Learning Device for Abnormal Class Discrimination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition techniques using domain adaptation struggle when training data for certain classes, especially abnormal classes, are insufficient, leading to biased discriminator training and poor classification performance.
Innovation Solution
A learning device and method that includes feature extraction, class discrimination, normal/abnormal discrimination, AUC loss calculation, and domain adaptation mechanisms to update parameters, ensuring accurate classification even with limited data for abnormal classes by using a combination of class discriminative loss, AUC loss, and domain discriminative loss.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If domain adaptation is used to train the discriminator with limited abnormal class data, then the training can proceed with available data, but the discriminator training becomes concentrated in normal classes and fails to correctly discriminate abnormal classes
Solution Approach 1:
The patent segments the training process into two distinct phases: (1) pre-training the discriminator on source domain data with abundant normal class samples, and (2) fine-tuning with domain adaptation using limited target domain data. This segmentation allows the discriminator to first learn general features from sufficient data, then adapt to the target domain with limited abnormal class samples, preventing training concentration on normal classes alone.
Solution Approach 2:
The patent performs preliminary action by pre-training the discriminator on the source domain before applying domain adaptation. This preliminary training establishes a baseline discriminative capability using abundant normal class data, which then serves as a foundation for the subsequent domain adaptation phase where limited abnormal class data is effectively utilized.
2Ease of manufacture
If training data is collected from public sources for domain adaptation, then training can be performed, but the data may not be prepared satisfactorily and evenly for all classes especially abnormal classes
Solution Approach 1:
The patent introduces domain adaptation as an intermediary mechanism that bridges the gap between source domain data (easily obtained from public sources) and target domain requirements (balanced class distribution). The domain adaptation technique acts as a mediator that transforms the easily obtainable but imbalanced source data into effective training data for the target domain, maintaining ease of data preparation while improving class distribution balance through the adaptation process.
3Productivity
If the discriminator is trained using conventional domain adaptation, then training can proceed with available source and target domain data, but the discriminator cannot correctly discriminate abnormal classes when abnormal class samples are scarce
Solution Approach 1:
The patent performs preliminary training on the source domain discriminator before applying domain adaptation. This preliminary action ensures the discriminator develops basic discrimination capabilities using abundant source domain data, which then enables effective fine-tuning with limited target domain abnormal class samples, maintaining training efficiency while improving abnormal class discrimination reliability.
Solution Approach 2:
The patent segments the training process into source domain pre-training and target domain fine-tuning with domain adaptation. This segmentation allows the system to efficiently utilize abundant source domain data for initial learning while subsequently focusing on improving abnormal class discrimination in the target domain with limited samples, thereby maintaining productivity while enhancing reliability for abnormal class detection.
Data Source
AI summary
In a learning device, a feature extraction means extracts image features from an input image. A class discrimination means discriminate a class of the input image based on the image features, and generates a class discriminative result. A class discriminative loss calculation means calculates a class discriminative loss based on the class discriminative result. A normal/abnormal discrimination means discriminates whether the class is a normal class or an abnormal class, based on the image features, and generates a normal/abnormal discriminative result. The AUC loss calculation means calculates an AUC loss based on the normal/abnormal result. A first learning means updates parameters of the feature extraction means, a class discrimination means, and the normal/abnormal discrimination means, based on the class discriminative loss and the AUC loss.