Packed tobacco authenticity identification method based on image acquisition and algorithm identification
By employing image acquisition and deep learning-based methods, efficient and accurate identification of tobacco packaging has been achieved, solving the problems of high cost and low efficiency of manual identification and adapting to changing market demands.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-05
- Publication Date
- 2026-03-24
AI Technical Summary
In existing technologies, the identification of genuine and counterfeit tobacco packaging relies on manual inspection, which has problems such as high labor costs, low efficiency, and difficulty in guaranteeing accuracy. Especially when tobacco products are rapidly iterated and there are many categories, it is difficult to achieve large-scale, batch identification.
An identification method based on image acquisition and deep learning algorithms is adopted. The front-end interaction module standardizes image acquisition, the back-end processing module performs feature extraction and analysis, and the deep learning model is combined to output the true and false identification results. The back-end management module manages the archiving and verification of the results.
It enables efficient and accurate identification of genuine and counterfeit tobacco packaging, reduces labor costs, improves identification efficiency and accuracy, supports large-scale batch testing, and its systematic process adapts to market changes.
Smart Images
Figure CN121724637A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of information technology, and in particular to a packaging tobacco authenticity identification method based on image acquisition and algorithm recognition. BACKGROUND
[0002] In the production and circulation, market supervision and other links of tobacco products, the cigarette box packaging as the core carrier of product information, its text identification (such as brand information, tar content, warning language), pattern design, compliance identification and other elements not only need to meet the requirements of industry standards and laws and regulations, but also can be used as key features for identifying and determining the authenticity of cigarettes.
[0003] At present, the identification of cigarette box information mainly depends on manual completion, and the audit personnel need to check the images or real objects of cigarette boxes one by one, and manually check whether the information is complete, accurate and compliant. There are the following defects: High labor cost: the cigarette box information is complex and needs to be checked one by one, and a large number of professional personnel need to be equipped, especially in the scene of fast iteration and multiple categories of tobacco products, the long-term labor cost is huge.
[0004] Low identification efficiency: manual inspection of each item requires a lot of time, and the single box identification process is complicated, which is difficult to meet the large-scale and batch identification demand.
[0005] The accuracy is difficult to guarantee: different professional personnel have different understandings of the standards, and the correlation check of information (such as whether the brand name is consistent, whether the warning language position is compliant, etc.) depends on manual memory and comparison, which is easy to cause misjudgment and omission due to subjective judgment deviation or negligence.
[0006] High training cost: the relevant standards are updated frequently, and there are many brands and versions, so new and on-the-job personnel need to be continuously trained and examined, and it takes time and cost to establish a unified judgment standard and professional ability, which is difficult to quickly copy and scale.
[0007] Therefore, how to ensure the consistency and integrity of feature information in multi-view image acquisition and processing, and comprehensively capture the key features of tobacco packaging through effective technical means, has become a key problem to be solved in the current tobacco industry anti-counterfeiting management. SUMMARY
[0008] The present application provides a packaging tobacco authenticity identification method based on image acquisition and algorithm recognition, mainly comprising: S1. Obtain multi-view image data of tobacco packaging through a front-end interactive module, the front-end interactive module provides a guided acquisition interface to standardize the image acquisition process; S2. Transmit the multi-view image data to a back-end processing module, the back-end processing module pre-processes and extracts features from the image data to generate a feature map; S3. analyzing the feature map by a deep learning model in the back-end processing module, the deep learning model outputting a result of identifying the authenticity of the tobacco packaging; S4. feeding the result of identifying the authenticity to a front-end interactive module and a back-end management module, the back-end management module archiving and reviewing the result of identifying the authenticity.
[0009] Further, the front-end interactive module acquires multi-view image data of the tobacco packaging, including: the front-end interactive module displays a plurality of view acquisition guide interfaces, the guide interfaces providing shooting views and environmental prompts; the acquisition state of each view is recorded in the guide interface, the acquisition state presenting the views that have completed acquisition in the form of an identifier; when the image acquisition of all predetermined views is completed, a submission function is activated, the submission function being used for integrating the acquired image data; a unique association identifier and a user identifier are added to the integrated image data, the unique association identifier being used for data tracking, and the user identifier being used for distinguishing the upload source; the image data is transmitted to the back-end processing module in an encrypted manner, the encrypted manner protecting the security in the data transmission process.
[0010] Further, the multi-view image data is transmitted to the back-end processing module, including: receiving the multi-view image data transmitted by the front-end interactive module, the image data containing image contents of a plurality of predetermined views; performing direction correction processing on the image data, the direction correction being based on image size and shooting angle analysis to unify the image direction; identifying the tobacco packaging main body region in the image data by a pre-trained object detection model, the object detection model outputting a bounding box of the main body region; performing cropping processing on the image data according to the bounding box, the cropping processing removing background interference to retain the main body region content; taking the cropped image data as an input for subsequent feature extraction, the input being used to generate a feature map.
[0011] Further, the deep learning model in the back-end processing module analyzes the feature map, including: classifying and identifying the image features in the feature map, the classification and identification determining the brand information of the tobacco packaging; matching a corresponding feature template according to the brand information, the feature template being used to provide a comparison benchmark for the authenticity features; extracting key feature points in the feature map by a feature extraction algorithm, the key feature points being used to represent the detailed information of the tobacco packaging; comparing the key feature points with the feature template in terms of similarity, the similarity comparison result being used as a basis for authenticity determination; comprehensively analyzing the similarity comparison result by the deep learning model, the model outputting a final result of identifying the authenticity.
[0012] Further, the feedback of the authenticity identification result to the front-end interaction module and the back-end management module comprises: transmitting the authenticity identification result to the front-end interaction module, and the transmission result is used for user to view the identification state; and archiving the authenticity identification result to the back-end management module, and the archiving record contains image data and identification details; for suspicious data in the authenticity identification result, triggering an artificial review process, the artificial review process performs secondary confirmation on the suspicious data; and updating the identification record according to the result of the artificial review process, and the updated record is used for optimizing the training data of the subsequent identification model.
[0013] Further, the analysis of the feature graph by the deep learning model in the back-end processing module comprises: performing splicing processing on the feature graph according to a predetermined size, the splicing processing integrates feature information of multiple perspectives; inputting the splicing-processed feature graph into a pre-trained deep learning model, the model is trained based on a large amount of sample data; performing classification analysis on the feature graph by the deep learning model, the classification analysis outputs a judgment result representing authenticity; and recording confidence data of the judgment result, the confidence data is used for evaluating the reliability of the identification result.
[0014] Further, the transmission of the multi-perspective image data to the back-end processing module comprises: performing format verification on the multi-perspective image data, the format verification confirms the integrity of the image data; if the format verification is passed, storing the image data to a temporary database, the temporary database is used for subsequent processing calling; extracting image data from the temporary database for preprocessing, the preprocessing includes image resolution adjustment and noise filtering; and taking the preprocessed image data as input data for feature extraction, the input data is used for constructing a feature graph.
[0015] Another purpose of the present application is to provide a packaging tobacco authenticity identification system based on image acquisition and algorithm identification, comprising: An image directional acquisition and standardized uploading module is used for acquiring multi-perspective image data of tobacco packaging, and provides a guided acquisition interface to standardize the image acquisition process. A back-end processing module is used for receiving multi-perspective image data, pre-processing and feature extraction of the multi-perspective image data to generate a feature graph. A deep learning efficient detection algorithm module is used for analyzing the feature graph and outputting an authenticity identification result of the tobacco packaging. A process management and feedback module is used for feeding back the authenticity identification result to the front-end interaction module and the back-end management module, and the back-end management module archives and reviews the identification result.
[0016] This invention provides an integrated solution of "targeted data collection via mini-program + deep learning recognition + backend platform management," which offers a more efficient and faster approach compared to existing technologies. Specific advantages are as follows: 1. Convenience and traceability in the data collection process The program's targeted data collection function is easy to operate, and users can quickly get started without professional training. It supports taking and uploading cigarette box pictures anytime, anywhere, greatly reducing the threshold for data collection and solving the problems of traditional data collection methods that rely on professional equipment and are complicated to operate.
[0017] The mini-program automatically records the time, location, and image information of each collection, forming a complete collection archive, which facilitates subsequent traceability and verification, avoiding the shortcomings of manual recording that is prone to omissions and difficult to track.
[0018] 2. Efficiency and intelligence in the identification process Deep learning models are trained on a large number of samples and combined with targeted images to accurately extract and match cigarette box features. They can also effectively distinguish cigarette boxes with similar appearances, and the recognition accuracy is significantly improved compared with traditional manual recognition.
[0019] The automated identification process significantly reduces the processing time for a single cigarette box, can efficiently meet the needs of large-scale batch inspection, reduces reliance on manual identification, and lowers labor costs.
[0020] 3. Systematization and controllability of management processes The back-end platform integrates collected data, identification results, and manual review records to form a complete information loop, supporting data statistics, trend analysis, and traceability management, thus solving the problems of easy loss and difficulty in tracing traditional manual records.
[0021] The manual review module is seamlessly integrated with the automatic recognition module, which quickly marks and reconfirms unrecognized or questionable results. This retains the efficiency of machine recognition while ensuring the reliability of results through human intervention, achieving a collaborative model of "machine as the main and human as the auxiliary".
[0022] 4. Standardization and scalability of the overall process The integrated solution establishes a standardized process from data collection to identification and management, unifies operating procedures and judgment criteria, avoids subjective differences in human identification, and ensures consistency of results under different scenarios and by different personnel.
[0023] The system architecture supports model iteration and upgrades as well as functional expansion. It can quickly update the identification library based on the emergence of new brands and packaging in the tobacco industry to adapt to changing market demands. Attached Figure Description
[0024] Figure 1The present invention provides a flowchart of a method for identifying genuine and counterfeit packaged tobacco based on image acquisition and algorithm recognition. Figure 1 .
[0025] Figure 2 The present invention provides a flowchart of a method for identifying genuine and counterfeit packaged tobacco based on image acquisition and algorithm recognition. Figure 2 . Detailed Implementation
[0026] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.
[0027] like Figures 1-2 This embodiment of a method for identifying the authenticity of packaged tobacco based on image acquisition and algorithm recognition may specifically include: S1, the system achieves standardized acquisition of tobacco packaging images through a front-end applet. The applet provides a guided interactive interface to ensure that users complete image acquisition from a predetermined perspective and encrypts and transmits the acquired image data to the back-end server.
[0028] Specifically, the front-end mini-program design adopts a standardized image acquisition process. Through guided interaction, it presents all images to be acquired from the front, back, right, left, top, and bottom of the cigarette box in a grid format. The guided interaction ensures that users complete the image acquisition of the necessary perspectives and key elements as required.
[0029] The interface will provide prompts such as shooting angle and shooting environment. After entering the shooting interface for any part (such as the "front of the cigarette pack"), the interface retains a "Return to Part Selection" button, allowing users to exit the current interface at any time during the shooting process and return to the selection list to switch to other parts that have not been captured (for example, if the lighting is poor when shooting the side, users can temporarily return to select the front to shoot and then reshoot the side later). Parts that have been captured will automatically update with a green "Captured" icon next to their diagram and include a shooting timestamp, making it easy for users to clearly understand the capture progress.
[0030] The onboarding interface includes a "Submit Photos" confirmation button. This button is only clickable when 6 photos have been uploaded. When the upload is incomplete, it is grayed out and displays the "Number of Photos Uploaded" to prevent users from missing key perspectives.
[0031] After the data collection is completed, a unique association identifier (such as timestamp + ID) and a unique user identifier (openId) are added to the image group, and the data is encrypted and transmitted to the service backend to create data records. The authenticity of the data is then automatically identified through deep learning algorithms.
[0032] S2, the backend server receives tobacco packaging image data uploaded by the frontend applet. The image data contains image content from multiple perspectives. The backend server preprocesses and extracts features from the image data using a deep learning algorithm to generate a feature map for authenticity identification.
[0033] The backend server obtains uploaded tobacco packaging image data from the frontend mini-program. This image data contains original images from multiple perspectives. First, these original images undergo orientation correction to ensure consistent orientation across different perspectives, providing a standardized first image set for subsequent processing. For this first image set, a pre-trained deep learning tool is used to detect tobacco packaging areas within the images, determining the boundary range of the tobacco packaging in each perspective image. The detected areas are then cropped to generate a second image set containing only the main body of the tobacco packaging. Key feature points are extracted from the second image set. For each cropped image from each perspective, specific areas for authenticity verification are identified, forming a third image set containing feature distributions. This third image set is compared with pre-established genuine product template features to generate a feature map for authenticity verification, ensuring that the feature map accurately reflects the subtle differences in tobacco packaging details.
[0034] In one possible implementation, the backend server obtains the uploaded tobacco packaging image data from the frontend applet. This image data includes original image content from multiple perspectives such as the front, back, and sides. Orientation correction processing is performed on these original images, for example, by automatically rotating the images by analyzing the image size and shooting angle to ensure that the orientation of all images is consistent, thereby forming a standardized first image group. This can eliminate orientation deviations caused by different user shooting habits, which is beneficial to improving the accuracy and efficiency of subsequent processing.
[0035] For example, for the first image group, a pre-trained deep learning tool is used to detect the tobacco packaging area in the image. The deep learning tool here refers to an object detection framework based on a convolutional neural network. It learns from a large number of labeled images to identify object boundaries. The specific process includes: after inputting the image, the framework extracts image features and predicts the position and size of the bounding box, determines the boundary range of the tobacco packaging in each view image, and then crops these detected areas to generate a second image group containing only the main body of the tobacco packaging. This cropping helps to remove background noise, making feature extraction more focused on packaging details, thereby improving the reliability of authenticity identification.
[0036] In one possible implementation, key feature points are extracted from the second image group. For each cropped image from each viewpoint, specific areas for authenticity verification are identified, such as brand logos or warning labels on the front image. The texture and edges of these areas are detected by feature extraction algorithms to form a third image group containing feature distributions. This approach can capture microscopic differences in packaging, which is beneficial to the accuracy of subsequent comparisons and avoids subjective errors from human judgment.
[0037] For example, the third image group is compared with the pre-established genuine product template features, which are a standardized set of features extracted from real tobacco packaging samples in advance. The comparison process involves calculating the similarity between feature points, such as comparing the distribution and intensity of key points through matching algorithms, and generating a feature map for authenticity identification. This ensures that the feature map can accurately reflect the detailed differences in tobacco packaging. This comparison is beneficial for quickly identifying counterfeit products, improving identification efficiency and reducing costs.
[0038] In one possible implementation, the specific implementation of orientation correction processing can take into account that the images uploaded by users may differ depending on whether the phone is in landscape or portrait orientation. For example, if the frontal image is in landscape orientation, it can be automatically rotated 90 degrees to the left to make it uniformly portrait orientation. This helps to standardize the input and avoid deviations in the algorithm when processing inconsistent orientations, thereby ensuring the consistency of the entire process.
[0039] For example, when detecting tobacco packaging areas, the training process of a pre-trained deep learning tool involves iteratively optimizing parameters using a labeled dataset, enabling the tool to adapt to various lighting and background conditions. The second image set generated after cropping focuses on the subject, which helps reduce computational resource consumption and improve processing speed.
[0040] In one possible implementation, when extracting key feature points, a suitable extractor can be selected for regions with rich textures, such as identifying the color and shape distribution of brand patterns to form a third image group. This is beneficial for highlighting key elements that differentiate between genuine and counterfeit products and supports comprehensive identification from multiple perspectives.
[0041] For example, when comparing the third image group with the genuine template, if the similarity is high, it is determined to be genuine; otherwise, it is counterfeit. This process of generating feature maps is beneficial for visualizing differences, helping users understand the basis for identification, and providing an efficient tool for large-scale supervision.
[0042] S3, the backend server analyzes the feature map based on a pre-trained deep learning model. The model outputs the authenticity identification result based on the brand and model characteristics of the tobacco packaging and synchronously feeds the result back to the front-end mini-program and the back-end management platform.
[0043] The process involves acquiring raw image data from multiple sides of the tobacco packaging image to be inspected. The raw image of each side is then standardized to a pre-defined uniform size. These processed images are then stitched together to form a complete feature map image for subsequent feature analysis. The stitched feature map image is input into a pre-established deep learning framework. This framework, trained on a large number of samples for specific tobacco brands and models, can extract key features from the packaging image and classify the feature map image, outputting preliminary authenticity identification results. These preliminary results are compared with a pre-defined brand and model feature library. If the result matches the authenticity criteria in the feature library, a final authenticity conclusion is determined and formatted as structured data for transmission to other modules. This structured data is synchronously transmitted to the front-end mini-program interface and the back-end management platform, ensuring that the authenticity identification results are presented in real time. This facilitates the relevant modules in obtaining the results and performing subsequent processing, completing the closed-loop process for tobacco packaging authenticity identification.
[0044] For example, when processing tobacco packaging images, after obtaining the original image data from multiple sides, the data is adjusted to a uniform size through size standardization, such as scaling the image of each side to a fixed pixel grid. This ensures the consistency of the subsequent stitching process and helps improve the accuracy of feature extraction, because inconsistent sizes may lead to information loss or distortion, thereby affecting the reliability of the overall analysis.
[0045] Specifically, this standardization process aims to unify the image proportions, making the stitched feature map image easier to input into the model, avoiding computational deviations caused by size differences, and thus resulting in a more stable recognition effect.
[0046] In one embodiment, the images of the six sides of a cigarette box are adjusted and stitched together to form a complete map, which can capture packaging details from multiple dimensions, such as texture and color distribution. This helps the model to better identify subtle differences and achieve accuracy in judging authenticity.
[0047] In one possible implementation, the stitched feature map image is input into a pre-established deep learning framework. This framework is trained on a large number of samples for specific tobacco brands and models, such as using tens of thousands of genuine and counterfeit cigarette box images for iterative optimization. The framework contains convolutional layers and pooling layers to extract key features such as edge contours and color gradients. Then, it outputs a preliminary result of authenticity by classification. The advantage of this framework is that it can automatically learn complex patterns, which is more efficient than manual inspection and reduces human error.
[0048] Specifically, the training process of the framework involves backpropagation to adjust the weights in order to minimize the classification error, thereby enabling fast image processing in practical applications and bringing the advantage of real-time recognition. For example, given an input map, the framework can output results within seconds, supporting high-throughput detection scenarios.
[0049] It should be noted that the preliminary authenticity identification results are compared with a preset brand and model feature library. For example, the feature library stores standardized texture templates of genuine cigarette boxes and common defect patterns of counterfeit products. If the result matches the genuine product standard, such as a specific QR code pattern, the final conclusion is determined and formatted as structured data. The purpose of this comparison is to verify the model output, enhance robustness, avoid potential biases from a single framework, and bring more reliable authenticity conclusions.
[0050] In one embodiment, for a suspected counterfeit image, if the result does not match the color consistency standard in the feature library, it is formatted as JSON data containing probability scores for easy subsequent transmission and auditing.
[0051] Specifically, structured data is synchronously transmitted to the front-end mini-program interface and the back-end management platform, for example, through real-time push via API interface, to ensure that the results are presented immediately. This makes it convenient for users to view the data and achieves a closed-loop process, improving the system's response speed and user experience. For example, while displaying genuine and counterfeit labels on the mini-program, the back-end records logs to track counterfeit trends, thereby improving the overall anti-counterfeiting efficiency.
[0052] S4, the back-end management platform receives the recognition results and performs closed-loop data management. The management includes support for manual review of unrecognized or questionable results, as well as continuous updates to the recognition model and rule base to improve system adaptability and accuracy.
[0053] Image data of cigarette boxes to be identified is acquired from the front-end device. Features are extracted from the images using a pre-established deep learning architecture to generate corresponding feature vector sets for subsequent matching and judgment. The generated feature vector sets are compared with a preset feature library to determine if a match exists. If the match result is lower than a preset confidence threshold, the image data is marked as a questionable image and pushed to the back-end management platform. The back-end management platform receives the data marked as questionable images, triggers a manual review process, and stores the reviewed annotation information in the database for subsequent feature library updates. The reviewed annotation information is retrieved from the database, and the feature library is periodically incrementally adjusted to ensure the continuous expansion of the feature vector set coverage, improving its adaptability to cigarette box images and completing closed-loop data management.
[0054] The back-end management platform archives and performs statistical analysis on the recognition results. The archiving includes the retention of collected data and audit logs, and the statistical analysis is used to optimize the update strategy of the recognition model and rule base.
[0055] The backend management platform acquires raw data from the acquisition device for each collection, including timestamps, geographic locations, and corresponding image content. This raw data is then initially archived chronologically to form initial archive records. For these initial archive records, the backend management platform correlates and matches the raw data with the corresponding recognition results, while simultaneously recording log information from manual review, creating a complete dataset containing multi-dimensional information. Key fields are extracted from the complete archive dataset, and the backend management platform categorizes and statistically analyzes the accuracy and error types of the recognition results, generating statistical reports for subsequent rule adjustments. Based on the categorized statistical results in the reports, the backend management platform structurally stores the archived content of the recognition results, ensuring the integrity of the collected data and review logs, while also providing data support for optimizing the update of the recognition rule base.
[0056] In addition, the back-end management platform supports seamless integration between the manual review module and the automatic recognition module. The manual review module performs secondary confirmation on questionable results to ensure the reliability and consistency of the recognition results.
[0057] The backend management platform, through a pre-established interface mechanism, transmits unconfirmed or questionable result data from the automatic identification process to the manual review module in real time. A list of tasks to be processed is generated for these questionable results for subsequent handling. The manual review module retrieves the questionable result data from the task list and uses pre-defined annotation tools to quickly label and initially assess this data. If the labeling results differ from the automatic identification results, the labeled data is categorized as a dataset requiring secondary confirmation. For datasets requiring secondary confirmation, the manual review module uses a built-in comparison tool to compare the labeled data with the original automatic identification results item by item to determine the final identification result and ensure consistency. The backend management platform integrates the processing records of the final identification results and questionable results and updates them to the database for subsequent tracking and optimization of the accuracy of the automatic identification process, ensuring seamless integration between the manual review module and the automatic identification process.
[0058] Another embodiment of the present invention provides a system for identifying the authenticity of packaged tobacco based on image acquisition and algorithm recognition, comprising: The image orientation acquisition and standardized upload module is used to acquire multi-view image data of tobacco packaging and provides a guided acquisition interface to standardize the image acquisition process. The backend processing module is used to receive multi-view image data, preprocess the multi-view image data and extract features to generate feature maps; A deep learning-based high-efficiency detection algorithm module is used to analyze the feature map and output the authenticity identification result of tobacco packaging; The process management and feedback module is used to feed back the authenticity identification results to the front-end interaction module and the back-end management module. The back-end management module archives and verifies the identification results.
[0059] The above embodiments are merely one of the preferred embodiments of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.
Claims
1. A method for identifying genuine and counterfeit packaged tobacco based on image acquisition and algorithm recognition, characterized in that, include: S1. Acquire multi-view image data of tobacco packaging through a front-end interaction module, wherein the front-end interaction module provides a guided acquisition interface to standardize the image acquisition process; S2. The multi-view image data is transmitted to the back-end processing module, which performs preprocessing and feature extraction on the image data to generate a feature map; S3. The feature map is analyzed using a deep learning model, and the deep learning model outputs the authenticity identification result of the tobacco packaging; S4. The authenticity recognition result is fed back to the front-end interaction module and the back-end management module. The back-end management module archives and verifies the recognition result.
2. The method for identifying genuine and counterfeit packaged tobacco based on image acquisition and algorithm recognition as described in claim 1, characterized in that, The process of acquiring multi-view image data of tobacco packaging through the front-end interaction module includes: The front-end interactive module displays a collection guidance interface with multiple perspectives, which provides shooting perspective and environmental prompts. The acquisition status of each viewpoint is recorded in the guidance interface, and the acquisition status is presented in the form of an identifier for the viewpoints that have been acquired. Once all images from the predetermined viewpoints have been acquired, the submission function is activated. This submission function is used to integrate the acquired image data. A unique association identifier and a user identifier are added to the integrated image data. The unique association identifier is used for data tracking, and the user identifier is used to distinguish the upload source. The image data is transmitted to the backend processing module using an encrypted method, which protects the security of the data transmission process.
3. The method for identifying genuine and counterfeit packaged tobacco based on image acquisition and algorithm recognition as described in claim 1, characterized in that, The step of transmitting the multi-view image data to the back-end processing module includes: Receive multi-view image data transmitted by the front-end interaction module, wherein the image data contains image content from multiple predetermined viewpoints; The image data is subjected to orientation correction processing, which is based on the analysis of image size and shooting angle to unify the image orientation; The object detection model identifies the tobacco packaging main body region in the image data using a pre-trained object detection model, and the object detection model outputs the bounding box of the main body region. The image data is cropped based on the bounding box, and the cropping process removes background interference to retain the main content area. The cropped image data is used as input for subsequent feature extraction, and the input is used to generate a feature map.
4. The method for identifying genuine and counterfeit packaged tobacco based on image acquisition and algorithm recognition as described in claim 1, characterized in that, The analysis of the feature map using the deep learning model in the backend processing module includes: The image features in the feature map are classified and identified, and the classification and identification determines the brand information of the tobacco packaging; The corresponding feature template is matched based on the brand information, and the feature template is used to provide a benchmark for comparing genuine and counterfeit features; Key feature points are extracted from the feature map using a feature extraction algorithm. These key feature points are used to characterize the detailed information of the tobacco packaging. The key feature points are compared with the feature template in terms of similarity, and the similarity comparison result is used as the basis for determining authenticity. The deep learning model comprehensively analyzes the similarity comparison results and outputs the final result for identifying whether a match is genuine or fake.
5. The method for identifying genuine and counterfeit packaged tobacco based on image acquisition and algorithm recognition as described in claim 1, characterized in that, The step of feeding back the authenticity identification result to the front-end interaction module and the back-end management module includes: The authenticity recognition result is transmitted to the front-end interaction module, and the transmission result is used by the user to view the recognition status; Simultaneously, the authenticity recognition results are archived to the backend management module, and the archived record includes image data and recognition details; For questionable data in the authenticity identification results, a manual review process is triggered, in which the questionable data is reconfirmed. The identification record is updated based on the results of the manual review process, and the updated record is used to optimize the training data for subsequent identification models.
6. The method for identifying genuine and counterfeit packaged tobacco based on image acquisition and algorithm recognition as described in claim 1, characterized in that, The analysis of the feature map using the deep learning model in the backend processing module includes: The feature maps are stitched together at a predetermined size, and the stitching process integrates feature information from multiple perspectives. The spliced feature map is input into a pre-trained deep learning model, which is trained based on a large amount of sample data; The deep learning model is used to classify and analyze the feature map, and the classification and analysis outputs the determination result of the authenticity of the representation; Record the confidence level data of the determination result, which is used to evaluate the reliability of the identification result.
7. The method for identifying genuine and counterfeit packaged tobacco based on image acquisition and algorithm recognition as described in claim 1, characterized in that, The step of transmitting the multi-view image data to the back-end processing module includes: The format of the multi-view image data is validated to ensure the integrity of the image data. If the format verification passes, the image data is stored in a temporary database, which is used for subsequent processing. Image data is extracted from the temporary database and preprocessed, including image resolution adjustment and noise filtering. The preprocessed image data is used as input data for feature extraction, and the input data is used to construct a feature map.
8. A system for identifying genuine and counterfeit packaged tobacco based on image acquisition and algorithm recognition, characterized in that, The system applies the method according to any one of claims 1-7, including: The image orientation acquisition and standardized upload module is used to acquire multi-view image data of tobacco packaging and provides a guided acquisition interface to standardize the image acquisition process. The backend processing module is used to receive multi-view image data, preprocess the multi-view image data and extract features to generate feature maps; A deep learning-based high-efficiency detection algorithm module is used to analyze the feature map and output the authenticity identification result of tobacco packaging; The process management and feedback module is used to feed back the authenticity identification results to the front-end interaction module and the back-end management module. The back-end management module archives and verifies the identification results.