AI Unseen-Class Item Detection With Class-Specific Codebooks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence models struggle with accurately classifying items with high visual similarity between classes and are affected by data noise, leading to reduced detection performance and quality of service in home appliances.
Innovation Solution
An artificial intelligence apparatus using a codebook for each pre-learned item class to encode and decode item images, enabling accurate classification of items with high similarity and data noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If an existing artificial intelligence model is used to classify items, then the model can recognize pre-learned items of seen classes, but it makes errors when discriminating unseen class items that have high visual similarity to seen class items
Solution Approach 1:
The codebook is divided into multiple codebooks, each corresponding to a specific item class. This segmentation allows the system to separately learn and store encoding patterns for each class, improving the ability to distinguish between visually similar items of different classes while maintaining recognition accuracy for pre-learned items.
2Device complexity
If an existing artificial intelligence model is used for item classification, then the system can operate with a simple structure, but detection performance deteriorates when data noise is present
Solution Approach 1:
The codebook for each item class is pre-trained and stored beforehand. This preliminary action allows the system to have ready-made encoding patterns for various item classes, enabling it to quickly and accurately classify items even in the presence of noise, without requiring complex real-time processing structures.
3Device complexity
If a single codebook is used for all item classes, then the system structure remains simple, but the ability to distinguish between visually similar items of different classes is reduced
Solution Approach 1:
The codebook is divided into multiple codebooks, each corresponding to a specific item class. This segmentation allows the system to separately learn and store encoding patterns for each class, improving the ability to distinguish between visually similar items of different classes while maintaining recognition accuracy for pre-learned items.
Solution Approach 2:
Each codebook is specialized for a specific item class, giving local quality to each codebook. This specialization allows each codebook to optimize its encoding patterns for its specific class, improving classification accuracy for visually similar items while keeping the overall system structure manageable.
Data Source
AI summary
The present disclosure relates to an artificial intelligence apparatus capable of classifying and detecting an item of an unseen class having high visual similarity to an item of a seen class that has been learned and known in advance, and a method for detecting an unseen class items thereof, and when the item image is input, generate encoded data by encoding the item image, generate decoded data by decoding the encoded data using a codebook of a plurality of codebooks in which encoded data for each item class is stored, corresponding to a pre-learned item class, and detect the unknown item by classifying the class of the item image based on the generated decoded data.