Dog food production information intelligent verification system and method based on big data

Abstract

The present application relates to the technical field of product information management, and discloses a dog food production information intelligent verification system and method based on big data, which breaks the inherent loophole of traditional anti-counterfeit codes being true if being real by virtue of the technical features of one-to-one mapping of production end single product traceability codes and multi-dimensional visual features, combined with the verification logic of multi-dimensional feature weighting matching of the consumer end; the technical features of the lightweight mobile end verification entrance and the automatic feature extraction of the convolutional neural network greatly reduce the operation threshold of the consumer end anti-counterfeit verification; the consumer only needs to scan the code through the mobile end and upload the real image according to the instruction, and the system can automatically complete image preprocessing, feature extraction and comparison and determination, without the need of professional pet food identification knowledge, and the verification page synchronously displays the product full-link production traceability information, which not only improves the convenience and anti-counterfeit coverage of user verification, but also realizes the transparent display of dog food production information.

Description

Technical Field

[0001] This invention relates to the technical field of product information management, specifically to a smart verification system and method for dog food production information based on big data. Background Technology

[0002] The quality of dog food products is directly related to pet health and the legal rights of pet owners. As the domestic pet food market continues to expand, counterfeit dog food production methods are constantly evolving and upgrading. Among these, using genuine product codes to package counterfeit goods has become a prevalent and large-scale form of counterfeiting. Counterfeiters recycle genuine packaging, reuse valid anti-counterfeiting codes, and then repackage inferior dog food into genuine packaging before it enters the market. Because dogs are the primary consumers of dog food and cannot provide feedback on their eating experience or quality differences to human consumers, it is difficult for humans to directly identify the authenticity of the product. This makes such counterfeiting highly concealed and difficult to detect. Conventional anti-counterfeiting methods are not suitable for the unique characteristics of the dog food industry; therefore, building a precise and efficient intelligent verification system for production information has become a rigid requirement for the industry. Existing anti-counterfeiting and production information verification technologies for dog food have several shortcomings, making it difficult to effectively address scenarios where genuine codes are used to counterfeit products: First, the current anti-counterfeiting system relies on a single digital anti-counterfeiting code as its core logic. It can only verify whether the anti-counterfeiting code itself exists in the official database, but it cannot establish a strong binding relationship between the anti-counterfeiting code and the physical product. This makes it completely ineffective in scenarios where genuine codes are affixed to counterfeit packaging, resulting in frequent incidents where consumers scan the code to see the genuine product but actually purchase a counterfeit. Second, consumers lack the ability to independently verify the product. Ordinary consumers do not have professional knowledge of dog food identification and cannot identify counterfeit products using packaging details or kibble appearance. They can only rely on manual feedback from official channels, which has a high verification threshold and poor timeliness, making it difficult to improve the coverage and popularity of brand anti-counterfeiting. Third, the anti-counterfeiting verification system lacks a data-driven closed-loop optimization mechanism. The identification models and risk control rules are mostly based on fixed parameters and cannot be continuously iterated and upgraded based on massive consumer verification data. Faced with constantly evolving counterfeiting methods, the false negative rate and false positive rate will continue to rise, making it impossible to maintain stable verification accuracy in the long term.

[0003] Chinese invention patent application CN120338256A, published on July 18, 2025, discloses a quality inspection system and method for industrial products based on data monitoring. The system is connected to a product quality inspection center via a data entry module, a data acquisition module, a database module, a data processing module, a data analysis module, and an early warning analysis module. The data entry module is used to input product production information and set production process sub-sequences. The data acquisition module is used to set data detection points and collect relevant product data information. The data processing module is used to construct a product quality inspection model. The data analysis module is used to generate verification data based on the product quality inspection model and the obtained product data information. The early warning analysis module analyzes the verification data and maintains production equipment based on the analysis results. This invention, to a certain extent, avoids resource waste in the industrial product production process. However, the above technical solution cannot achieve intelligent verification and management of product information for counterfeit goods at the consumer end. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the issues mentioned above, the existing anti-counterfeiting system, which relies on a single digital anti-counterfeiting code as its core logic, can only verify whether the anti-counterfeiting code exists in the official database. It cannot establish a strong binding relationship between the anti-counterfeiting code and the physical product, rendering it completely ineffective in scenarios involving counterfeit products using genuine codes on fake packaging. This leads to frequent incidents where consumers scan the code to see the genuine product but actually purchase a counterfeit. Furthermore, consumers lack the ability to independently verify the product. Ordinary consumers lack the professional knowledge to identify dog ​​food and cannot distinguish counterfeit products using code-switching based on packaging details or kibble appearance. They can only rely on manual feedback from official channels, resulting in high verification thresholds and poor timeliness, hindering the improvement of brand anti-counterfeiting coverage and popularity. Finally, the anti-counterfeiting verification system lacks a data-driven closed loop. The optimization mechanism addresses the issue that the identification models and risk control rules, which often rely on fixed parameters, cannot continuously iterate and upgrade based on massive consumer verification data. Faced with evolving counterfeiting methods, the false negative and false positive rates continue to rise, making it impossible to maintain stable verification accuracy in the long term. This problem is addressed by mapping the traceability codes of individual products at the production end to multi-dimensional visual features, combined with a verification logic based on weighted matching of multi-dimensional features at the consumer end. This breaks the inherent loophole of traditional anti-counterfeiting codes that guarantee authenticity. Through a lightweight mobile verification entry point and automatic feature extraction via convolutional neural networks, along with a closed-loop mechanism of full verification data accumulation and periodic incremental optimization, the verification system achieves adaptive iterative upgrades.

[0006] (II) Technical Solution

[0007] This invention is achieved through the following technical solution: a smart verification method for dog food production information based on big data, comprising the following steps:

[0008] Step S101: Collect multi-dimensional visual feature data of each batch of genuine dog food, generate product traceability code for each dog food product, establish a one-to-one binding relationship between the product traceability code and the visual feature data of the corresponding product, and build a big data resource library of genuine product features.

[0009] Step S102: Receive the QR code verification request initiated by the consumer, verify the legality of the product traceability code, retrieve the corresponding filing information after verification and open the image verification entry, and guide the consumer to upload the physical image of the dog food to be verified.

[0010] Step S103: Perform standardized preprocessing on the uploaded physical object image, and extract multi-dimensional visual feature vectors of the image to be tested through a visual feature extraction network to generate a test feature dataset;

[0011] Step S104: Retrieve the genuine product registration feature set corresponding to the product traceability code from the genuine product feature big data resource library, calculate the comprehensive matching degree between the test feature dataset and the registration feature set through the feature matching algorithm, and generate the code-item binding consistency verification result based on the comparison result;

[0012] Step S105: Retrieve all historical scanning behavior data corresponding to the product traceability code, perform risk control verification from three dimensions: time logic, geographical distribution, and scanning frequency, and generate risk level judgment results based on the verification results;

[0013] Step S106: Output the final verification conclusion based on the consistency verification result of the joint code item binding and the risk level determination result, and simultaneously update all the data of this verification to the genuine product feature big data resource library.

[0014] Step S107: Extract all verification data and labeled risk samples at fixed intervals, perform incremental optimization on the visual feature extraction network and risk control rules, and deploy and go online after verification to complete the closed-loop iteration of the verification system.

[0015] Preferably, step S101 specifically includes:

[0016] During the production and packaging process, industrial vision acquisition equipment is used to collect the packaging appearance features and the characteristics of the dog food pellets in the same batch of products one by one. The packaging appearance features include the outline of the printing pattern, the heat-pressed sealing texture, the shape of the inkjet characters, and the location coordinates of the traceability code. The dog food pellet features include the pellet size range, surface texture features, color value range, and shape regularity parameters.

[0017] Generate a globally unique product traceability code for each dog food product, and establish a strong binding relationship between the packaging appearance characteristics of the product, the characteristics of the dog food pellets in the same batch, and the product traceability code;

[0018] The bound data is categorized and indexed according to production batch, production line number, and production date, and stored in a distributed database to build a structured big data resource library of genuine product characteristics.

[0019] Preferably, step S102 specifically includes:

[0020] Receive product traceability code information uploaded by consumers scanning product packaging, and verify whether the traceability code exists in the genuine product feature big data resource database;

[0021] When the traceability code does not exist in the resource library, it is determined to be an invalid traceability code. The consumer is directly output with a warning message that the anti-counterfeiting code is invalid and the product is not a genuine product, and the verification process is terminated.

[0022] When the traceability code exists in the resource library, it is determined to be a legitimate traceability code. The product basic information and filing feature items corresponding to the traceability code are retrieved, and the image verification entry is displayed on the verification page. Consumers are prompted to upload an overall picture of the packaging, a detailed picture of the seal, and a real picture of the dog food pellets in sequence according to the standardized shooting guidelines to complete the physical image collection.

[0023] Preferably, step S103 specifically includes:

[0024] The uploaded physical images are sequentially processed with Gaussian filtering for noise reduction, histogram equalization for contrast enhancement, and size normalization to generate standardized images with uniform resolution.

[0025] The standardized image is input into a pre-trained convolutional neural network visual feature extraction model to extract the layout feature vector, texture feature vector, and character shape feature vector of the packaging area, as well as the size feature vector, color feature vector, and texture feature vector of the particle area.

[0026] Min-max normalization is performed on the feature vectors of all dimensions to eliminate the dimensional differences between different features and generate a structured dataset of features to be tested.

[0027] Preferably, step S104 specifically includes:

[0028] Based on the product traceability code, retrieve the corresponding set of genuine product registration features from the genuine product feature big data resource library;

[0029] The cosine similarity algorithm is used to calculate the similarity scores of each dimension of features between the feature dataset to be tested and the feature set to be filed.

[0030] The similarity scores of each dimension are weighted and summed according to preset weight coefficients to obtain the comprehensive matching degree; among which, the weight of sealing hot pressing texture is 0.3, the weight of inkjet character shape is 0.25, the weight of particle comprehensive feature is 0.25, and the weight of printing layout is 0.2.

[0031] The overall matching degree is compared with the preset qualified threshold;

[0032] When the overall matching degree is greater than or equal to the preset qualified threshold, it is determined that the product under test is consistent with the characteristics of the registered genuine product, and a verification result of consistent code-item binding is generated.

[0033] When the overall matching degree is less than the preset qualified threshold, it is determined that the product under test does not match the characteristics of the registered genuine product, the generated code and item binding are inconsistent, and there is a risk of genuine code being used to impersonate counterfeit goods.

[0034] Preferably, step S105 specifically includes:

[0035] Retrieve the full historical scanning records of the product's traceability code from the genuine product feature big data resource library, and extract the scanning timestamp, geographic location information, and device unique identifier field for each record;

[0036] Execution time logic verification: Determine whether the first scan time is earlier than the corresponding product's manufacturing time; if the first scan time is earlier than the manufacturing time, it is determined that the time logic is abnormal; if the first scan time is later than or equal to the manufacturing time, it is determined that the time logic is not abnormal.

[0037] Perform a geographic distribution verification: Determine whether there are two or more cross-regional QR code scanning records that are more than a preset distance threshold apart within a preset time window; if a corresponding record exists, determine that a geographic reuse anomaly has been triggered; if no corresponding record exists, determine that a geographic reuse anomaly has not been triggered.

[0038] Perform a scan frequency check: Determine whether the cumulative number of scans for the traceability code exceeds a preset threshold; if the cumulative number of scans exceeds the threshold, a high-frequency reuse anomaly is triggered; if the cumulative number of scans does not exceed the threshold, a high-frequency reuse anomaly is not triggered.

[0039] When at least one of the three checks triggers an anomaly, a high-risk level determination result is generated;

[0040] If none of the three checks trigger an anomaly, a low-risk risk level determination result is generated.

[0041] Preferably, step S106 specifically includes:

[0042] Obtain the consistency verification results of the code-item binding and the risk level determination results, and implement a joint determination using a veto mechanism;

[0043] When the code-item binding consistency verification result is consistent with the code-item binding and the risk level judgment result is low risk, it is judged as genuine. The final conclusion that the verification has passed and it is genuine is output to the consumer. The product production information, quality inspection report and full-chain traceability data are displayed simultaneously. The scanning record is archived to the normal verification dataset and the scanning flow of the genuine product feature big data resource library is updated.

[0044] When the code-item binding consistency verification result is inconsistent, or the risk level judgment result is high risk, or both anomalies exist simultaneously, it is judged as a risky product. A warning conclusion of potential counterfeit goods with code-cloning is output to the consumer, and the specific anomaly dimension and explanation are marked simultaneously. At the same time, the traceability code, image to be tested, anomaly type, and scanning information corresponding to this verification are associated and stored in the risk sample library, and a risk label is added to the corresponding traceability code and the anomaly record is updated in the genuine product feature big data resource library.

[0045] Preferably, step S107 specifically includes:

[0046] The full verification data and manually labeled risk samples within a fixed period are extracted to construct an incremental training set. The visual feature extraction network is then incrementally trained, and the weight parameters of each layer and the matching weights of each feature dimension are adjusted.

[0047] The error rate and missed rate of each risk control rule within the statistical period are used to adjust the threshold parameters and judgment logic of risk control verification accordingly, and generate optimized risk control rules.

[0048] Perform grayscale testing on the optimized visual feature extraction network and risk control rules, and calculate the verification accuracy.

[0049] When the test accuracy reaches the preset threshold, the optimized model and rules will be officially deployed and launched to replace the original version.

[0050] When the test accuracy does not reach the preset threshold, the original version is retained, and the sample is expanded before entering the next round of optimization iteration.

[0051] A big data-based intelligent verification system for dog food production information is used to implement the big data-based intelligent verification method for dog food production information. The system includes:

[0052] The feature library module is used to collect multi-dimensional visual feature data of each batch of genuine dog food, generate product traceability codes and establish a strong binding relationship between the code and the product, and build a big data resource library of genuine product features.

[0053] The QR code interaction module is used to receive QR code verification requests from consumers, verify the legality of the traceability code, open the image verification entry, and guide consumers to upload images of physical objects.

[0054] The image preprocessing module is used to perform standardized preprocessing on the uploaded physical object images, extract multi-dimensional visual feature vectors through a visual feature extraction network, and generate a dataset of features to be tested.

[0055] The feature matching module is used to retrieve the genuine product registration feature set corresponding to the traceability code, calculate the comprehensive matching degree, and generate the code-item binding consistency verification result;

[0056] The risk control verification module is used to retrieve historical scanning behavior data of the traceability code, perform risk control verification from multiple dimensions, and generate risk level judgment results.

[0057] The results output module is used to combine the two types of verification results to output the final conclusion, and at the same time, the verification data is updated to the genuine product feature big data resource library.

[0058] The iterative optimization module is used to extract full verification data and risk samples periodically, perform incremental optimization on the identification model and risk control rules, and deploy and go online after verification to complete the closed-loop iteration of the system.

[0059] (III) Beneficial Effects

[0060] This invention provides an intelligent verification system and method for dog food production information based on big data. It has the following beneficial effects:

[0061] I. By mapping the traceability code of a single product at the production end to multi-dimensional visual features, and combining this with the verification logic of weighted matching of multi-dimensional features at the consumer end, the inherent loophole of traditional anti-counterfeiting codes guaranteeing authenticity is broken. This solution binds each product with multi-dimensional genuine product features such as packaging style, sealing texture, inkjet pattern, and particle characteristics during the production process. During verification, the cosine similarity algorithm is used to calculate the comprehensive matching degree to determine the consistency of the code and the product. This can accurately identify the counterfeiting behavior of recycling genuine packaging and filling inferior dog food with mismatched codes, filling the technical gap in intelligent verification of genuine codes for counterfeit products in the dog food industry, and significantly improving the accuracy of anti-counterfeiting verification.

[0062] Second, by leveraging a lightweight mobile verification portal and the technical features of automatic feature extraction via convolutional neural networks, the operational threshold for anti-counterfeiting verification at the consumer end has been significantly reduced. Consumers only need to scan the code on their mobile devices and upload images of the actual product as instructed. The system can then automatically complete image preprocessing, feature extraction, and comparison and judgment, without requiring professional knowledge of pet food identification. At the same time, the verification page simultaneously displays the product's full-chain production traceability information, which not only improves the convenience of user verification and the coverage of anti-counterfeiting measures, but also achieves transparent display of dog food production information, fully protecting consumers' right to know.

[0063] Third, through the closed-loop mechanism of full-scale verification data accumulation and periodic incremental optimization, the verification system achieves adaptive iterative upgrades. The system synchronously archives normal and risk samples from each verification into the database, and periodically performs incremental training on the visual feature extraction network based on labeled samples and optimizes the risk control threshold parameters. After gray-scale testing verifies that the accuracy rate meets the standards, it is officially launched, enabling the verification system to continuously adapt to constantly upgraded fraud methods, maintain high accuracy and low false judgment rate in the long term, and have stronger scenario adaptability and long-term operation capabilities. Attached Figure Description

[0064] Figure 1 A flowchart of the intelligent verification method for dog food production information based on big data provided by the present invention;

[0065] Figure 2 This is a schematic diagram of the modules of the intelligent verification system for dog food production information based on big data provided by the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] An example of the intelligent verification system and method for dog food production information based on big data is as follows:

[0068] Example 1:

[0069] Please see Figure 1 The intelligent verification method for dog food production information based on big data includes the following steps:

[0070] Step S101: Collect multi-dimensional visual feature data of each batch of genuine dog food, generate product traceability code for each dog food product, establish a one-to-one binding relationship between the product traceability code and the visual feature data of the corresponding product, and build a big data resource library of genuine product features.

[0071] Step S102: Receive the QR code verification request initiated by the consumer, verify the legality of the product traceability code, retrieve the corresponding filing information after verification and open the image verification entry, and guide the consumer to upload the physical image of the dog food to be verified.

[0072] Step S103: Perform standardized preprocessing on the uploaded physical object image, and extract multi-dimensional visual feature vectors of the image to be tested through a visual feature extraction network to generate a test feature dataset;

[0073] Step S104: Retrieve the genuine product registration feature set corresponding to the product traceability code from the genuine product feature big data resource library, calculate the comprehensive matching degree between the test feature dataset and the registration feature set through the feature matching algorithm, and generate the code-item binding consistency verification result based on the comparison result;

[0074] Step S105: Retrieve all historical scanning behavior data corresponding to the product traceability code, perform risk control verification from three dimensions: time logic, geographical distribution, and scanning frequency, and generate risk level judgment results based on the verification results;

[0075] Step S106: Output the final verification conclusion based on the consistency verification result of the joint code item binding and the risk level determination result, and simultaneously update all the data of this verification to the genuine product feature big data resource library.

[0076] Step S107: Extract all verification data and labeled risk samples at fixed intervals, perform incremental optimization on the visual feature extraction network and risk control rules, and deploy and go online after verification to complete the closed-loop iteration of the verification system.

[0077] Step S101 specifically includes:

[0078] During the production and packaging process, industrial vision acquisition equipment is used to collect the packaging appearance features of each dog food product and the features of the dog food pellets in the same batch. The packaging appearance features include the outline of the printing pattern, the heat-pressed sealing texture, the shape of the inkjet characters, and the location coordinates of the traceability code. The dog food pellet features include the pellet size range, surface texture features, color value range, and shape regularity parameters.

[0079] Generate a globally unique product traceability code for each dog food product, and establish a strong binding relationship between the packaging appearance characteristics of the product, the characteristics of the dog food pellets in the same batch, and the product traceability code;

[0080] The bound data is categorized and indexed according to production batch, production line number, and production date, and stored in a distributed database to build a structured big data resource library of genuine product characteristics.

[0081] Specifically, a high-definition industrial area array camera and a ring-shaped shadowless light source are deployed at the final packaging station of the dog food production line to capture multi-angle images of each heat-sealed dog food product. For the packaging appearance, three types of images are captured: a panoramic view of the front of the packaging, a close-up of the sealing area, and a close-up of the production date inkjet printing area. Machine vision algorithms are used to extract corresponding features: For the printed layout outline features, Canny edge detection is used to extract the pixel coordinates and relative positional deviations of the brand logo, product name, and ingredient list patterns. Edge detection determines edge pixels by calculating image gradients, and then performs non-maximum suppression and double thresholding on the edge pixels to obtain clear layout outline boundaries. For the sealing heat-press texture features, gray-level co-occurrence matrix (GLCM) is used to extract four dimensions of parameters: texture contrast, correlation, energy, and entropy. GLCM statistically analyzes the joint gray-level distribution of pixel pairs at specific distances and directions in the image to quantify the spatial correlation of the texture. For the inkjet printing character morphology features, the font height, stroke width, character spacing, and tilt angle parameters of the characters are extracted. Simultaneously, the precise pixel position coordinates of the product traceability code on the packaging are recorded to form a packaging appearance feature set. By using industrial vision acquisition equipment, the automatic extraction of multi-dimensional features of genuine product packaging is achieved, which solves the problem that traditional anti-counterfeiting relies only on a single digital code and cannot establish a code-item binding relationship, providing benchmark data support for subsequent code-item consistency verification.

[0082] For dog food kibble, samples were taken at a rate of 0.5% from the same batch. High-resolution top-view images were captured after the kibble was evenly laid on a white stage. Individual kibble outlines were identified using an instance segmentation algorithm. The size range and average of the major and minor axes of the kibble were calculated. The gray-level co-occurrence matrix features of the surface pore texture were extracted. The mean and standard deviation of the chromaticity of the L, a, and b channels in the Lab color space were quantified. Kibble roundness was statistically analyzed as a shape regularity parameter, forming a feature set for the dog food kibble. In the Lab color space, the L channel represents luminance, ranging from 0 to 100; the a channel represents the range from red to green, ranging from -128 to 127; and the b channel represents the range from yellow to blue, ranging from -128 to 127. The color characteristics of the dog food kibble were comprehensively quantified using these three chromaticity parameters. Kibble roundness was also assessed. The calculation formula is: ;in For particle area, The circumference of the kibble is used; a roundness score closer to 1 indicates a more regular kibble shape. Incorporating these kibble characteristics into the genuine product feature system supplements the anti-counterfeiting verification dimension from the perspective of the contents, solving the problem that verifying only the packaging cannot confirm the quality of the dog food inside, and improving the comprehensiveness of identifying counterfeit products using mislabeling.

[0083] Each dog food product is assigned a globally unique product traceability code, consisting of a 6-digit production batch code, a 2-digit production line number, and an 8-digit serial number. This coding structure ensures that each product has a unique identifier, and the total number of traceability codes can reach 10.16 The system is designed to meet the uniqueness requirements of coding in large-scale production scenarios. The traceability code is printed on the inside opening of the product packaging using a scratch-off silver coating process. The coating design ensures that the traceability code is not visible before the product is first opened, preventing it from being copied and misused during production and distribution. A strong one-to-one mapping relationship is established between the product's packaging appearance features, the characteristics of the dog food pellets in the same batch, and the product's traceability code, ensuring that one traceability code corresponds to only one set of genuine product characteristics. Finally, a three-level classification index is established according to production batch, production line number, and production date. All bound data is stored in a distributed cloud database, constructing a structured big data resource library of genuine product characteristics, providing benchmark data support for subsequent consumer verification. This three-level indexing mechanism improves big data retrieval efficiency, with the response time for a single traceability code query controlled within 100ms, meeting the real-time requirements of high-concurrency consumer verification scenarios.

[0084] Step S102 specifically includes:

[0085] Receive product traceability code information uploaded by consumers scanning product packaging, and verify whether the traceability code exists in the genuine product feature big data resource database;

[0086] When the traceability code does not exist in the resource library, it is determined to be an invalid traceability code. The consumer is directly output with a warning message that the anti-counterfeiting code is invalid and the product is not a genuine product, and the verification process is terminated.

[0087] When the traceability code exists in the resource library, it is determined to be a legitimate traceability code. The product basic information and filing feature items corresponding to the traceability code are retrieved, and the image verification entry is displayed on the verification page. Consumers are prompted to upload an overall picture of the packaging, a detailed picture of the seal, and a real picture of the dog food pellets in sequence according to the standardized shooting guidelines to complete the physical image collection.

[0088] Specifically, after purchasing a product, consumers scratch off the traceability code coating inside the packaging and scan the product's traceability code via WeChat mini-program or the brand's official app. Upon receiving the scan request, the system first searches for matching records of the traceability code in the genuine product feature big data resource database. The scan request carries fields such as the traceability code string, scanning device identifier, scanning timestamp, and IP geolocation. The system performs a precise matching query based on the traceability code's primary key.

[0089] If no matching entry is found, the traceability code is deemed invalid. A red warning message is returned directly to the consumer: "Invalid anti-counterfeiting code, product from unofficial channels. Please purchase with caution." The message also provides access to the brand's official customer service to terminate the verification process. This invalid code detection logic directly blocks counterfeit traceability codes, filtering out fake products with unofficial codes at the source.

[0090] If a matching entry is found and deemed a legitimate traceability code, the system retrieves the corresponding product name, net content, production batch, production date, and quality inspection certificate number, displaying this information at the top of the verification page. Simultaneously, an AI image verification entry pops up. The page includes standardized shooting guidelines, prompting consumers to take a frontal view of the packaging, a close-up view of the sealed heat-pressed area, and a flat view of the dog food kibble after opening. The guidelines explicitly require that the images be well-lit, free of glare, and without blurring or ghosting. After the consumer completes the shooting and submits the image according to the guidelines, the system transmits the image to the backend image processing module to complete the physical image acquisition. These standardized shooting guidelines ensure that images uploaded by consumers adhere to uniform acquisition standards, reducing errors in subsequent feature extraction caused by differences in shooting angle and lighting, and improving the stability and accuracy of the verification results.

[0091] Step S103 specifically includes:

[0092] The uploaded physical images are sequentially processed with Gaussian filtering for noise reduction, histogram equalization for contrast enhancement, and size normalization to generate standardized images with uniform resolution.

[0093] The standardized image is input into a pre-trained convolutional neural network visual feature extraction model to extract the layout feature vector, texture feature vector, and character shape feature vector of the packaging area, as well as the size feature vector, color feature vector, and texture feature vector of the particle area.

[0094] Min-max normalization is performed on the feature vectors of all dimensions to eliminate the dimensional differences between different features and generate a structured dataset of features to be tested.

[0095] Specifically, for the three physical product images uploaded by consumers, three preprocessing steps are performed sequentially: First, a 5×5 Gaussian filter algorithm is used to perform convolution operations on the images to remove random noise and graininess generated by mobile phone shooting, and smooth areas of abrupt changes in pixel grayscale. The Gaussian filter generates a convolution kernel through a two-dimensional Gaussian function, and performs a weighted average on the surrounding pixels. The weights are Gaussian distributed with respect to the distance from the center pixel, which preserves the image edge details well while removing noise. Second, a histogram equalization algorithm is used to adjust the image grayscale distribution, stretching the originally concentrated grayscale range to the full range of 0-255, enhancing image contrast, and strengthening details such as packaging edges and grain texture. Histogram equalization maps pixel grayscale values ​​through a cumulative distribution function, making the grayscale histogram of the output image approximately uniformly distributed. Finally, all images are scaled to a uniform size of 640×640 pixels and converted to a standard RGB three-channel format to generate standardized images. The three-step preprocessing process transforms non-standardized images captured by consumers into standardized data that meets the model's input requirements. This solves the problem of inconsistent image quality on mobile devices affecting the accuracy of feature extraction, and provides a standardized data foundation for the subsequent accurate extraction of multi-dimensional visual features.

[0096] Standardized images are input into a pre-trained convolutional neural network (CNN) visual feature extraction model. This model consists of 5 convolutional layers, 3 max-pooling layers, and 2 fully connected layers, using 3×3 convolutional kernels to mine local features from the image. The convolutional layers extract features within the local receptive field by sliding the kernel; shallow convolutions extract low-level features such as edges and textures, while deep convolutions extract semantic-level abstract features. The max-pooling layer downsamples by taking the maximum value within a window, reducing feature dimensionality while preserving key features and improving the model's translation invariance. For the overall packaging image and the sealing detail image, a 128-dimensional layout feature vector, a 128-dimensional texture feature vector, and a 64-dimensional character shape feature vector are output, respectively. For the particle image, a 64-dimensional size feature vector, a 64-dimensional chroma feature vector, and a 128-dimensional texture feature vector are output. All feature vectors are floating-point numerical vectors.

[0097] Finally, min-max normalization is performed on the feature vectors of all dimensions, mapping the values ​​of each dimension to the 0-1 range using a formula; eliminating the dimensional differences between different features and generating a structured dataset of features to be tested; min-max normalization ensures that features of each dimension have the same numerical scale in subsequent similarity calculations, avoiding features with larger dimensions from dominating the matching results, and improving the scientificity and accuracy of multi-dimensional feature fusion matching.

[0098] Step S104 specifically includes:

[0099] Based on the product traceability code, retrieve the corresponding set of genuine product registration features from the genuine product feature big data resource library;

[0100] The cosine similarity algorithm is used to calculate the similarity scores of each dimension of features between the feature dataset to be tested and the feature set to be filed.

[0101] The similarity scores of each dimension are weighted and summed according to preset weight coefficients to obtain the comprehensive matching degree; among which, the weight of sealing hot pressing texture is 0.3, the weight of inkjet character shape is 0.25, the weight of particle comprehensive feature is 0.25, and the weight of printing layout is 0.2.

[0102] The overall matching degree is compared with the preset qualified threshold;

[0103] When the overall matching degree is greater than or equal to the preset qualified threshold, it is determined that the product under test is consistent with the characteristics of the registered genuine product, and a verification result of consistent code-item binding is generated.

[0104] When the overall matching degree is less than the preset qualified threshold, it is determined that the product under test does not match the characteristics of the registered genuine product, the generated code and item binding are inconsistent, and there is a risk of genuine code being used to impersonate counterfeit goods.

[0105] Specifically, based on the product traceability code currently scanned, the genuine product registration feature set bound to the code is retrieved from the genuine product feature big data resource library. This set contains six feature vectors that completely correspond to the feature dimensions to be tested, ensuring that the feature dimensions correspond one-to-one and that the matching calculation is comparable.

[0106] The cosine similarity algorithm is used to calculate the similarity score of each dimension of the feature dataset to be tested and the feature set to be filed. The calculation formula is as follows:

[0107] ,in The first feature vector to be tested Dimensional value, For the first feature vector of the filing Dimensional value, For feature vectors, the calculated similarity score ranges from 0 to 1, with higher scores indicating stronger feature consistency. Cosine similarity measures the degree of similarity between two vectors in a direction, is unaffected by the absolute length of the vectors, and is suitable for similarity calculation scenarios after the feature vector numerical scale is unified.

[0108] The similarity scores for each dimension are weighted and summed according to preset weighting coefficients to calculate the overall matching degree. The calculation formula is as follows: ;in For the similarity of the hot-pressed sealing texture, For the similarity of the inkjet characters, The similarity of the overall characteristics of dog food pellets. The similarity is determined by the print layout. The weighting is based on the anti-counterfeiting differentiation of each feature. The heat-pressed seal texture is given the highest weight because it is the most difficult to copy, while the print layout is given a lower weight because it is relatively easy to counterfeit.

[0109] A preset threshold of 0.8 is used to determine the overall matching degree. The calculated overall matching degree is compared with this threshold. When the overall matching degree is ≥0.8, the product under test is determined to be consistent with the registered features of the genuine product, generating a verification result indicating that the code and the product are bound together. When the overall matching degree is <0.8, the product under test is determined to be inconsistent with the registered features of the genuine product, generating a verification result indicating that the code and the product are not bound together and there is a risk of counterfeit goods using genuine codes. Through multi-dimensional weighted matching, the misjudgment caused by the influence of the shooting environment on a single feature is reduced, improving the accuracy of verification. The multi-dimensional feature weighted matching mechanism verifies the consistency of the code and the product from both the packaging and the contents, solving the core pain point of traditional anti-counterfeiting methods that only verify the traceability code itself and cannot identify counterfeit goods using genuine codes, thus achieving accurate identification of counterfeit code-using scenarios.

[0110] Step S105 specifically includes:

[0111] Retrieve the full historical scanning records of the product's traceability code from the genuine product feature big data resource library, and extract the scanning timestamp, geographic location information, and device unique identifier field for each record;

[0112] Execution time logic verification: Determine whether the first scan time is earlier than the corresponding product's manufacturing time; if the first scan time is earlier than the manufacturing time, it is determined that the time logic is abnormal; if the first scan time is later than or equal to the manufacturing time, it is determined that the time logic is not abnormal.

[0113] Perform a geographic distribution verification: Determine whether there are two or more cross-regional QR code scanning records that are more than a preset distance threshold apart within a preset time window; if a corresponding record exists, determine that a geographic reuse anomaly has been triggered; if no corresponding record exists, determine that a geographic reuse anomaly has not been triggered.

[0114] Perform a scan frequency check: Determine whether the cumulative number of scans for the traceability code exceeds a preset threshold; if the cumulative number of scans exceeds the threshold, a high-frequency reuse anomaly is triggered; if the cumulative number of scans does not exceed the threshold, a high-frequency reuse anomaly is not triggered.

[0115] When at least one of the three checks triggers an anomaly, a high-risk level determination result is generated;

[0116] If none of the three checks trigger an anomaly, a low-risk risk level determination result is generated.

[0117] Specifically, the system retrieves all historical scanning records of the product's traceability code since its entry into the genuine product feature big data resource library, extracts three core fields for each record: scanning timestamp, IP location latitude and longitude, and unique device identifier, and performs three risk control checks in sequence:

[0118] Time logic verification: Retrieve the manufacturing timestamp of the corresponding batch of the product and compare it with the first scan timestamp of the traceability code. If the first scan timestamp is earlier than the manufacturing timestamp, a time logic anomaly is triggered, indicating a risk of premature leakage and misuse of the traceability code. Under normal circumstances, it is impossible for consumers to scan and verify the code before the product leaves the factory. If the first scan timestamp is later than or equal to the manufacturing timestamp, no time logic anomaly is triggered. Time logic verification identifies abnormal leakage scenarios of traceability codes from a temporal perspective, intercepting counterfeit activities such as premature copying and misuse of traceability codes in the production and distribution process.

[0119] Geographical Distribution Verification: A 24-hour time window is set, and the geographical locations of all scanned records within this time window are statistically analyzed. The straight-line distance between any two scanned locations is calculated. Geographical locations are obtained based on IP address resolution, achieving city-level accuracy. Distance calculation uses the Haversine formula, which calculates the spherical distance between two points on the Earth's surface based on latitude and longitude. The formula is as follows:

[0120] ,in For the Earth's radius, , For the first point, latitude and longitude, , The second point refers to latitude and longitude. When the straight-line distance between two or more scanning locations exceeds 500 kilometers, a regional reuse anomaly is triggered, indicating a risk of cross-regional reuse and counterfeiting of the traceability code. It is impossible for a single product to appear in multiple geographically distant locations within a short period. When no corresponding record exists, a regional reuse anomaly is not triggered. Regional distribution verification identifies counterfeiting scenarios where the traceability code is copied and used simultaneously in multiple locations from a spatial perspective, supplementing the behavioral dimension verification that image verification cannot cover.

[0121] Scan frequency verification: This function counts the cumulative number of scans of the traceability code since its activation. If the cumulative number of scans exceeds 3, a high-frequency reuse anomaly is triggered, indicating a risk of the traceability code being copied and reused multiple times. Normally, 1-2 scans per product by a consumer is reasonable; more than 3 scans strongly suggests the code has been copied and distributed. If the cumulative number of scans is ≤3, no high-frequency reuse anomaly is triggered. This scan frequency verification identifies counterfeit scenarios involving the mass copying of traceability codes based on usage frequency. The frequency threshold is set in conjunction with statistical data on normal consumer behavior, balancing detection and false positive rates.

[0122] After completing the three verifications, a comprehensive judgment is performed: if at least one of the three verifications triggers an anomaly, a high-risk level judgment result is generated; if none of the three verifications trigger an anomaly, a low-risk level judgment result is generated. Through multi-dimensional behavioral risk control verification, it supplements the identification of code-swapping and counterfeiting scenarios from the behavioral logic level, forming a dual guarantee with image feature verification. The three-dimensional risk control verification system constructs a behavioral anomaly identification model from the three dimensions of time, space, and frequency, solving the problem that single image verification cannot identify the reuse and theft of traceability codes. It forms a complementary dual anti-counterfeiting mechanism with code-item consistency verification, comprehensively improving the ability to intercept counterfeiting.

[0123] Step S106 specifically includes:

[0124] Obtain the consistency verification results of the code-item binding and the risk level determination results, and implement a joint determination using a veto mechanism;

[0125] When the code-item binding consistency verification result is consistent with the code-item binding and the risk level judgment result is low risk, it is judged as genuine. The final conclusion that the verification has passed and it is genuine is output to the consumer. The product production information, quality inspection report and full-chain traceability data are displayed simultaneously. The scanning record is archived to the normal verification dataset and the scanning flow of the genuine product feature big data resource library is updated.

[0126] When the code-item binding consistency verification result is inconsistent, or the risk level judgment result is high risk, or both anomalies exist simultaneously, it is judged as a risky product. A warning conclusion of potential counterfeit goods with code-cloning is output to the consumer, and the specific anomaly dimension and explanation are marked simultaneously. At the same time, the traceability code, image to be tested, anomaly type, and scanning information corresponding to this verification are associated and stored in the risk sample library, and a risk label is added to the corresponding traceability code and the anomaly record is updated in the genuine product feature big data resource library.

[0127] Specifically, the system obtains the consistency verification results of the code-item binding and the risk level judgment results, and adopts a veto mechanism to perform joint judgment. That is, if any one of the verifications fails, the whole product is judged as a risky product, ensuring the rigor of anti-counterfeiting judgment.

[0128] When the code-item binding consistency verification result shows that the code and item are bound correctly, and the risk level assessment result is low risk, the product is ultimately determined to be genuine. A green verification pass is displayed on the consumer-facing page, indicating a genuine product. Simultaneously, the entire production traceability information for the product is displayed, including the production team, raw material source, quality inspection report number, and logistics nodes. The timestamp, geographical location, and equipment identification information of this scan are archived into the normal verification dataset, and the corresponding traceability code's scan log is updated in the genuine product feature big data resource library. This full-chain traceability information display achieves transparency in production information, allowing consumers to fully understand the product's origin and quality assurance. At the same time, the continuous accumulation of normal verification data provides positive sample support for subsequent model iterations.

[0129] When the code-item binding consistency verification result is inconsistent, or the risk level assessment result is high risk, or both anomalies exist simultaneously, the product is ultimately determined to be a risky product. A red warning about the risk of counterfeit products with mismatched codes is displayed on the consumer-facing page, along with specific anomalies such as inconsistent sealing textures with genuine product registration or the anti-counterfeiting code having been scanned multiple times. This clearly informs consumers of the risk dimensions and provides an official customer service feedback portal for further manual verification. The backend stores the traceability code number, the image to be tested, the anomaly type, the scanning time, and the geographical information associated with this verification in a risk sample library. A risk tag is added to the corresponding traceability code in the genuine product feature big data resource library, and the anomaly record entry is updated, providing data support for subsequent model optimization and counterfeit clue tracing. The structured storage of risk samples provides a negative sample data source for subsequent incremental model training and risk control rule optimization, supporting the continuous iteration and upgrading of the verification system. A veto-all joint judgment mechanism ensures that anomalies in any dimension are effectively intercepted, solving the problem that single-dimensional verification is easily circumvented by counterfeiters, and constructing a multi-layered, comprehensive anti-counterfeiting verification system.

[0130] Step S107 specifically includes:

[0131] The full verification data and manually labeled risk samples within a fixed period are extracted to construct an incremental training set. The visual feature extraction network is then incrementally trained, and the weight parameters of each layer and the matching weights of each feature dimension are adjusted.

[0132] The error rate and missed rate of each risk control rule within the statistical period are used to adjust the threshold parameters and judgment logic of risk control verification accordingly, and generate optimized risk control rules.

[0133] Perform grayscale testing on the optimized visual feature extraction network and risk control rules, and calculate the verification accuracy.

[0134] When the test accuracy reaches the preset threshold, the optimized model and rules will be officially deployed and launched to replace the original version.

[0135] When the test accuracy does not reach the preset threshold, the original version is retained, and the sample is expanded before entering the next round of optimization iteration.

[0136] Specifically, with a 7-day optimization cycle, all verification data within the cycle, along with manually reviewed and labeled risk samples, misjudged positive samples, and missed negative samples, are extracted to construct an incremental training set. Manual labeling is performed by brand anti-counterfeiting specialists who verify and confirm the system's judgment results, categorizing them into three types: correct judgment, misjudgment (genuine product judged as risky), and missed judgment (risky product judged as genuine). The labeled data serves as the baseline ground truth for model optimization.

[0137] For the visual feature extraction network, incremental training is performed using transfer learning, with the cross-entropy loss function as the optimization objective. The formula for calculating the cross-entropy loss function is as follows: ;in For real labels, The model predicts probabilities; it adjusts the weight parameters of each layer of the network using the backpropagation algorithm, focusing on optimizing the discriminative power of easily confused features; and it recalculates the optimal matching weights for each feature dimension based on new samples, improving the model's accuracy in identifying new counterfeit packaging and differences in particles from different batches. Transfer learning is based on the low-level feature weights of the pre-trained model, only fine-tuning the parameters of the high-level feature extraction and classification layers, enabling rapid adaptation to new data in scenarios with small sample increments, avoiding the high cost of training from scratch.

[0138] For risk control rules, the false positive rate and false negative rate of each of the three rules—time logic verification, geographical distribution verification, and scanning frequency verification—were statistically analyzed within the specified period. The false positive rate is calculated as: False positive rate = Number of false positives / Total number of genuine products, reflecting the proportion of normal products incorrectly identified as risky. The false negative rate is calculated as: False negative rate = Number of false negatives / Total number of risky products, reflecting the proportion of counterfeit products not identified. For example, if the false positive rate of geographical distribution verification is too high, the distance threshold is appropriately increased to 600 kilometers or the time window is narrowed to 12 hours; if the false negative rate of high-frequency reuse verification is too high, the number of false negatives is lowered to 2. Through data-driven parameter optimization, the adaptability of risk control rules is continuously improved. Dynamically adjusting risk control parameters based on statistical feedback from real-world operational data solves the problems of fixed-parameter rules failing to adapt to market changes and accuracy declining after long-term operation.

[0139] After model training and rule tuning are completed, accuracy verification is conducted in a gray-scale testing environment. The overall verification accuracy is calculated based on a manually labeled sample set. When the test accuracy is ≥95%, the optimization is deemed successful, and the new version of the visual feature extraction model and risk control rules are officially deployed to the production environment, replacing the original running version. When the test accuracy is <95%, the optimization is deemed unsuccessful, the original running version is retained, and the labeled samples are supplemented and expanded before entering the next round of optimization iterations. The gray-scale release mechanism ensures that the new version is fully verified before being officially launched, avoiding model degradation from affecting the online verification experience. Through a closed-loop iteration mechanism, the verification system can continuously adapt to constantly evolving counterfeiting methods, maintaining a high verification accuracy rate in the long term. The periodic iterative optimization mechanism constructs a data-driven self-improvement closed loop, solving the problem that traditional anti-counterfeiting systems with fixed parameters cannot cope with the continuous upgrading of counterfeiting methods, ensuring the long-term effectiveness and adaptability of the verification system.

[0140] Example 2:

[0141] Please see Figures 1-2 A big data-based intelligent verification system for dog food production information is used to implement a big data-based intelligent verification method for dog food production information. The system includes:

[0142] The feature library module is used to collect multi-dimensional visual feature data of each batch of genuine dog food, generate product traceability codes and establish a strong binding relationship between the code and the product, and build a big data resource library of genuine product features.

[0143] The QR code interaction module is used to receive QR code verification requests from consumers, verify the legality of the traceability code, open the image verification entry, and guide consumers to upload images of physical objects.

[0144] The image preprocessing module is used to perform standardized preprocessing on the uploaded physical object images, extract multi-dimensional visual feature vectors through a visual feature extraction network, and generate a dataset of features to be tested.

[0145] The feature matching module is used to retrieve the genuine product registration feature set corresponding to the traceability code, calculate the comprehensive matching degree, and generate the code-item binding consistency verification result;

[0146] The risk control verification module is used to retrieve historical scanning behavior data of the traceability code, perform risk control verification from multiple dimensions, and generate risk level judgment results.

[0147] The results output module is used to combine the two types of verification results to output the final conclusion, and at the same time, the verification data is updated to the genuine product feature big data resource library.

[0148] The iterative optimization module is used to extract full verification data and risk samples periodically, perform incremental optimization on the identification model and risk control rules, and deploy and go online after verification to complete the closed-loop iteration of the system.

[0149] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent verification of dog food production information based on big data, characterized in that, Includes the following steps: Step S101: Collect multi-dimensional visual feature data of each batch of genuine dog food, generate product traceability code for each dog food product, establish a one-to-one binding relationship between the product traceability code and the visual feature data of the corresponding product, and build a big data resource library of genuine product features. Step S102: Receive the QR code verification request initiated by the consumer, verify the legality of the product traceability code, retrieve the corresponding filing information after verification and open the image verification entry, and guide the consumer to upload the physical image of the dog food to be verified. Step S103: Perform standardized preprocessing on the uploaded physical object image, and extract multi-dimensional visual feature vectors of the image to be tested through a visual feature extraction network to generate a test feature dataset; Step S104: Retrieve the genuine product registration feature set corresponding to the product traceability code from the genuine product feature big data resource library, calculate the comprehensive matching degree between the test feature dataset and the registration feature set through the feature matching algorithm, and generate the code-item binding consistency verification result based on the comparison result; Step S105: Retrieve all historical scanning behavior data corresponding to the product traceability code, perform risk control verification from three dimensions: time logic, geographical distribution, and scanning frequency, and generate risk level judgment results based on the verification results; Step S106: Output the final verification conclusion based on the consistency verification result of the joint code item binding and the risk level determination result, and simultaneously update all the data of this verification to the genuine product feature big data resource library. Step S107: Extract all verification data and labeled risk samples at fixed intervals, perform incremental optimization on the visual feature extraction network and risk control rules, and deploy and go online after verification to complete the closed-loop iteration of the verification system.

2. The intelligent verification method for dog food production information based on big data according to claim 1, characterized in that, Step S101 specifically includes: During the production and packaging process, industrial vision acquisition equipment is used to collect the packaging appearance features and the characteristics of the dog food pellets in the same batch of products one by one. The packaging appearance features include the outline of the printing pattern, the heat-pressed sealing texture, the shape of the inkjet characters, and the location coordinates of the traceability code. The dog food pellet features include the pellet size range, surface texture features, color value range, and shape regularity parameters. Generate a globally unique product traceability code for each dog food product, and establish a strong binding relationship between the packaging appearance characteristics of the product, the characteristics of the dog food pellets in the same batch, and the product traceability code; The bound data is categorized and indexed according to production batch, production line number, and production date, and stored in a distributed database to build a structured big data resource library of genuine product characteristics.

3. The intelligent verification method for dog food production information based on big data according to claim 1, characterized in that, Step S102 specifically includes: Receive product traceability code information uploaded by consumers scanning product packaging, and verify whether the traceability code exists in the genuine product feature big data resource database; When the traceability code does not exist in the resource library, it is determined to be an invalid traceability code. The consumer is directly output with a warning message that the anti-counterfeiting code is invalid and the product is not a genuine product, and the verification process is terminated. When the traceability code exists in the resource library, it is determined to be a legitimate traceability code. The product basic information and filing feature items corresponding to the traceability code are retrieved, and the image verification entry is displayed on the verification page. Consumers are prompted to upload an overall picture of the packaging, a detailed picture of the seal, and a real picture of the dog food pellets in sequence according to the standardized shooting guidelines to complete the physical image collection.

4. The intelligent verification method for dog food production information based on big data according to claim 1, characterized in that, Step S103 specifically includes: The uploaded physical images are sequentially processed with Gaussian filtering for noise reduction, histogram equalization for contrast enhancement, and size normalization to generate standardized images with uniform resolution. The standardized image is input into a pre-trained convolutional neural network visual feature extraction model to extract the layout feature vector, texture feature vector, and character shape feature vector of the packaging area, as well as the size feature vector, color feature vector, and texture feature vector of the particle area. Min-max normalization is performed on the feature vectors of all dimensions to eliminate the dimensional differences between different features and generate a structured dataset of features to be tested.

5. The intelligent verification method for dog food production information based on big data according to claim 1, characterized in that, Step S104 specifically includes: Based on the product traceability code, retrieve the corresponding set of genuine product registration features from the genuine product feature big data resource library; The cosine similarity algorithm is used to calculate the similarity scores of each dimension of features between the feature dataset to be tested and the feature set to be filed. The similarity scores of each dimension are weighted and summed according to preset weight coefficients to obtain the comprehensive matching degree; among which, the weight of sealing hot pressing texture is 0.3, the weight of inkjet character shape is 0.25, the weight of particle comprehensive feature is 0.25, and the weight of printing layout is 0.

2. The overall matching degree is compared with the preset qualified threshold; When the overall matching degree is greater than or equal to the preset qualified threshold, it is determined that the product under test is consistent with the characteristics of the registered genuine product, and a verification result of consistent code-item binding is generated. When the overall matching degree is less than the preset qualified threshold, it is determined that the product under test does not match the characteristics of the registered genuine product, the generated code and item binding are inconsistent, and there is a risk of genuine code being used to impersonate counterfeit goods.

6. The intelligent verification method for dog food production information based on big data according to claim 1, characterized in that, Step S105 specifically includes: Retrieve the full historical scanning records of the product's traceability code from the genuine product feature big data resource library, and extract the scanning timestamp, geographic location information, and device unique identifier field for each record; Execution time logic verification: Determine whether the first scan time is earlier than the corresponding product's manufacturing time; if the first scan time is earlier than the manufacturing time, it is determined that the time logic is abnormal; if the first scan time is later than or equal to the manufacturing time, it is determined that the time logic is not abnormal. Perform a geographic distribution verification: Determine whether there are two or more cross-regional QR code scanning records that are more than a preset distance threshold apart within a preset time window; if a corresponding record exists, determine that a geographic reuse anomaly has been triggered; if no corresponding record exists, determine that a geographic reuse anomaly has not been triggered. Perform a scan frequency check: Determine whether the cumulative number of scans for the traceability code exceeds a preset threshold; if the cumulative number of scans exceeds the threshold, a high-frequency reuse anomaly is triggered; if the cumulative number of scans does not exceed the threshold, a high-frequency reuse anomaly is not triggered. When at least one of the three checks triggers an anomaly, a high-risk level determination result is generated; If none of the three checks trigger an anomaly, a low-risk risk level determination result is generated.

7. The intelligent verification method for dog food production information based on big data according to claim 1, characterized in that, Step S106 specifically includes: Obtain the consistency verification results of the code-item binding and the risk level determination results, and implement a joint determination using a veto mechanism; When the code-item binding consistency verification result is consistent with the code-item binding and the risk level judgment result is low risk, it is judged as genuine. The final conclusion that the verification has passed and it is genuine is output to the consumer. The product production information, quality inspection report and full-chain traceability data are displayed simultaneously. The scanning record is archived to the normal verification dataset and the scanning flow of the genuine product feature big data resource library is updated. When the code-item binding consistency verification result is inconsistent, or the risk level judgment result is high risk, or both anomalies exist simultaneously, it is judged as a risky product. A warning conclusion of potential counterfeit goods with code-cloning is output to the consumer, and the specific anomaly dimension and explanation are marked simultaneously. At the same time, the traceability code, image to be tested, anomaly type, and scanning information corresponding to this verification are associated and stored in the risk sample library, and a risk label is added to the corresponding traceability code and the anomaly record is updated in the genuine product feature big data resource library.

8. The intelligent verification method for dog food production information based on big data according to claim 1, characterized in that, Step S107 specifically includes: The full verification data and manually labeled risk samples within a fixed period are extracted to construct an incremental training set. The visual feature extraction network is then incrementally trained, and the weight parameters of each layer and the matching weights of each feature dimension are adjusted. The error rate and missed rate of each risk control rule within the statistical period are used to adjust the threshold parameters and judgment logic of risk control verification accordingly, and generate optimized risk control rules. Perform grayscale testing on the optimized visual feature extraction network and risk control rules, and calculate the verification accuracy. When the test accuracy reaches the preset threshold, the optimized model and rules will be officially deployed and launched to replace the original version. When the test accuracy does not reach the preset threshold, the original version is retained, and the sample is expanded before entering the next round of optimization iteration.

9. A smart verification system for dog food production information based on big data, characterized in that: The system is used to implement the intelligent verification method for dog food production information based on big data as described in any one of claims 1-8, the system comprising: The feature library module is used to collect multi-dimensional visual feature data of each batch of genuine dog food, generate product traceability codes and establish a strong binding relationship between the code and the product, and build a big data resource library of genuine product features. The QR code interaction module is used to receive QR code verification requests from consumers, verify the legality of the traceability code, open the image verification entry, and guide consumers to upload images of physical objects. The image preprocessing module is used to perform standardized preprocessing on the uploaded physical object images, extract multi-dimensional visual feature vectors through a visual feature extraction network, and generate a dataset of features to be tested. The feature matching module is used to retrieve the genuine product registration feature set corresponding to the traceability code, calculate the comprehensive matching degree, and generate the code-item binding consistency verification result; The risk control verification module is used to retrieve historical scanning behavior data of the traceability code, perform risk control verification from multiple dimensions, and generate risk level judgment results. The results output module is used to combine the two types of verification results to output the final conclusion, and at the same time, the verification data is updated to the genuine product feature big data resource library. The iterative optimization module is used to extract full verification data and risk samples periodically, perform incremental optimization on the identification model and risk control rules, and deploy and go online after verification to complete the closed-loop iteration of the system.

Images