Cross-border e-commerce infringement goods patrol method and system based on image recognition

Abstract

This invention discloses a method and system for inspecting infringing goods in cross-border e-commerce based on image recognition, involving the application of image recognition technology in the field of cross-border e-commerce. The system includes an OpenClaw intelligent agent, a feature library storage module, an AI visual large model module, and an early warning unit. The OpenClaw intelligent agent collects product data across platforms in compliance with regulations and performs deduplication filtering. The feature library stores standardized visual feature vectors of genuine and infringing products. The AI ​​visual large model module preprocesses product images, extracts features, and compares them with the feature library for similarity, determining risk through dual thresholds. The early warning unit pushes structured early warning information through multiple channels. The method achieves automated inspection through a closed-loop process of parameter configuration, data collection, feature extraction, risk determination, and early warning push. This solution solves the problems of low efficiency, incomplete coverage, high missed detection rate, and delayed response of traditional manual inspections, improving the comprehensiveness, efficiency, and accuracy of infringement inspections and reducing the business risks for merchants.

Description

Technical Field

[0001] This invention relates to the application of image recognition technology in the field of cross-border e-commerce, and in particular to a method and system for inspecting infringing goods in cross-border e-commerce based on image recognition. Background Technology

[0002] During the rapid development of cross-border e-commerce, issues such as product image theft, trademark infringement, and design counterfeiting have become increasingly prominent. Because cross-border e-commerce covers multiple countries and regions, involving numerous e-commerce platforms and sites such as Amazon, AliExpress, Shopee, and Lazada, the number of products sold is enormous, the frequency of new product launches is high, and the pace of iteration is rapid. Traditional methods relying on manual inspection of each platform, category, and product suffer from drawbacks such as low inspection efficiency, limited coverage, high missed detection rates, and delayed response times.

[0003] If merchants fail to promptly detect infringement of their products, or if their products are deemed infringing, they are highly susceptible to penalties such as product removal, store deductions, and store closures, resulting in significant economic losses and damage to their brand reputation. Currently, the industry generally lacks a comprehensive, automated, and highly accurate product infringement inspection and early warning mechanism that can be implemented across platforms, making it difficult to meet the actual needs of cross-border e-commerce merchants for intellectual property protection and compliant operations. Summary of the Invention

[0004] This invention aims to overcome the problems of slow, comprehensive, and accurate inspection and early warning of image theft and brand infringement in cross-border e-commerce scenarios in existing technologies. It solves the pain points of low efficiency, incomplete coverage, high missed detection rate, and slow response of manual inspection, and realizes automated inspection and intelligent identification of infringement of products on multiple platforms and sites, effectively reducing the risk of store penalties and store closures for merchants due to infringement or being infringed.

[0005] To address the aforementioned technical problems, the present invention provides a cross-border e-commerce infringing goods inspection system based on image recognition, comprising a data acquisition unit, an image recognition unit, and an early warning unit. Its distinguishing feature is that it further includes an OpenClaw intelligent agent, a feature library storage module, and an AI vision large model module. The OpenClaw intelligent agent establishes a compliant data interaction channel with the open API interface of at least one cross-border e-commerce platform. The interface interaction follows the platform's data access protocol and calling specifications. It is used to traverse each site of the target cross-border e-commerce platform according to a preset inspection strategy to obtain image data of the target product and related product information. The related product information includes product ID, product title, product category, store information, and product link. The feature library storage module is used to pre-build and store a genuine product feature library and an infringing product feature library. The genuine product feature library stores standardized visual feature vectors of genuine products and brand logos, and the infringing product feature library stores standardized visual feature vectors of known infringing products. The feature vectors are all fixed-dimensional numerical vectors generated by feature extraction algorithms. The AI ​​vision big model module establishes a bidirectional communication connection with the OpenClaw agent and the feature library storage module, respectively, to receive product image data transmitted by the OpenClaw agent, extract multi-dimensional visual features from the product images to generate a standardized feature vector to be tested, compare the standardized feature vector to be tested with the feature vectors in the genuine product feature library and the infringing product feature library, and determine whether the current product has the risk of infringement or being infringed based on the comparison results and a preset judgment threshold. The warning unit is electrically connected to the AI ​​vision big model module and is used to trigger the warning mechanism when the AI ​​vision big model module determines that there is a risk of infringement or being infringed, and to push structured warning information to the target user. The structured warning information is a set of feature information in a standardized format.

[0006] As a preferred implementation, the OpenClaw agent supports multi-node distributed concurrent access. Each node allocates inspection tasks according to a load balancing strategy. The agent executes timed inspection tasks according to a preset inspection cycle. The inspection cycle supports a combination configuration of global full-category periodic inspection and high-frequency inspection of key risk categories. Global full-category inspection and key risk category inspection are independent inspection task processes, and the inspection time and inspection frequency can be configured separately.

[0007] In a preferred embodiment, the OpenClaw agent has a built-in data filtering unit, which is a data cleaning and processing module. During the process of acquiring product information, it performs multi-level deduplication and redundancy filtering on duplicate or highly similar product data in the product list data. The deduplication filtering is based on the unique identifier of the product, image feature values, and text feature values. Only one image acquisition and feature comparison process is performed for the same product. Filtered invalid product data does not enter the subsequent feature extraction process.

[0008] In a preferred embodiment, before extracting features from the product image, the AI ​​vision large model module first performs a standardized preprocessing operation on the input product image. The preprocessing operation is based on an image processing algorithm and includes image size adjustment, pixel distribution normalization, and image noise filtering. The image size is adjusted to a tensor format with a fixed resolution, the pixel distribution normalization maps the pixel values ​​to a preset value range, and the image noise filtering uses a filtering algorithm to eliminate interference noise in the image.

[0009] In a preferred embodiment, the AI ​​visual large model module adopts a deep learning convolutional neural network architecture. The convolutional neural network includes convolutional layers, pooling layers, fully connected layers, and residual connection structures, which are used to extract global visual features and local key features from product images in multiple dimensions. The features include the product's color features, texture features, shape and outline features, and brand logo features. The global visual features are the overall visual attribute features of the product, and the local key features are the iconic visual attribute features of the product and brand.

[0010] As a preferred implementation, the AI ​​vision large model module adopts a high-dimensional feature vector similarity matching algorithm, combined with the fast retrieval capability of the vector database. The vector database is equipped with an approximate nearest neighbor retrieval index to realize full matching and comparison between the standardized feature vector to be tested and the feature library of genuine products and the feature library of infringing products. The matching and comparison process is to calculate the similarity between the feature vector to be tested and all feature vectors in the feature library one by one, and output the highest similarity value and the corresponding feature vector.

[0011] As a preferred implementation, the AI ​​visual large model module has a built-in infringement judgment threshold, which is a preset similarity value. When the highest similarity between the standardized feature vector to be tested and the feature vector in the feature library of infringing products is higher than the preset infringement judgment threshold, the current product is judged to be a suspected infringing product, and the judgment result is synchronously transmitted to the early warning unit and triggers the corresponding early warning process.

[0012] In a preferred embodiment, the AI ​​visual large model module has a built-in genuine matching threshold, which is a preset similarity value. When the similarity between the standardized feature vector to be tested and the corresponding feature vector in the genuine product feature library is lower than the preset genuine matching threshold, and the matching degree with the infringing product feature library meets the preset risk condition, it is determined that the corresponding genuine product is at risk of being infringed. The preset risk condition is that the similarity between the feature vector to be tested and the infringing product feature library is higher than the preset infringement judgment threshold.

[0013] In a preferred embodiment, the early warning unit has a built-in multi-channel push interface, including an SMS gateway interface, an email sending interface, and a platform intra-site message interface. The structured early warning information pushed by the early warning unit includes the name of the risky product, the platform site to which it belongs, the product link address, the type of infringement, and the inspection and identification time. The early warning unit supports pushing early warning information through at least one of SMS, email, and platform intra-site messages, and the early warning information from different push channels has the same standardized format content.

[0014] This invention also proposes the image recognition-based method for inspecting infringing goods in cross-border e-commerce, which is applied to the aforementioned image recognition-based system for inspecting infringing goods in cross-border e-commerce. The method includes the following steps: S1. Complete the system parameter configuration, which includes the OpenClaw agent's inspection platform, inspection cycle, interface parameters, and push parameters of the early warning unit. At the same time, construct and initialize the genuine product feature library and the infringing product feature library. The genuine product feature library stores standardized visual feature vectors of genuine products and brand logos, and the infringing product feature library stores standardized visual feature vectors of known infringing products. Feature library initialization includes feature data entry and feature vector index construction. S2. Establish a compliant data interaction channel with the OpenClaw intelligent agent and the open API interfaces of multiple cross-border e-commerce platforms. According to the preset inspection strategy, traverse each site of the target cross-border e-commerce platform, obtain product list data through API interface calls, and after deduplicating and filtering the product list data, obtain the image data of the target product and related product information. Establish a one-to-one correspondence between the related product information and the image data. S3. The AI ​​vision big model module performs preprocessing and feature extraction operations on the acquired product images in sequence. First, the product images are processed into standardized data that meet the model input requirements. Then, the convolutional neural network is used to extract multi-dimensional visual features from the standardized images to generate a fixed-dimensional standardized feature vector to be tested. S4. Input the standardized feature vector to be tested into the genuine product feature library and the infringing product feature library respectively. Calculate the similarity between the feature vector to be tested and the feature vector in the feature library through the similarity matching algorithm, obtain the highest similarity value, and compare the calculated similarity with the preset judgment threshold to determine whether the current product has the risk of infringement or being infringed. The preset judgment threshold includes the infringement judgment threshold and the genuine matching threshold. S5. When an infringement risk or the risk of being infringed is determined, the AI ​​vision big model module sends a trigger signal to the early warning unit. After receiving the signal, the early warning unit triggers the early warning mechanism and pushes structured early warning information to the target user according to the preset push channel and push format. The early warning information includes the core identification information of the risky product and the judgment result information.

[0015] Compared with related technologies, the image recognition-based cross-border e-commerce infringing goods inspection system provided by this invention has the following beneficial effects: 1. This solution uses OpenClaw intelligent agents to build a multi-node distributed architecture, supporting cross-platform compliance data collection and concurrent access by multiple nodes. It also configures independent task processes for global full-category periodic inspections and high-frequency inspections of key risk categories, achieving full coverage inspections of product categories across multiple platforms and sites. Furthermore, it rationally allocates inspection tasks through load balancing strategies to ensure the stability and efficiency of the inspection process. This replaces the traditional manual platform-by-platform and category-by-category inspection method, significantly improving the overall execution efficiency of cross-border e-commerce infringing product inspections and breaking through the scope and efficiency limitations of manual inspections.

[0016] 2. This solution utilizes the multi-level data filtering unit built into the OpenClaw intelligent agent to deduplicatize the collected product data based on unique product identifiers, text features, and image features, eliminating invalid and redundant data. Simultaneously, the AI ​​vision large model module performs standardized preprocessing and multi-dimensional visual feature extraction on product images. Combined with a high-dimensional feature vector similarity matching algorithm and a dual-threshold judgment mechanism, it achieves accurate judgment of infringing products and the risk of infringement of genuine products, avoiding invalid data from interfering with the recognition results, reducing the false negative rate and false positive rate, and improving the accuracy of infringing product identification.

[0017] 3. This solution constructs a fully automated closed loop from product data collection, feature extraction, risk assessment to early warning push. After the AI ​​vision big model module completes the risk assessment, it immediately sends a trigger signal to the early warning unit. The early warning unit realizes the rapid push of structured early warning information through multi-channel push interfaces, while monitoring the push status and executing a retry mechanism to ensure that the early warning information reaches the target users in a timely manner. This allows merchants to quickly know about the infringement risks and take countermeasures, effectively reducing losses such as store penalties and store closures caused by delayed inspection response, and protecting the brand reputation and economic interests of merchants.

[0018] In summary, this solution, through the collaborative operation of the OpenClaw intelligent agent, AI vision large model module, feature library storage module, and early warning unit, constructs a cross-platform, automated, and high-precision cross-border e-commerce infringing product inspection system. It completely solves the core problems of traditional manual inspections, such as low efficiency, incomplete coverage, high missed detection rate, and delayed response. This achieves full coverage, high efficiency, and precision in infringing product inspection. Simultaneously, through a fully automated risk warning mechanism, it enables rapid identification and notification of infringement risks, effectively reducing various losses incurred by merchants due to infringement or being infringed upon, and meeting the actual needs of cross-border e-commerce merchants for intellectual property protection and compliant operation. Attached Figure Description

[0019] Figure 1 A diagram showing the connection relationships of the core modules of the image recognition-based cross-border e-commerce infringing goods inspection system provided by this invention. Figure 2 This is a flowchart of the infringing goods inspection method provided by the present invention; Figure 3 This is a block diagram illustrating the principle of the AI ​​vision large model module provided by the present invention. Figure 4 The OpenClaw intelligent agent inspection logic block diagram provided by this invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] The image recognition-based cross-border e-commerce infringing goods inspection system and method disclosed in this embodiment rely on the OpenClaw intelligent agent to realize cross-platform compliant data collection and combine AI vision large model to realize high-precision infringing feature identification and comparison. It solves the pain points of low efficiency, incomplete coverage, insufficient identification accuracy and high data redundancy in the cross-border e-commerce scenario. The following provides a complete and sufficient disclosure of the implementation method of the system and method.

[0023] Please refer to the following: Figures 1-4 The image recognition-based cross-border e-commerce infringing goods inspection system includes a data acquisition unit, an image recognition unit, an early warning unit, an OpenClaw intelligent agent, a feature library storage module, and an AI vision large model module; The OpenClaw intelligent agent establishes a compliant data interaction channel with the open API interface of at least one cross-border e-commerce platform. The interface interaction follows the platform's data access protocol and calling specifications. It is used to traverse each site of the target cross-border e-commerce platform according to the preset inspection strategy to obtain the image data of the target product and related product information. The related product information includes product ID, product title, product category, store information, and product link. The feature library storage module is a high-dimensional vector storage architecture used to pre-build and store the feature library of genuine products and the feature library of infringing products. The feature library of genuine products stores standardized visual feature vectors of genuine products and brand logos, while the feature library of infringing products stores standardized visual feature vectors of known infringing products. All feature vectors are fixed-dimensional numerical vectors generated by feature extraction algorithms. The AI ​​vision big model module establishes bidirectional communication connections with the OpenClaw agent and the feature library storage module, respectively. The communication link adopts encrypted transmission. It is used to receive product image data transmitted by the OpenClaw agent, extract multi-dimensional visual features from the product images to generate standardized feature vectors to be tested, compare the standardized feature vectors to be tested with feature vectors in the genuine product feature library and the infringing product feature library, and determine whether the current product has the risk of infringement or being infringed based on the comparison results and the preset judgment threshold. The early warning unit is electrically connected to the AI ​​vision big model module. When the AI ​​vision big model module determines that there is a risk of infringement or being infringed, it triggers the early warning mechanism and pushes structured early warning information to the target user. The structured early warning information is a set of feature information in a standardized format.

[0024] It should be further explained that the OpenClaw intelligent agent is an automated data collection and task scheduling intelligent agent built on a distributed architecture. It has four core functions: cross-platform API compliance adaptation, multi-node concurrent task execution, dynamic configuration of inspection strategies, and pre-filtering of data. It establishes a compliant data interaction channel with the open API interface of at least one cross-border e-commerce platform, and the interface interaction follows the platform's data access protocol and calling specifications throughout the process.

[0025] In this invention, the OpenClaw agent adopts a master-slave node distributed architecture, comprising one master scheduling node and N slave execution nodes. The master scheduling node and the slave execution nodes establish a bidirectional communication link through a long TCP connection. The master scheduling node is responsible for the splitting, allocation, and full lifecycle status monitoring of inspection tasks, while the slave execution nodes are responsible for compliant interaction with the open API interfaces of the corresponding cross-border e-commerce platforms and the collection and execution of product data. It supports distributed concurrent access by multiple nodes, enabling parallel inspections across multiple platforms and sites. It should be noted that the master-slave architecture decouples task scheduling and data collection, avoiding single-node performance bottlenecks. The long TCP connection ensures low latency and stability in communication between nodes, and the concurrent access of multiple nodes significantly improves the efficiency of product data collection across platforms and sites, ensuring that the business needs of multi-platform inspections in cross-border e-commerce are met. The master scheduling node has a built-in load balancing scheduling module, which uses a weighted round-robin load balancing algorithm to dynamically allocate inspection tasks, ensuring that the task load of each node matches its own computing power and network capabilities. The algorithm implementation logic is as follows: The master scheduling node collects the running status parameters of each slave execution node in real time. The running status parameters include node CPU utilization, memory usage, network bandwidth utilization, API request success rate, and task execution latency. Based on the running status parameters of each slave execution node, the task allocation weight of each slave execution node is dynamically updated using a weighted value calculation formula. The weighted value calculation formula is as follows: ;in, Assign weights to the i-th task from the execution node. The overall performance score for the i-th execution node is given. The overall performance score is given to the j-th slave execution node, where n is the total number of slave execution nodes; overall performance score The comprehensive performance score is obtained by normalizing and weighting the node's running status parameters. It is positively correlated with the node's available computing power and network stability, and negatively correlated with task execution latency. It should be noted that the node running status parameters collected in real time must be collected at a rate of seconds to ensure the real-time performance of the weight calculation. The parameters are normalized and then weighted and summed after weighted value calculation. Computing power and network stability are the core weighting factors to ensure that the task allocation is highly adapted to the actual capabilities of the nodes. The main scheduling node breaks down the global inspection task into multiple independent sub-task units. Each sub-task unit corresponds to the inspection task of a single category of products on a single site of the target platform. The sub-task units are allocated to the corresponding slave execution nodes according to the weight ratio of each slave execution node to avoid node overload or resource idleness and ensure the stability and execution efficiency of concurrent inspection. The inspection cycle of the OpenClaw intelligent agent supports configurable management, and supports a combination of global, full-category cycle inspections and high-frequency inspections of key risk categories. The specific implementation method is as follows: Global and category-wide periodic inspection: For all product categories of the target cross-border e-commerce platform, a fixed global inspection cycle is set, and a full-scale traversal inspection task is launched at preset time intervals to achieve a full-coverage investigation of infringement risks for all products on the platform. High-frequency inspections of key risk categories: For categories with a high history of infringement, categories that are key protected by brands, and high-value product categories, a high-frequency inspection cycle is set up independently of the overall inspection. The inspection frequency is higher than that of the overall category inspection, so as to achieve key monitoring and rapid response of high-risk product categories. Among them, categories with a high history of infringement are specifically determined based on statistical analysis of infringing product data from the platform's historical inspections. Categories that are key protected by brands are configured and designated by the brands themselves. High-value product categories are specifically determined based on a comprehensive assessment of product pricing, market value, and industry category value rating system. Global category-wide inspections and key risk category inspections are independent task processes, each configured with its own task execution queue, inspection time window, and API call frequency limit. The two types of tasks execute in parallel without interference, preventing high-frequency inspection tasks from affecting the stable execution of the global category-wide inspection task. A configurable inspection cycle balances comprehensive coverage of all products across the platform with focused monitoring of high-risk categories. Independent task processes and resource configurations prevent interference between tasks, ensuring comprehensive inspections while improving response speed to high-incidence infringement scenarios, adapting to the differentiated infringement risks across cross-border e-commerce product categories. The configurable management of this inspection cycle and the independent configuration and parallel execution of different types of inspection tasks are conventional task scheduling and configuration methods in existing technologies and are not innovative features of this solution. This solution only adapts and applies these conventional methods.

[0026] In this invention, the OpenClaw intelligent agent has a built-in data filtering unit, which is a multi-level data cleaning and processing module based on multi-feature fusion. During the product data collection process, the acquired product list data undergoes three-level deduplication and redundancy filtering, retaining only valid product data for subsequent feature extraction and comparison processes. The specific implementation steps are as follows: The first level is hard deduplication based on the unique identifier of the product: For the obtained product list data, the unique product ID of each product data is extracted. Using the product ID as the unique primary key, the duplicate product IDs of the same product data under the same platform and the same site are directly deduplicated, and only one valid product data is retained, filtering out redundant data that are repeatedly entered. The second level is text feature-based deduplication: For the product data after the first level of deduplication, text features of product titles, product categories, and store information are extracted. The TF-IDF algorithm is used to calculate text feature vectors, and the similarity of text feature vectors of different product data is calculated using cosine similarity. When the text similarity of two product data is higher than the preset text deduplication threshold, they are judged as highly similar products, and only one product data is retained to enter the next level of filtering to avoid duplicate processing of duplicate products. The TF-IDF calculation formula for text feature vectors is: TF−IDF(t,d,D)=TF(t,d)×IDF(t,D). Where TF(t,d) is the term frequency of term t in product text d, and IDF(t,D) is the inverse document frequency of term t in the full product text corpus D, which is used to measure the distinguishability of terms to product text; The third level is precise deduplication based on image features: For the product data after the first two levels of deduplication, the perceptual hash features of the main product image are extracted, and the Hamming distance is used to calculate the feature differences between different main product images. When the Hamming distance between two main product images is less than the preset image deduplication threshold, they are determined to be products with duplicate or highly similar images. Only one product data is retained to perform subsequent image acquisition and feature comparison processing, and the remaining similar product data is directly filtered out and does not enter the subsequent feature extraction process. After the above three-level deduplication filtering process, the redundancy of product data is effectively reduced, avoiding duplicate feature extraction and comparison calculations for duplicate and similar products, reducing system computing power and network resource consumption, and improving the overall processing efficiency of subsequent infringement identification processes.

[0027] In this invention, the AI ​​visual large-scale model module is specifically a visual feature extraction and infringement intelligent judgment module built based on a deep learning convolutional neural network. It establishes encrypted bidirectional communication links with the OpenClaw intelligent agent and the feature library storage module, respectively. The communication links adopt the TLS 1.3 encrypted transmission protocol. The TLS 1.3 encrypted transmission protocol covers end-to-end encryption and authentication of data transmission, effectively preventing the theft and tampering of core data such as product images and feature vectors during transmission, meeting the security and compliance requirements of cross-border e-commerce data transmission. The core functions of the AI ​​visual large-scale model module include image standardization preprocessing, multi-dimensional visual feature extraction, high-dimensional vector similarity matching, and intelligent infringement risk judgment. Specific implementation methods are as follows: Before performing feature extraction on product images, the AI ​​vision large model module first performs a standardization preprocessing operation on the input product images to eliminate the impact of image size, pixel distribution, and noise interference on the accuracy of feature extraction. This converts the product images into standardized tensor data that meets the model's input requirements. Standardization preprocessing unifies the input format of product images, eliminates interference from various non-feature factors, and ensures the consistency and accuracy of convolutional neural network feature extraction, laying a data foundation for subsequent high-precision similarity matching. The specific steps for preprocessing are as follows: Step 1, Image Resizing: The input product images are resized and padded using bicubic interpolation to resize images with different resolutions and aspect ratios into fixed-resolution three-channel RGB images. These images are then converted to tensor formats that meet the model's input requirements, ensuring consistent dimensions across all input images and satisfying the input specifications for convolutional neural networks. The pixel value calculation formula for bicubic interpolation is as follows: In the formula, This represents the interpolated pixel value of the target image at coordinates (x, y). These represent the neighboring pixel indices in the horizontal and vertical directions of the source image, respectively (both ranging from 0 to 3). This represents the original pixel value of the neighboring pixel at coordinates (p, q) in the source image. This represents the bicubic interpolation kernel function, used to perform smooth weighted calculations of neighboring pixels. This represents the weighted component of the bicubic interpolation kernel function in the horizontal direction, used to measure the contribution of the horizontal neighboring pixels of the source image to the target pixel (x,y). This represents the weighted component of the bicubic interpolation kernel function in the vertical direction, used to measure the contribution of the vertical neighbor pixels of the source image to the target pixel (x,y). Step 2, Pixel Distribution Normalization: Perform pixel value normalization on the resized image tensor, mapping the RGB three-channel pixel values ​​of the image from the original 0-255 value range to a preset standardized value range. This eliminates the impact of differences in pixel value magnitudes on model feature extraction. The normalization calculation formula is: ,in, For the normalized pixel tensor, The original pixel tensor is μ, the preset pixel mean is σ1, and the preset pixel standard deviation is σ1. The mean and standard deviation are set based on the pre-trained statistical values ​​of a large-scale image dataset to ensure the fit with the pre-trained weights of the model. Step 3, Image Noise Filtering: A Gaussian filtering algorithm is used to filter noise in the normalized image, eliminating Gaussian noise, salt-and-pepper noise, and other interfering noise while preserving the core visual features of the product, thus improving the accuracy of subsequent feature extraction. The formula for Gaussian filtering is: Where G(x,y) is the weight value of the two-dimensional Gaussian kernel function at coordinates (x,y), and σ2 is the standard deviation of the Gaussian kernel. By performing convolution operation between the Gaussian kernel and the image pixel matrix, the image is smoothed and denoised, filtering out interference noise while retaining key visual features such as the outline and texture of the product.

[0028] In this invention, the AI ​​vision large model module adopts a deep learning convolutional neural network architecture based on residual connections. The convolutional neural network includes an input layer, multiple convolutional modules, pooling layers, fully connected layers, an output layer, and a residual connection structure across convolutional modules. This effectively solves the gradient vanishing problem during deep network training and enables multi-dimensional and accurate extraction of global visual features and local key features from product images. The specific implementation method is as follows: Each convolutional module of a convolutional neural network contains two consecutive convolutional layers, one batch normalization layer, one ReLU activation function layer, and one max pooling layer. The convolutional layers use convolutional kernels of different sizes to perform sliding window convolution operations on the input image tensor to extract visual features at different scales. The batch normalization layer normalizes the feature map output by the convolution, accelerating model convergence and improving the stability of feature extraction. The ReLU activation function layer introduces non-linear expressive power into the model through non-linear activation, fitting complex visual feature mapping relationships. The max pooling layer downsamples the feature map, reducing the feature dimensionality while retaining key feature information. The residual connection structure is placed between adjacent convolutional modules, directly skipping the input feature map of the preceding convolutional module to the output of the following convolutional module. The residual is then added element-wise with the output feature map of the following convolutional module before being input to subsequent network layers. The residual calculation expression is as follows: ,in, This represents the final output feature of the residual module. X represents the residual mapping features learned by the convolutional layers in the residual module, and X represents the input features of the residual module. Through the residual connection structure, it is ensured that feature information can be effectively transferred as the number of layers in the deep network increases, avoiding gradient vanishing and improving the model's ability to extract fine-grained visual features. The convolutional neural network performs multi-dimensional feature extraction on the preprocessed product images. The extracted visual features include the product's color features, texture features, shape and contour features, and brand logo features, among which: Global visual feature extraction: Through the deep convolutional module of the network, the overall visual attributes of the product are encoded to generate global visual features that represent the overall appearance, color distribution and overall outline of the product, reflecting the overall visual attributes of the product. Local Key Feature Extraction: Using a spatial attention mechanism, key areas in product images, such as brand logos, design patents, and iconic patterns, are weighted and their proportion is increased. This extracts local key features representing the product's core identity and iconic design. The weight calculation formula for the spatial attention mechanism is as follows: ;in, Let F be the spatial attention weight map corresponding to the input feature map, and σ3 be the sigmoid activation function. For 7×7 size convolution operations, The average pooling result for feature map F, To achieve the maximum pooling result of feature map F, the key region features are enhanced by element-wise multiplication of the spatial attention weight map with the original feature map, thereby improving the extraction accuracy of key infringing features such as brand logos and patent design points. Finally, the convolutional neural network fuses and reduces the dimensionality of the extracted global visual features and local key features through a fully connected layer, and outputs a standardized feature vector of fixed dimension to be tested. The feature vector is a floating-point numerical vector that can accurately represent the core visual attributes of the product and is used for subsequent similarity matching and infringement determination. It should be further explained that the spatial attention mechanism generates an attention weight map by performing pooling and convolution operations on the feature map of the product image. This gives higher weights to key infringing areas such as brand logos, making the model more focused on the core visual features of infringement determination and weakening the interference of irrelevant background features.

[0029] In this invention, the AI ​​vision large model module employs a high-dimensional feature vector similarity matching algorithm, combined with the fast retrieval capability of the vector database in the feature library storage module, to achieve full and rapid matching and comparison of the standardized feature vector to be tested with the feature libraries of genuine and infringing products. The specific implementation method is as follows: The feature library storage module adopts a high-dimensional vector storage architecture and is equipped with a high-performance vector database. The vector database constructs an approximate nearest neighbor retrieval index for standardized feature vectors in the genuine product feature library and the infringing product feature library. The index adopts an inverted file system combined with an IVF-PQ index structure with product quantization. The high-dimensional feature vectors are clustered, bucketed and quantized in blocks, transforming the full vector retrieval into bucket retrieval, which greatly reduces the computational load of vector retrieval. It achieves efficient retrieval of high-dimensional vectors at the expense of minimal precision, and is suitable for the storage and retrieval needs of massive feature library data. The specific implementation steps for similarity matching and comparison are as follows: Vector retrieval trigger: The AI ​​vision big model module will input the generated standardized feature vector to be tested into the vector retrieval interface corresponding to the genuine product feature library and the infringing product feature library respectively, triggering a full-database approximate nearest neighbor retrieval; Fast candidate set recall: The vector database uses IVF-PQ indexing to quickly recall the Top-K candidate feature vectors that are closest to the feature vector to be tested, generating a candidate feature set and narrowing the scope of subsequent accurate similarity calculation; Precise similarity calculation: For each feature vector in the recalled candidate feature set, the cosine similarity algorithm is used to calculate the similarity value between the normalized feature vector to be tested and the candidate feature vectors. The formula for calculating cosine similarity is: ;in, The feature vector to be tested With candidate feature vectors The cosine similarity is in the range of [0,1]. The closer the value is to 1, the higher the visual feature similarity between the two feature vectors, and the higher the visual overlap of the corresponding products. These are the dot products of the feature vector to be tested and the candidate feature vectors, respectively. and These are the L2 norms of the feature vector to be tested and the candidate feature vector, respectively. Matching result output: Sort the similarity values ​​of all candidate feature vectors in descending order, output the highest similarity value, as well as the feature library type to which the feature vector with the highest similarity belongs, and the association information of the corresponding genuine product / known infringing product, thus completing the full matching comparison process.

[0030] In this invention, the AI ​​vision large model module incorporates a dual-threshold judgment mechanism, including a preset infringement judgment threshold and a genuine product matching threshold. Based on the comparison results of similarity matching results and the dual thresholds, it achieves intelligent judgment of the risk of product infringement and the risk of genuine products being infringed. The specific implementation method is as follows: The infringement determination threshold is a preset similarity value used to determine the degree of feature matching between the product under test and known infringing products. The specific determination logic is as follows: When the highest similarity value between the standardized feature vector to be tested and the feature vector in the feature library of infringing products is higher than the preset infringement judgment threshold, it is determined that the visual features of the current product to be tested are highly overlapping with those of known infringing products, which conforms to the visual feature rules of infringing products, and the current product is determined to be a suspected infringing product. When the highest similarity value between the standardized feature vector to be tested and the feature vector in the infringing product feature library is lower than or equal to the preset infringement judgment threshold, it is determined that the current product does not have a highly matching feature with the known infringing products, and the infringing product judgment process is not triggered. After completing the identification of suspected infringing products, the AI ​​vision big model module will synchronously transmit the identification results, the corresponding product association information, and similarity matching data to the early warning unit, triggering the corresponding infringement risk early warning process; The authenticity matching threshold is a preset similarity value used to determine the degree of feature matching between the tested product and the corresponding authentic product. The determination of the risk of infringement of authentic products adopts a dual threshold combination judgment logic, specifically: First, a genuine feature matching verification is performed: when the similarity value between the standardized feature vector to be tested and the feature vector of the corresponding genuine product in the genuine product feature library is lower than the preset genuine matching threshold, it is determined that the product to be tested has not obtained the genuine authorization to use the visual features and does not have the compliant visual feature authorization of genuine products. Simultaneous execution of infringement feature matching verification: When the highest similarity value between the above-mentioned standardized feature vector to be tested and the feature vector in the infringing product feature library is higher than the preset infringement judgment threshold, the preset risk condition is met; When both of the above verification conditions are met, the product under test is determined to be an infringing product that counterfeits the corresponding genuine product, and the corresponding genuine product is at risk of being infringed. The expression for the above dual-threshold combination judgment logic is: ; in, This is the result of the risk assessment for the infringement of genuine products. A value of true indicates that there is a risk of infringement, and a value of false indicates that there is no risk of infringement. The similarity between the feature vector to be tested and the feature vector of the corresponding genuine product is denoted as . The threshold for matching genuine content; The highest similarity between the feature vector to be tested and the feature library of infringing products. ∧ represents the threshold for infringement determination; ∧ is the logical AND operator, which determines the result as true only when both conditions are met simultaneously. After completing the risk assessment of the infringement of genuine products, the AI ​​vision big model module will synchronously transmit the assessment results, the corresponding genuine product information, the infringing product association information, and the similarity matching data to the early warning unit, triggering the corresponding infringement risk early warning process. It should be noted that by setting up a dual-threshold judgment mechanism, through dual threshold verification of infringement and genuine products, the risk of suspected infringing products and genuine products being infringed can be accurately distinguished and judged, avoiding the problem of excessively high false judgment rate of single threshold judgment, and improving the accuracy and reliability of infringement risk judgment.

[0031] In this invention, the early warning unit is electrically connected to the AI ​​visual large model module and adopts an event-driven triggering mechanism. Upon receiving a risk assessment trigger signal sent by the AI ​​visual large model module, it immediately executes the early warning information generation and multi-channel push process. The specific implementation method is as follows: The early warning unit has a built-in standardized early warning information generation module and a multi-channel push interface module, including: The standardized early warning information generation module is used to generate structured early warning information in a standardized format based on the risk judgment results and product association information transmitted by the AI ​​vision big model module. The structured early warning information is a set of feature information with fixed fields, including the following core fields: risky product ID, risky product name, cross-border e-commerce platform and site, product link address, store entity information, infringement type (suspected infringement / genuine product being infringed), highest similarity value, inspection and identification time, and association information of the corresponding genuine product / known infringing product. The multi-channel push interface module has built-in standardized push interfaces of various types, including SMS gateway interface, email sending interface, and platform internal message interface. Each interface follows the corresponding communication protocol and interface specification, supporting seamless adaptation with third-party push services. It should be noted that the structured warning information is generated using a fixed field format to ensure information consistency across different push channels. Each push interface follows industry-standard communication protocols, enabling integration with third-party push services without additional development, thus improving system compatibility and ease of use. The SMS gateway interface supports integration with the SMS gateways of the three major telecom operators and third-party SMS service platforms to enable batch push of warning SMS messages; the email sending interface follows the SMTP email transmission protocol and supports the push of warning emails with attachments, which can include screenshots of risky products, similarity matching reports, and other supporting materials; the platform internal messaging interface supports integration with the merchant backend internal messaging system of the corresponding cross-border e-commerce platform and the user-end internal messaging system of this system to enable real-time push of warning information within the platform.

[0032] The push execution logic of the early warning unit is as follows: Warning Trigger: Upon receiving a risk assessment trigger signal from the AI ​​vision big model module, the warning process is immediately activated, and the full information and assessment results of the corresponding risky goods are read. Warning information generation: Based on a preset standardized format, structured warning information is generated. For different push channels, warning content adapted to the format of the corresponding channel is generated. The core warning information of all channels remains consistent to ensure the uniformity of information. Push channel execution: Based on the push channel pre-configured by the user, push warning information to the designated receiving terminal of the target user through the corresponding push interface. It supports simultaneous push through multiple channels such as SMS, email, and platform in-site message to ensure that users can receive warning information in a timely manner. Push status feedback: Real-time monitoring of the sending status of each push channel, recording push results, sending time, and receiving status, retrying push for a preset number of times for failed push information, and generating push logs for subsequent auditing and tracing; It should be noted that the delivery rate of early warning information is ensured through push status monitoring and retry mechanisms, avoiding the situation where users do not receive risk information in a timely manner due to push failures from a single channel; the retention of push logs enables the entire process of early warning behavior to be auditable and traceable, meeting the evidence retention requirements for cross-border e-commerce intellectual property rights protection.

[0033] This invention discloses an image recognition-based method for inspecting infringing goods in cross-border e-commerce. Applied to the aforementioned image recognition-based cross-border e-commerce infringing goods inspection system, this method achieves automated, cross-platform, high-precision inspection and risk warning of infringing goods on cross-border e-commerce platforms through the coordinated operation of various modules. It solves the technical problems of low efficiency, incomplete coverage, and high missed detection rate associated with traditional manual inspections. The specific implementation steps of this method are detailed below. Each step is executed sequentially to form a complete closed loop for infringing goods inspection. The specific implementation method is as follows: S1. Complete the system parameter configuration, which includes the OpenClaw intelligent agent's inspection platform, inspection cycle, interface parameters, and push parameters of the early warning unit. At the same time, construct and initialize the genuine product feature library and the infringing product feature library. The genuine product feature library stores the standardized visual feature vectors of genuine products and brand logos, while the infringing product feature library stores the standardized visual feature vectors of known infringing products. Feature library initialization includes feature data entry and feature vector index construction. S2. Establish a compliant data interaction channel with the OpenClaw intelligent agent and the open API interfaces of multiple cross-border e-commerce platforms. According to the preset inspection strategy, traverse each site of the target cross-border e-commerce platform, obtain product list data through API interface calls, and after deduplicating and filtering the product list data, obtain the image data of the target product and related product information. Establish a one-to-one correspondence between the related product information and the image data. S3. The AI ​​vision big model module performs preprocessing and feature extraction operations on the acquired product images in sequence. First, the product images are processed into standardized data that meet the model input requirements. Then, the convolutional neural network is used to extract multi-dimensional visual features from the standardized images to generate a fixed-dimensional standardized feature vector to be tested. S4. Input the standardized feature vector to be tested into the genuine product feature library and the infringing product feature library respectively. Calculate the similarity between the feature vector to be tested and the feature vector in the feature library through the similarity matching algorithm, obtain the highest similarity value, and compare the calculated similarity with the preset judgment threshold to determine whether the current product has the risk of infringement or being infringed. The preset judgment threshold includes the infringement judgment threshold and the genuine matching threshold. S5. When it is determined that there is a risk of infringement or being infringed, the AI ​​vision big model module sends a trigger signal to the early warning unit. After receiving the signal, the early warning unit triggers the early warning mechanism and pushes structured early warning information to the target user according to the preset push channel and push format. The early warning information includes the core identification information of the risky product and the judgment result information. By deeply integrating the method with system modules, a fully automated closed loop is formed, encompassing "parameter configuration, data collection, feature extraction, risk assessment, and early warning push." ​​This completely eliminates the reliance on manual inspections, enabling comprehensive inspections across platforms and multiple sites, significantly improving the efficiency and accuracy of infringement inspections, and reducing the missed detection rate and response lag.

[0034] Furthermore, the image recognition-based cross-border e-commerce infringing goods inspection method of this invention achieves full-process automation from system configuration, data collection, feature extraction, risk assessment to early warning push through a closed loop of steps S1 to S5. This effectively improves the efficiency and accuracy of cross-border e-commerce infringing goods inspection, expands the inspection coverage, and significantly reduces the missed detection rate and response lag, providing technical support for the intellectual property protection and compliant operation of cross-border e-commerce merchants. It should be noted that each step is executed sequentially with unidirectional data flow. Data validity checks are set between steps to ensure that only valid data output from the previous step enters the next step, thus guaranteeing the stability and reliability of the entire inspection method. In addition, the parameters of each step can be flexibly configured according to actual business needs to adapt to the inspection requirements of different brands and cross-border e-commerce platforms.

[0035] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

[0036] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. An image recognition-based cross-border e-commerce infringement product patrol system, comprising a data acquisition unit, an image recognition unit and a warning unit, characterized in that, It also includes the OpenClaw intelligent agent, a feature library storage module, and an AI vision large model module; The OpenClaw intelligent agent establishes a compliant data interaction channel with the open API interface of at least one cross-border e-commerce platform. The interface interaction follows the platform's data access protocol and calling specifications. It is used to traverse each site of the target cross-border e-commerce platform according to a preset inspection strategy to obtain image data of the target product and related product information. The related product information includes product ID, product title, product category, store information, and product link. The feature library storage module is used to pre-build and store a genuine product feature library and an infringing product feature library. The genuine product feature library stores standardized visual feature vectors of genuine products and brand logos, and the infringing product feature library stores standardized visual feature vectors of known infringing products. The feature vectors are all fixed-dimensional numerical vectors generated by feature extraction algorithms. The AI ​​vision big model module establishes a bidirectional communication connection with the OpenClaw agent and the feature library storage module, respectively, to receive product image data transmitted by the OpenClaw agent, extract multi-dimensional visual features from the product images to generate a standardized feature vector to be tested, compare the standardized feature vector to be tested with the feature vectors in the genuine product feature library and the infringing product feature library, and determine whether the current product has the risk of infringement or being infringed based on the comparison results and a preset judgment threshold. The warning unit is electrically connected to the AI ​​vision big model module and is used to trigger the warning mechanism when the AI ​​vision big model module determines that there is a risk of infringement or being infringed, and to push structured warning information to the target user. The structured warning information is a set of feature information in a standardized format. 2.The image recognition based cross-border e-commerce infringement product patrol system according to claim 1, characterized in that, The OpenClaw agent supports multi-node distributed concurrent access. Each node allocates inspection tasks according to a load balancing strategy. The agent executes timed inspection tasks according to a preset inspection cycle. The inspection cycle supports a combination configuration of global full-category periodic inspection and high-frequency inspection of key risk categories. Global full-category inspection and key risk category inspection are independent inspection task processes, and the inspection time and inspection frequency can be configured separately. 3.The image recognition based cross-border e-commerce infringement product patrol system according to claim 1, characterized in that, The OpenClaw intelligent agent has a built-in data filtering unit, which is a data cleaning and processing module. During the process of acquiring product information, it performs multi-level deduplication and redundancy filtering on duplicate or highly similar product data in the product list data. The deduplication filtering is based on the unique identifier of the product, image feature value, and text feature value. Only one image acquisition and feature comparison processing is performed for the same product. Filtered invalid product data does not enter the subsequent feature extraction process. 4.The image recognition based cross-border e-commerce infringement product patrol system according to claim 1, characterized in that, Before extracting features from product images, the AI ​​vision large model module first performs standardized preprocessing operations on the input product images. The preprocessing operations are based on image processing algorithms and include image size adjustment, pixel distribution normalization, and image noise filtering. The image size is adjusted to a fixed resolution tensor format, pixel distribution normalization maps pixel values ​​to a preset value range, and image noise filtering uses filtering algorithms to eliminate interference noise in the image. 5.The image recognition based cross-border e-commerce infringement product patrol system according to claim 1, characterized in that, The AI ​​vision large model module adopts a deep learning convolutional neural network architecture, which includes convolutional layers, pooling layers, fully connected layers, and residual connection structures. It is used to extract global visual features and local key features from product images in multiple dimensions. The features include the product's color features, texture features, shape and outline features, and brand logo features. The global visual features are the overall visual attribute features of the product, and the local key features are the iconic visual attribute features of the product and brand. 6.The image recognition based cross-border e-commerce infringement product patrol system according to claim 1, characterized in that, The AI ​​vision large model module adopts a high-dimensional feature vector similarity matching algorithm, combined with the fast retrieval capability of the vector database. The vector database is equipped with an approximate nearest neighbor retrieval index to realize full matching and comparison between the standardized feature vector to be tested and the feature library of genuine products and the feature library of infringing products. The matching and comparison process is to calculate the similarity between the feature vector to be tested and all feature vectors in the feature library one by one, and output the highest similarity value and the corresponding feature vector. 7.The image recognition based cross-border e-commerce infringement product patrol system according to claim 1, characterized in that, The AI ​​vision big model module has a built-in infringement judgment threshold, which is a preset similarity value. When the highest similarity between the standardized feature vector to be tested and the feature vector in the feature library of infringing products is higher than the preset infringement judgment threshold, the current product is judged to be a suspected infringing product. The judgment result is synchronously transmitted to the early warning unit and triggers the corresponding early warning process. 8.The image recognition based cross-border e-commerce infringement product patrol system according to claim 1, characterized in that, The AI ​​vision large model module has a built-in genuine matching threshold, which is a preset similarity value. When the similarity between the standardized feature vector to be tested and the corresponding feature vector in the genuine product feature library is lower than the preset genuine matching threshold, and the matching degree with the infringing product feature library meets the preset risk condition, it is determined that the corresponding genuine product is at risk of being infringed. The preset risk condition is that the similarity between the feature vector to be tested and the infringing product feature library is higher than the preset infringement judgment threshold.

9. The image recognition-based cross-border e-commerce infringing goods inspection system according to claim 1, characterized in that, The early warning unit has built-in multi-channel push interfaces, including SMS gateway interface, email sending interface, and platform in-site message interface. The structured early warning information pushed by the early warning unit includes the name of the risky product, the platform site to which it belongs, the product link address, the infringement type, and the inspection and identification time. The early warning unit supports pushing early warning information through at least one of SMS, email, and platform in-site message. The early warning information from different push channels has the same standardized format content.

10. A method for inspecting infringing goods in cross-border e-commerce based on image recognition, characterized in that: The method applied to the image recognition-based cross-border e-commerce infringing goods inspection system described in claims 1-9 includes the following steps: S1. Complete the system parameter configuration, which includes the OpenClaw agent's inspection platform, inspection cycle, interface parameters, and push parameters of the early warning unit. At the same time, construct and initialize the genuine product feature library and the infringing product feature library. The genuine product feature library stores standardized visual feature vectors of genuine products and brand logos, and the infringing product feature library stores standardized visual feature vectors of known infringing products. Feature library initialization includes feature data entry and feature vector index construction. S2. Establish a compliant data interaction channel with the OpenClaw intelligent agent and the open API interfaces of multiple cross-border e-commerce platforms. According to the preset inspection strategy, traverse each site of the target cross-border e-commerce platform, obtain product list data through API interface calls, and after deduplicating and filtering the product list data, obtain the image data of the target product and related product information. Establish a one-to-one correspondence between the related product information and the image data. S3. The AI ​​vision big model module performs preprocessing and feature extraction operations on the acquired product images in sequence. First, the product images are processed into standardized data that meet the model input requirements. Then, the convolutional neural network is used to extract multi-dimensional visual features from the standardized images to generate a fixed-dimensional standardized feature vector to be tested. S4. Input the standardized feature vector to be tested into the genuine product feature library and the infringing product feature library respectively. Calculate the similarity between the feature vector to be tested and the feature vector in the feature library through the similarity matching algorithm, obtain the highest similarity value, and compare the calculated similarity with the preset judgment threshold to determine whether the current product has the risk of infringement or being infringed. The preset judgment threshold includes the infringement judgment threshold and the genuine matching threshold. S5. When an infringement risk or the risk of being infringed is determined, the AI ​​vision big model module sends a trigger signal to the early warning unit. After receiving the signal, the early warning unit triggers the early warning mechanism and pushes structured early warning information to the target user according to the preset push channel and push format. The early warning information includes the core identification information of the risky product and the judgment result information.

Images