Class-Specific Semantic Reconstruction for Open Set Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Open set recognition is crucial for deep neural networks (DNNs) to identify unknown classes while maintaining accuracy on known classes.
• Auto-encoders (AE) and prototype learning are promising existing methods for this task.
• Current approaches face challenges in effectively distinguishing between known and unknown data distributions.

Purpose of the Study:

• To propose a novel method, Class-Specific Semantic Reconstruction (CSSR), integrating auto-encoder and prototype learning for enhanced open set recognition.
• To improve the ability of DNNs to accurately classify both known and unknown samples.
• To develop a flexible and robust framework for open set recognition tasks.

Main Methods:

• CSSR replaces traditional prototype points with class-specific auto-encoder (AE) manifolds.
• Each known class is modeled on an individual AE manifold, with class belongingness determined by reconstruction error.
• Class-specific AEs are integrated with the DNN backbone to reconstruct semantic representations, not raw images.

Main Results:

• The proposed CSSR method demonstrates outstanding performance in both closed and open set recognition tasks.
• End-to-end learning allows the DNN and AEs to mutually enhance the learning of discriminative and representative information.
• Experimental results across multiple datasets validate the effectiveness of the CSSR approach.

Conclusions:

• CSSR offers a powerful and effective solution for open set recognition by leveraging class-specific semantic reconstruction.
• The method is simple, flexible, and can be readily incorporated into existing deep learning frameworks.
• CSSR significantly advances the capabilities of DNNs in handling data with unknown classes.