Product information intelligent verification method and system

Abstract

The application discloses a product information intelligent verification method and system. The method comprises the following steps: obtaining product information and carrying out image-text diversion processing to obtain real object images and text information respectively; carrying out self-adaptive preprocessing on the real object images, extracting multi-dimensional visual features, and obtaining visual feature extraction results; carrying out structured analysis on the text information, screening out key fields highly related to verification, and obtaining text key fields; carrying out correlation verification on the visual feature extraction results and the text key fields to obtain preliminary verification results; and carrying out multi-dimensional matching on the preliminary verification results and a standardized genuine product feature benchmark library, adopting a weighted scoring mechanism to calculate feature matching degrees, and outputting verification results of complete consistency, high similarity or obvious inconsistency according to a preset threshold. Through the technical architecture of image-text double-path parallel processing, multi-dimensional feature comparison and data closed-loop management, the application realizes automatic verification and risk early warning of product information uploaded by customers.

Description

Technical Field

[0001] This invention relates to the field of product information verification and intelligent identification technology, and more specifically to a product information intelligent verification method and system. Background Technology

[0002] Currently, there are various solutions in the field of product information verification, but all of them have significant technical shortcomings and are unable to meet the needs of large-scale, efficient, and cross-entity product information consistency verification in scenarios such as enterprise centralized procurement, merchant warehousing, and testing agency audits. The specific problems are as follows: Traditional manual verification mode: Relying on manual comparison of product text information and physical information, the verification efficiency is extremely low and cannot meet the large-scale needs of bulk procurement; at the same time, the verification results are easily affected by factors such as the subjective experience and fatigue of the staff, making it difficult to guarantee the accuracy and consistency of information verification, and the manually recorded verification data is difficult to achieve efficient aggregation and analysis.

[0003] Isolated barcode scanning tool solution: Existing barcode scanning tools can only read basic coding information such as barcodes and QR codes on products, and can only retrieve basic product information. They cannot complete the linkage comparison between product physical image information and text information submitted by merchants. The verification dimension is single and cannot verify the authenticity and accuracy of product information.

[0004] Brand-owned closed anti-counterfeiting systems: Each brand's anti-counterfeiting and traceability system is exclusive to its own brand, with strong closedness and exclusivity. It lacks cross-brand and cross-supply chain universality and cannot achieve collaborative verification by multiple entities such as physical merchants, e-commerce merchants, supply chain merchants, testing institutions, and purchasing companies. It is difficult to adapt to product circulation and quality control scenarios involving multiple entities.

[0005] Insufficient image data processing capabilities: Some existing solutions involving product image recognition have not been optimized for the actual product images collected. When faced with actual collection problems such as uneven lighting, insufficient resolution, and angle deviation, the image data quality is poor, which leads to a significant reduction in the accuracy of subsequent feature extraction and comparison, affecting the reliability of verification results.

[0006] In summary, existing technologies cannot construct a universal, intelligent, and closed-loop multi-entity product information verification system, failing to meet the market's core demand for efficient verification of product information authenticity and accuracy, and also unable to provide comprehensive and reliable technical support for procurement decisions, quality control, and market supervision. Therefore, designing an intelligent product information verification method and system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides an intelligent product information verification method and system, aiming to solve the industry challenge of verifying the consistency of product information in scenarios such as enterprise centralized procurement, merchant warehousing, and testing agency audits. It overcomes the technical shortcomings of existing verification schemes, such as low efficiency, limited dimensions, poor universality, and insufficient image data processing capabilities, and constructs an intelligent product information verification system for multiple stakeholders, including physical merchants, e-commerce merchants, and supply chain merchants. By integrating intelligent image processing, machine vision recognition, and data storage and analysis technologies, it achieves accurate and efficient comparison between the textual information and physical image information submitted by merchants, ensuring the authenticity and accuracy of product information. Simultaneously, it provides multi-dimensional verification data support and decision-making references for purchasing enterprises, testing agencies, and market supervision departments, adapting to the needs of large-scale, cross-stakeholder product information verification.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A product information intelligent verification method includes: Step 1: Obtain product information and perform sorting processing to obtain physical images and text information of the product; Step 2: Perform adaptive preprocessing on the physical image and extract multi-dimensional visual features to obtain the visual feature extraction results; perform structured parsing on the text information and extract the key verification fields to obtain the text key fields; Step 3: Verify the visual feature extraction results by associating them with the key text fields to obtain preliminary verification results; Step 4: Match the obtained preliminary verification results with the standardized genuine product feature benchmark library, calculate the feature matching degree, and output the verification results based on the feature matching degree.

[0009] Preferably, step 2, the adaptive preprocessing of the physical image, includes: Local histogram equalization is used to address uneven lighting in real-world images, improving detail in dark areas and suppressing highlights. Super-resolution reconstruction or downsampling algorithms are used to unify the physical image to a preset resolution standard; By using perspective transformation and geometric correction, the main body of the product in the tilted real-world image is straightened to a standard perspective; A standardized image is output by employing denoising and blur restoration preprocessing methods.

[0010] Preferably, based on a deep convolutional neural network or a classic feature point algorithm, multi-dimensional visual features are extracted from the standardized image, including: the color distribution of the product, information on surface texture, key visual elements, and geometric features. The extracted features are encoded to generate feature vectors for matching, thereby obtaining the visual feature extraction results.

[0011] Preferably, the process of obtaining key text fields in step 2 includes: performing structured parsing of the text information, splitting it into fields such as product name, model, specifications, batch, and barcode, and filtering out core fields that are strongly related to product verification through key field extraction methods.

[0012] Preferably, step 3 includes: The visual feature extraction results are correlated and verified with the key text fields to obtain preliminary verification results, which are divided into pass and fail. If the verification passes, a multi-dimensional comparison will be performed; if it fails, a prompt will be made to resubmit the information. If the verification fails multiple times, a manual verification option will be automatically added.

[0013] Preferably, the standardized genuine product feature benchmark library is used to store the standard visual features and standard text information of products of various categories, brands and models, and adopts a brand-authorized update mechanism to ensure the authenticity and authority of product information.

[0014] Preferably, the matching verification results include: completely consistent, highly similar, and obviously inconsistent; A weighted scoring mechanism is adopted, which sets weights for features in different dimensions and calculates the final matching score.

[0015] Preferably, a product information intelligent verification system includes: an information collection and classification module, an image and text preprocessing and extraction module, an information consistency verification module, a multi-dimensional comparison algorithm module, a standardized genuine product feature benchmark library, and a record retention and data analysis module; The information collection and classification module is used to receive product information uploaded by customers, automatically identify and classify it into two categories: physical images and text information, and complete data integrity verification. The image and text preprocessing and extraction module is used to adaptively preprocess real object images and extract multi-dimensional visual features, and to perform structured parsing of text information and extract key verification fields. The information consistency verification module is used to perform correlation and matching verification between visual features and key text fields. If the verification passes, it will proceed to the multi-dimensional comparison stage; if it fails, it will prompt the information to be retransmitted or automatically transfer to manual verification. The multi-dimensional comparison algorithm module calls the standardized genuine product feature benchmark library, performs matching calculations between the extracted product features and the features in the benchmark library, and outputs the verification results. The record retention and data analysis module is used to retain and verify complete records and to periodically aggregate and analyze the data.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a product information intelligent verification method and system, with the following beneficial effects: This invention solves the problem of poor image data quality in existing solutions: by using an independently optimized image enhancement algorithm to achieve adaptive correction of product images, the clarity and standardization of image data are greatly improved, and the accuracy of product feature extraction is increased by more than 80%, providing a high-precision data foundation for subsequent comparison. This invention also achieves precise comparison of image and text information: through a standardized genuine product feature benchmark library and feature point matching algorithm, it completes multi-dimensional verification of product entity features and text information, breaking through the limitation of isolated barcode scanning tools with a single verification dimension, and the verification accuracy rate can reach over 99%. This invention also breaks down industry barriers and enables multi-entity collaborative verification: the multi-entity interconnected architecture is adapted to multiple entities such as physical merchants, e-commerce merchants, and supply chain merchants, and supports universal verification across brands and supply chains, solving the problem of poor universality of brand owners' own anti-counterfeiting systems. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 The method flowchart provided by the present invention.

[0019] Figure 2 This is an overall structural diagram provided for the present invention. Detailed Implementation

[0020] 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.

[0021] The embodiments of the present invention address the image quality problem in non-standard acquisition scenarios by implementing adaptive preprocessing and multi-dimensional visual feature extraction for real object images. Structured parsing and key field extraction of text information enable consistency verification with image information; Based on a standardized genuine product feature benchmark library, a multi-dimensional comparison algorithm is used to achieve accurate genuine product verification and risk classification. Construct a data closed loop of "collection-verification-retention-analysis" to support product information iteration and market risk early warning.

[0022] like Figure 1 As shown in the figure, an embodiment of the present invention discloses an intelligent verification method for product information, including: Step 1: Obtain product information and perform sorting processing to obtain physical images and text information of the product; Step 2: Perform adaptive preprocessing on the physical image and extract multi-dimensional visual features to obtain the visual feature extraction results; perform structured parsing on the text information and extract the key verification fields to obtain the text key fields; Step 3: Verify the visual feature extraction results by associating them with the key text fields to obtain preliminary verification results; Step 4: Match the obtained preliminary verification results with the standardized genuine product feature benchmark library, calculate the feature matching degree, and output the verification results based on the feature matching degree.

[0023] Specifically, step 2, the adaptive preprocessing of the physical image, includes: Local histogram equalization is used to address uneven lighting in real-world images, improving detail in dark areas and suppressing highlights. Super-resolution reconstruction or downsampling algorithms are used to unify the physical image to a preset resolution standard; By using perspective transformation and geometric correction, the main body of the product in the tilted real-world image is straightened to a standard perspective; A standardized image is output by employing denoising and blur restoration preprocessing methods.

[0024] Specifically, based on deep convolutional neural networks or classic feature point algorithms, multi-dimensional visual features are extracted from the standardized image, including: the product's color distribution, surface texture information, key visual elements, and geometric features. The extracted features are encoded to generate feature vectors for matching, thus obtaining the visual feature extraction results.

[0025] Specifically, step 2, which involves obtaining key text fields, includes: performing structured parsing of the text information to extract fields such as product name, model, specifications, batch number, and barcode; and using key field extraction methods to filter out core fields that are strongly related to product verification.

[0026] In a specific embodiment of the present invention, the verification of genuine imported toys (LEGO brick sets) is used as an example for illustration: Customers (merchants / purchasers) upload the following product information: Actual product image: A real photo of the front of the toy packaging box (including the LEGO logo, product model QR code, set number, and finished product appearance). Text information (text entered or uploaded): Brand: LEGO Product Model: 42143 Product Name: Ferrari Daytona SP3 Number of parts: 3778 pcs Barcode: 5702017412345 Production batch: S21B The system has completed steps 1-3 (stream splitting, feature extraction, and image-text association verification), and the preliminary verification result is "passed". Proceed to step 4.

[0027] Step 4-1: Call the standardized genuine product feature benchmark library

[0028] The system retrieves the standard authentic benchmark information for this set by searching the benchmark database based on "Brand=LEGO, Model=42143", as shown in Table 1: Table 1

[0029] Step 4-2: Calculate the feature similarity of each dimension

[0030] (1) Visual feature similarity (SvisSvis) is shown in Table 2: Table 2

[0031] Because the finished car in the image is slightly tilted, the car outline score is slightly lower.

[0032] (2) The text feature similarity (Stxt) is shown in Table 3: Table 3

[0033] The batch "S21B" conforms to the LEGO batch coding rules (S + year code + batch), but the benchmark library does not have this specific value, and the semantic similarity is judged to be 0.95 (close to the standard format).

[0034] Step 4-3: Weighted Overall Score

[0035] For the toy category, the system has preset weights (experience values): The weight of the visual feature group is wvis=0.65 (the appearance of the toy and the integrity of the packaging are more important for identifying authenticity). The weight of the text feature group is wtxt = 0.35. The subdivision weights within the visual group (primarily based on packaging and appearance) are shown in Table 4: Table 4:

[0036] The weighted sum within the visual group = 0.30 × 0.98 + 0.25 × 0.94 + 0.25 × 0.96 + 0.20 × 0.89 = 0.294 + 0.235 + 0.240 + 0.178 = 0.947

[0037] Text group: Since the brand / model / barcode / name are all exactly matched, the component quantity is exactly matched, the batch size is slightly lower (0.95), and the overall similarity of the text group is the average of the values ​​of each field: Stxt=(1.00+1.00+1.00+1.00+1.00+0.95) / 6=5.95 / 6=0.9917 Final overall score: Score = wvis × (visual group sum) + wtxt × Stxt Score=0.65×0.947+0.35×0.9917=0.61555+0.3471=0.96265 Step 4-4: Determining the Hierarchical Threshold System preset thresholds: Thigh = 0.95 (perfectly consistent) Tlow=0.70 (clearly incorrect) The calculated score is 0.96265, which is ≥ 0.95. Judgment result: "Completely consistent" → Output "Genuine product verification passed".

[0038] Furthermore, taking Lancôme Absolue Precious Cells Foundation as an example: Suppose the text information in the image uploaded by the customer contains the following original content: Product Name: Lancôme Absolue Precious Cells Foundation Product Model: LANCOME-AB123 Specification: 30ml Color code: PO-01 Production batch number: L2A38 Production date: 2024-08 Valid until: 2027-08 Product barcode: 3605531234567 Anti-counterfeiting code: 1234567890ABCDEF Distributor: [Company Name] Notes: None The system first performs structured parsing on the above text information to obtain the correspondence between field names and field values. Then, the system executes the "key field extraction method" to filter out core fields strongly related to product verification. This filtering process uses a preset rule engine combined with dynamic weight scoring, and the specific steps are as follows: Step 1: Define the importance level of each field. The system has a built-in field importance mapping table for cosmetic categories, as shown in Table 5: Table 5:

[0039] Step 2: Filter core fields based on rules

[0040] The system executes the following judgment logic: Rule 1: If a field has an importance score of 4 or higher, it will be forcibly included in the core field set.

[0041] Rule 2: If a field has an importance score of 3 points and its content is strongly correlated with the standard information in the benchmark library (e.g., specifications and models have a one-to-one correspondence in brand data), then it is also included in the core fields.

[0042] Rule 3: Fields with an importance score of ≤2 (such as production date, distributor, and remarks) are not included in the core fields and are only retained when the record is stored.

[0043] Step 3: Dynamically exclude redundant fields

[0044] If multiple fields exist for the same semantic information (such as "production batch number" and "production date"), the system will prioritize retaining the more accurate field (batch number) and remove the date field to avoid interference.

[0045] Step 4: Output the core field set

[0046] Based on the above rules, the core fields extracted for the Lancôme foundation in this example are shown in Table 6: Table 6:

[0047] (The product name "Lancôme Absolue Precious Cells Foundation" was not included in the core fields due to redundancy with the model number and its low importance; production date, distributor, etc. were not included.)

[0048] Specifically, step 3 includes: The visual feature extraction results are correlated and verified with the key text fields to obtain preliminary verification results, which are divided into pass and fail. If the verification passes, a multi-dimensional comparison will be performed; if it fails, a prompt will be made to resubmit the information. If the verification fails multiple times, a manual verification option will be automatically added.

[0049] Furthermore, the core task of correlation verification is to verify whether the visual features extracted from the physical image and the core fields extracted from the text information uploaded by the customer point to the same product.

[0050] In a specific embodiment of the present invention, a field matching mapping table is established.

[0051] The system pre-sets "vision - text" matching rules for the cosmetics category. Each rule includes: visual feature items, corresponding text fields, matching methods, and whether it is a mandatory match. As shown in Table 7: Table 7

[0052] Mandatory match item: All matches must be successful, otherwise the associated verification fails.

[0053] Non - mandatory match item: When the match fails or cannot be extracted, only the differences are recorded, and it does not affect passing or failing (considering that OCR may miss reading in actual shooting).

[0054] Step 2: Perform matching item by item

[0055] Assume that in this embodiment: Visual features extracted from the physical image: Brand logo: LANCOME Product barcode: 3605531234567 Product model: LANCOME - AB123 (OCR recognition on the bottle body) Specification: 30ml (OCR recognition on the bottle body) Color number: PO - 01 (OCR recognition on the bottle body) Core fields extracted from the text information: Brand name: LANCOME (standardized from "Lancôme") Product barcode: 3605531234567 Product model: LANCOME - AB123 Specification: 30ml Color number: PO - 01 The system compares item by item: Brand → Match successful √ Barcode → Match successful √ Model → Match successful √ Specification → Match successful √ (non - mandatory, record the match) Color number → Match successful √ (non - mandatory, record the match) Step 3: Comprehensive determination Determination rule: All mandatory match items are successful → The preliminary verification result is "passed".

[0056] In this example, all three mandatory items pass, so the result is passed.

[0057] Measurement criteria for passing and failing: All mandatory matching items (brand, barcode, model) were successfully matched (exact comparison or standardized code comparison were consistent). Non-mandatory items may have arbitrary results or not be recognized. Any mandatory matching item failed to match (including: the feature was not extracted visually, or it was extracted but did not match the text field).

[0058] Examples of scenarios that do not apply: Suppose that for the same Lancôme foundation, the model number in the text information uploaded by the customer is incorrect, written as LANCOME-XYZ999, but the OCR on the actual product image correctly identifies the bottle as LANCOME-AB123. All other information is correct.

[0059] Forced match comparison results: Brand: LANCOME vs LANCOME → √ Barcode: 3605531234567 vs 3605531234567 → √ Model: Image extraction LANCOME-AB123 vs text LANCOME-XYZ999 → × (inconsistent) Judgment result: Failed (because the mandatory item "Model" failed to match).

[0060] System response: The client is prompted: "The product model you uploaded (LANCOME-XYZ999) is inconsistent with the model (LANCOME-AB123) identified from the actual product image. Please check and resubmit." Record the failure event and the difference fields.

[0061] If the same customer submits three consecutive applications for the same product and all are rejected due to failure of the association verification, the system will automatically add a "manual verification option" to allow the user to switch to the manual review channel.

[0062] Specifically, the standardized genuine product feature benchmark library is used to store the standard visual features and standard text information of products of various categories, brands and models, and adopts a brand-authorized update mechanism to ensure the authenticity and authority of product information.

[0063] Specifically, the matching verification results include: completely identical, highly similar, and obviously inconsistent; A weighted scoring mechanism is adopted, which sets weights for features in different dimensions and calculates the final matching score.

[0064] In a specific embodiment of the present invention, the matching process includes: 1. Compare the preliminary verification results obtained in step 3 (including visual feature extraction results and text key fields) with the corresponding standard feature information such as category, brand, and model in the standardized genuine product feature benchmark library; 2. Calculate the feature similarity of each dimension respectively: Visual features: Compare the visual features to be verified (such as color texture, logo, structural outline, etc.) with the standard visual feature information in the reference library. Methods such as cosine similarity, Euclidean distance or inner product based on deep features can be used to obtain the visual similarity score Svis (the value range is 0~1). Text information: Compare the key text fields to be verified (such as model number, batch number, anti-counterfeiting code) with the standard text fields in the reference library. Exact match or semantic similarity (such as edit distance, BERT semantic vector similarity) can be used to obtain the text similarity score Stxt; 3. Calculate the final comprehensive matching score according to the preset weight coefficients: Among them, the weights can be dynamically adjusted according to the product category and verification scenario (for example, for products sensitive to appearance differences, increase the visual weight; for products with high requirements for batch consistency, increase the text weight).

[0065] 4. Preset two thresholds Thigh and Tlow (for example, Thigh = 0.95, Tlow = 0.70): If Score≥Thigh, it is determined as "completely consistent", and the verification result outputs "genuine product verification passed".

[0066] If Tlow≤Score<Thigh, it is determined as "highly similar", output "risk warning", and mark the difference items (such as "low texture feature matching degree"), and it is recommended to transfer to manual review.

[0067] If Score<Tlow, it is determined as "significantly inconsistent", output "verification failed", and automatically trigger the manual intervention process.

[0068] Furthermore, the feature matching degree is the original similarity of each dimension, the comprehensive calculation score is their weighted sum, and the matching verification result is the final judgment label output according to which threshold interval the comprehensive score falls into.

[0069] Even further, assume that the customer uploads the same set of Lego products, but due to printing problems on the packaging box image, the color of the LEGO logo is too light, resulting in the logo similarity dropping to 0.82; the color texture drops to 0.85 due to color cast caused by light; other visual items remain unchanged at 0.96 and 0.89. The text fields are exactly the same.

[0070] Recalculate the sum within the visual group: logo 0.82→0.30×0.82 = 0.246 Color texture 0.85→0.25×0.85 = 0.2125 Contour 0.96→0.240 Finished vehicle 0.89→0.178 The sum within the visual group = 0.246 + 0.2125 + 0.240 + 0.178 = 0.8765 Overall Score: Score=0.65×0.8765+0.35×0.9917=0.5697+0.3471=0.9168 Threshold comparison: 0.9168 in the interval [0.70, 0.95) → "highly similar".

[0071] System output: "Risk warning: The logo printing characteristics on the packaging differ from the standard of the genuine product (match rate 82%), and the color and texture deviate from the standard range. It is recommended to transfer it to manual review." Assuming the customer uploads a counterfeit product: text fields such as barcodes and model numbers may be copied and entered correctly, but the actual product image will reveal the problem. Logo misalignment and incorrect font → Similarity 0.45 The color and texture differ significantly from the genuine product (the counterfeit has a greenish tint) → Similarity: 0.40 Packaging outline proportions are out of proportion → Similarity 0.50 The finished vehicle's outline is significantly different → similarity 0.35 The sum of the visual groups = 0.30 × 0.45 + 0.25 × 0.40 + 0.25 × 0.50 + 0.20 × 0.35 =0.135 + 0.100 + 0.125 + 0.070 = 0.430 Overall score = 0.65 × 0.430 + 0.35 × 0.9917 = 0.2795 + 0.3471 = 0.6266 0.6266 < 0.70 → "Clearly inconsistent".

[0072] The system output: "Verification failed: The product's appearance features are seriously inconsistent with the genuine product standard. It has been automatically transferred to manual review."

[0073] Specifically, such as Figure 2 As shown, the "Customer Uploads Information → Categorize by Information Type" section at the top corresponds to the information collection and categorization module; The left side, "Object Image → Image Preprocessing Module → Output Standardized Image → Visual Feature Extraction," corresponds to the image preprocessing module and the visual feature extraction module. The right-hand "Text Information → Structured Parsing → Key Field Extraction" corresponds to the text information structured parsing module, which, together with the left-hand branch, is integrated into the information consistency verification module. "Calling the Standardized Genuine Product Feature Benchmark Library → Multi-Dimensional Comparison Algorithm" corresponds to the Standardized Genuine Product Feature Benchmark Library and Multi-Dimensional Comparison Algorithm modules; The "Retain Records → Periodically Aggregate Data Analysis" section below corresponds to the record retention and data analysis modules.

[0074] A product information intelligent verification system includes: an information collection and classification module, an image and text preprocessing and extraction module, an information consistency verification module, a multi-dimensional comparison algorithm module, a standardized genuine product feature benchmark library, and a record retention and data analysis module; The information collection and classification module is used to receive product information uploaded by customers, automatically identify and classify it into two categories: physical images and text information, and complete data integrity verification. The image and text preprocessing and extraction module is used to adaptively preprocess real object images and extract multi-dimensional visual features, and to perform structured parsing of text information and extract key verification fields. The information consistency verification module is used to perform correlation and matching verification between visual features and key text fields. If the verification passes, it will proceed to the multi-dimensional comparison stage; if it fails, it will prompt the information to be retransmitted or automatically transfer to manual verification. The multi-dimensional comparison algorithm module calls the standardized genuine product feature benchmark library, performs matching calculations between the extracted product features and the features in the benchmark library, and outputs the verification results. The record retention and data analysis module is used to retain and verify complete records and to periodically aggregate and analyze the data.

[0075] In a specific embodiment of the present invention, the information collection and classification module is used to receive product information uploaded by customers, automatically identify the information type, and classify it into two categories: physical images (including photos, real shots, etc.) and text information (including product name, model, specifications, batch, etc.). The system performs integrity checks on uploaded data and automatically prompts for completion when key information is missing.

[0076] This module is used to standardize and distribute input data of various types, laying the foundation for subsequent parallel processing of images and text.

[0077] Furthermore, the image preprocessing module and extraction module specifically include an image preprocessing module, a visual feature extraction module, and a text information structured parsing module: The image preprocessing module includes: 1) Includes adaptive lighting correction: It addresses uneven lighting issues through local histogram equalization, improving shadow details and suppressing highlights; 2) Resolution adjustment: Use super-resolution reconstruction or downsampling algorithms to unify the image to a preset resolution standard; 3) Angle correction: Through perspective transformation and geometric correction, the main body of the product in the tilted image is straightened to a standard viewing angle; 4) It can be expanded to include other preprocessing techniques (such as noise reduction, blur restoration, etc.) to finally output a standardized image.

[0078] Eliminate image defects in non-standard acquisition scenarios to improve the accuracy and stability of subsequent visual feature extraction.

[0079] The visual feature extraction module includes: extracting multi-dimensional visual features from standardized images based on deep convolutional neural networks or classic feature point algorithms. Color and texture features: Extract information such as color distribution and surface texture of the product; Identifying and extracting key visual elements such as product brand identity and anti-counterfeiting labels; Structural contour features: Extracting geometric features such as the product's external contour and component structure; The extracted features are encoded to generate feature vectors that can be used for matching.

[0080] Transforming unstructured physical images into computable feature data provides a core basis for subsequent comparison with a genuine benchmark library.

[0081] The text information structured parsing module includes: 1) Perform structured parsing on the text information uploaded by the customer, and extract fields such as product name, model, specifications, batch, and barcode; 2) By using key field extraction technology, core fields that are strongly related to product verification (such as model, batch, and anti-counterfeiting code) are selected.

[0082] Transforming unstructured text information into structured data provides a standard input for image-text consistency verification and text dimension comparison.

[0083] The information consistency verification module correlates and verifies the visual feature extraction results with key text fields (such as matching the brand logo extracted from the image with the brand name in the text, or matching the product outline in the image with the model number in the text). 2) The verification result is divided into pass and fail: Pass: Transferred to the next stage (multi-dimensional comparison); Failed: Prompt the customer to resubmit the information; if it fails multiple times, the option for manual verification will be automatically added.

[0084] Filtering out inconsistent and erroneous data at the source avoids invalid comparisons and improves overall verification efficiency.

[0085] The purpose of the standardized genuine product characteristic benchmark library is as follows: 1) Store the standard visual characteristics (colors, textures, logos, structural outlines, etc.) and standard textual information (official models, specifications, batch numbers, etc.) of products of various categories, brands, and models; 2) A brand-authorized update mechanism is adopted, and the authenticity and authority of the data are guaranteed through an audit process; 3) Provides an efficient feature retrieval interface, supporting fast invocation and matching.

[0086] It provides a unified and authoritative comparison benchmark, solving the problem of the lack of a universal standard for genuine product data in existing technologies.

[0087] The multi-dimensional comparison algorithm module includes: 1) Match the extracted product feature vectors with a standardized genuine product feature benchmark library, and output three types of comparison results: Completely identical: The feature matching degree reaches the preset threshold, and the product is determined to be genuine and passes the verification. High similarity: Minor differences in features trigger a risk warning, the differences are marked and the case is transferred for manual review; Obvious discrepancy: If the feature matching degree is far below the threshold, the verification is deemed unsuccessful, triggering manual intervention.

[0088] 2) A weighted scoring mechanism is adopted, which sets weights for different dimensional features (visual and textual) and calculates the final matching score.

[0089] The record retention and data analysis module is used to achieve accurate verification through the linkage of images and text, while also providing early warnings for high-risk suspected counterfeit products, balancing accuracy and risk control.

[0090] 1) Retain complete verification records, including core elements such as time, product information, verification results, and discrepancies; 2) Regularly aggregate and analyze retained data to extract key insights: Product information change patterns: Identify the update trends of brand's product models and specifications; Market preferences: Statistics on frequently verified product categories and regional distribution; Counterfeiting Trends: Analyze the differences in products, regions, and characteristics within the cluster of abnormal verifications; 3) Generate visual analysis reports to support business decisions and system optimization.

[0091] This creates a closed loop for verification data, transforming single verification results into long-term value and providing data support for product management and market supervision.

[0092] Furthermore, alternative image preprocessing solutions: In scenarios with limited computing power, bilateral filtering + brightness normalization can be used to replace adaptive lighting correction and resolution adjustment to achieve basic image optimization; Alternatives for visual feature extraction: When deep neural networks are unavailable, the classic SIFT / SURF feature point algorithm can be used to extract local features and complete basic feature matching. Alternative solution for consistency verification: introduce a rule engine + semantic similarity calculation to replace hard matching, and improve compatibility with synonymous text information; Alternative to the comparison algorithm: Cluster analysis combined with anomaly detection can be used to replace threshold determination, allowing for more refined risk classification of highly similar anomaly samples; Alternative deployment architecture: In the absence of a cloud environment, a local server private deployment can be adopted, where the feature library and algorithm engine are deployed within a local area network to ensure data privacy.

[0093] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0094] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent verification of product information, characterized in that, include: Step 1: Obtain product information and perform sorting processing to obtain physical images and text information of the product; Step 2: Perform adaptive preprocessing on the physical image and extract multi-dimensional visual features to obtain the visual feature extraction results; The text information is structured and parsed to extract key verification fields, thus obtaining the key text fields; Step 3: Verify the visual feature extraction results by associating them with the key text fields to obtain preliminary verification results; Step 4: Match the obtained preliminary verification results with the standardized genuine product feature benchmark library, calculate the feature matching degree, and output the verification results based on the feature matching degree.

2. The intelligent verification method for product information according to claim 1, characterized in that, Step 2, the adaptive preprocessing of the physical image, includes: Local histogram equalization is used to address uneven lighting in real-world images, improving detail in dark areas and suppressing highlights. Super-resolution reconstruction or downsampling algorithms are used to unify the physical image to a preset resolution standard; By using perspective transformation and geometric correction, the main body of the product in the tilted real-world image is straightened to a standard perspective; A standardized image is output by employing denoising and blur restoration preprocessing methods.

3. The intelligent verification method for product information according to claim 2, characterized in that, Based on deep convolutional neural networks or classic feature point algorithms, multi-dimensional visual features are extracted from the standardized image, including: product color distribution, surface texture information, key visual elements and geometric features. The extracted features are encoded to generate feature vectors for matching, and the visual feature extraction results are obtained.

4. The intelligent verification method for product information according to claim 1, characterized in that, The process of obtaining key text fields in step 2 includes: performing structured parsing of text information, separating the fields of product name, model, specifications, batch, and barcode, and filtering out core fields that are strongly related to product verification through key field extraction methods.

5. The intelligent verification method for product information according to claim 1, characterized in that, Step 3 includes: The visual feature extraction results are correlated and verified with the key text fields to obtain preliminary verification results, which are divided into pass and fail. If the verification passes, a multi-dimensional comparison will be performed; if it fails, a prompt will be made to resubmit the information. If the verification fails multiple times, a manual verification option will be automatically added.

6. The intelligent verification method for product information according to claim 1, characterized in that, The standardized genuine product feature benchmark library is used to store the standard visual features and standard text information of products of various categories, brands and models, and adopts a brand-authorized update mechanism to ensure the authenticity and authority of product information.

7. The intelligent verification method for product information according to claim 1, characterized in that, The matching verification results include: completely identical, highly similar, and obviously inconsistent; A weighted scoring mechanism is adopted, which sets weights for features in different dimensions and calculates the final matching score.

8. A product information intelligent verification system, characterized in that, include: Information collection and classification module, image and text preprocessing and extraction module, information consistency verification module, multi-dimensional comparison algorithm module, standardized genuine product feature benchmark library, record retention and data analysis module; The information collection and classification module is used to receive product information uploaded by customers, automatically identify and classify it into two categories: physical images and text information, and complete data integrity verification. The image and text preprocessing and extraction module is used to adaptively preprocess real object images and extract multi-dimensional visual features, and to perform structured parsing of text information and extract key verification fields. The information consistency verification module is used to perform correlation and matching verification between visual features and key text fields. If the verification passes, it will proceed to the multi-dimensional comparison stage; if it fails, it will prompt the information to be retransmitted or automatically transfer to manual verification. The multi-dimensional comparison algorithm module calls the standardized genuine product feature benchmark library, performs matching calculations between the extracted product features and the features in the benchmark library, and outputs the verification results. The record retention and data analysis module is used to retain and verify complete records and to periodically aggregate and analyze the data.

Images