on the preset structural similarity analysis method, a similarity analysis is performed on the standard test image dataset and the matching image dataset to obtain a structural similarity sequence; Based on the preset signal-to-noise ratio analysis method, a similarity analysis is performed on the standard test image dataset and the matched image dataset to obtain the peak signal-to-noise ratio similarity sequence; The similarity analysis is performed on the standard test image dataset and the matching image dataset according to the preset matcher to obtain the matching similarity. The average hash similarity sequence, differential hash similarity sequence, perceptual hash similarity sequence, matching similarity, structural similarity sequence, peak signal-to-noise ratio similarity sequence, prior feature similarity sequence, deep network feature similarity sequence, and reconstructed feature similarity sequence are calculated according to the weighted summation formula to obtain the total feature matching similarity value. Based on the voting method and the total feature matching similarity value, the average hash similarity sequence, differential hash similarity sequence, perceptual hash similarity sequence, matching similarity, structural similarity sequence, peak signal-to-noise ratio similarity sequence, prior feature similarity sequence, deep network feature similarity sequence, and reconstructed feature similarity sequence are analyzed to obtain auxiliary judgment results for the authenticity of liquor.

4. The method for auxiliary identification of genuine and counterfeit liquor as described in claim 3, characterized in that, The step of performing similarity analysis on the standard test image dataset and the matching image dataset according to a preset matcher to obtain the matching similarity includes: The matching image dataset is subjected to feature analysis based on the preset SIFT algorithm to obtain the first feature point descriptor; Feature analysis of the standard image dataset is performed using the SIFT algorithm to obtain the second feature point descriptor; The first and second feature point descriptors are compared and analyzed by the matcher to obtain the total number of matching pairs and the comparison analysis results. The number of correct matches is obtained from the total number of matches based on the comparative analysis results; The ratio of the number of correct matches to the total number of matches is calculated to obtain the matching similarity.

5. The method for auxiliary identification of genuine and counterfeit liquor as described in claim 1, characterized in that, The similarity analysis performed on the first prior feature, the second prior feature, the first deep network extracted feature, the second deep network extracted feature, the second 3D reconstructed feature, and the first 3D reconstructed feature to obtain the prior feature similarity sequence, the deep network feature similarity sequence, and the reconstructed feature similarity sequence includes: A similarity analysis is performed on the first and second prior features to obtain the prior feature similarity sequence; Similarity analysis is performed on the features extracted by the first deep network and the features extracted by the second deep network to obtain the deep network feature similarity sequence; A similarity analysis is performed on the first and second 3D reconstructed features to obtain a reconstructed feature similarity sequence.

6. A device for assisting in the identification of genuine and counterfeit liquor, characterized in that, include: The first data acquisition module is used to acquire a global view image dataset of the bottle body, a multi-angle image dataset of the bottle body, a multi-angle image dataset of the bottle cap, and a local view image dataset. The preprocessing module is used to preprocess the global view image dataset of the bottle body, the multi-angle image dataset of the bottle body, the multi-angle image dataset of the bottle cap, and the local view image dataset to obtain the standard view image dataset. The data storage module is used to store standard viewpoint image datasets into a preset identity database to obtain the wine bottle identity database; The data retrieval module is used to acquire the image dataset of the wine bottle to be tested and to preprocess the image dataset of the wine bottle to be tested to obtain a standard image dataset of the wine bottle to be tested. The matching image dataset is retrieved from the bottle identity database based on the preset OCR recognition method and the image dataset of the bottle to be tested. The similarity analysis module is used to perform similarity analysis on the dataset of wine bottle images to be tested and the dataset of matching images according to a preset voting method and a preset weighted summation formula, so as to obtain auxiliary judgment results on the authenticity of the liquor. Specifically, it includes: Two-dimensional feature extraction is performed on the matching image dataset to obtain the first prior features; A digital identity model for wine bottles is constructed based on a pre-defined unsupervised learning model and a pre-defined deep convolutional neural network model. Based on the digital identity model of the wine bottle and the first prior features, feature extraction is performed on the matching image dataset to obtain the first deep network extracted features and the first 3D reconstruction features; Two-dimensional feature extraction is performed on the standard dataset of images to be tested to obtain the second prior features; Based on the digital identity model of the wine bottle and the second prior features, feature extraction is performed on the standard image dataset to be tested, so as to obtain the features extracted by the second deep network and the second three-dimensional reconstruction features. Similarity analysis is performed on the first prior feature, the second prior feature, the first deep network extracted feature, the second deep network extracted feature, the second 3D reconstruction feature, and the first 3D reconstruction feature to obtain the prior feature similarity sequence, the deep network feature similarity sequence, and the reconstruction feature similarity sequence; Fast hashing is performed on the matched image dataset and the standard test image dataset to obtain the average hash similarity sequence, the difference hash similarity sequence, and the perceptual hash similarity sequence, specifically including: A fast hash operation is performed on the matching image dataset to obtain a first average hash sequence, a first difference hash sequence, and a first perceptual hash sequence; A fast hash operation is performed on the standard image dataset to obtain the second average hash sequence, the second differential hash sequence, and the second perceptual hash sequence; A similarity analysis is performed on the first average hash sequence and the second average hash sequence to obtain the average hash similarity sequence; The similarity between the first and second differential hash sequences is calculated to obtain the differential hash similarity sequence; A similarity analysis is performed on the first and second perceptual hash sequences to obtain a perceptual hash similarity sequence; Based on the voting method and weighted summation formula, the average hash similarity sequence, differential hash similarity sequence, perceptual hash similarity sequence, prior feature similarity sequence, deep network feature similarity sequence, and reconstructed feature similarity sequence are analyzed to obtain auxiliary judgment results for the authenticity of liquor.

7. A device for assisting in the identification of genuine and counterfeit liquor, characterized in that, The aforementioned auxiliary device for distinguishing genuine and counterfeit liquor includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the liquor authenticity identification device to perform the steps of the liquor authenticity identification method as described in any one of claims 1-5.

8. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the method for assisting in the identification of genuine and counterfeit liquor as described in any one of claims 1-5.

Images