Hierarchical Product Attribute Recognition for Self-Checkout Fraud Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition systems in self-checkout machines struggle with detecting fraudulent acts due to the need for extensive training data and difficulty in adapting to the frequent changes in product offerings, especially in stores like supermarkets and convenience stores where product lifecycles are short.
Innovation Solution
A specifying program and method using a machine learning model that employs a zero-shot image classifier, leveraging a hierarchical structure database and a contrastive language-image pre-training model to identify product attributes without requiring extensive training data updates, thereby reducing processing costs and adapting to product changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional image recognition systems are used in self-checkout machines, then they can detect objects, but they require extensive training data and cannot adapt to frequent product changes
Solution Approach 1:
The patent introduces an intermediary system consisting of a hierarchical attribute database and a contrastive language-image pre-training model. This intermediary layer translates product images into hierarchical attributes without requiring traditional extensive training, enabling the system to adapt to new products by simply updating the attribute database rather than retraining the entire recognition system
Solution Approach 2:
The system changes the parameters of the recognition approach by using a pre-trained contrastive language-image model that can be adapted to new products through parameter adjustments in the hierarchical attribute database, rather than requiring complete retraining with new training data for each product change
2Measurement precision
If traditional machine learning models are retrained for each product change, then recognition accuracy can be maintained, but processing costs and time increase
Solution Approach 1:
The system performs preliminary action by using a pre-trained contrastive language-image model that has already learned general object recognition capabilities. This pre-training eliminates the need for time-consuming retraining when products change, as the system can quickly adapt by updating the hierarchical attribute database with new product attributes
Solution Approach 2:
The contrastive language-image pre-training model serves multiple functions: it can recognize various types of objects across different product categories and adapt to new products through the hierarchical attribute system, making it a universal solution that maintains accuracy without requiring separate training for each product type
3Reliability
If comprehensive product attribute databases are maintained, then fraud detection accuracy improves, but system complexity increases
Solution Approach 1:
The system segments the product attribute database into a hierarchical structure with multiple levels (e.g., category, sub-category, specific attributes). This segmentation makes the database more manageable and reduces system complexity by organizing information in a structured way, while still maintaining comprehensive attributes for accurate fraud detection
Solution Approach 2:
The hierarchical attribute database acts as an intermediary between the image recognition model and the fraud detection logic. This intermediary structure simplifies the system architecture by providing a standardized interface for storing and retrieving product attributes, reducing the complexity of maintaining comprehensive product information
Data Source
AI summary
A non-transitory computer-readable storage medium storing a specifying program that causes at least one computer to execute a process, the process includes acquiring a video that includes an object; narrowing down, by inputting the acquired video to a machine learning model that refers to reference source data in which attributes of objects are associated with each of a plurality of hierarchies, attributes of the object included in the video among attributes of objects of a first hierarchy; identifying attributes of objects of a second hierarchy under the first hierarchy by using the attributes of the object obtained by the narrowing down; and specifying, by inputting the acquired video to the machine learning model, an attribute of the object included in the video among the attributes of the objects of the second hierarchy.