Divergent Angular Representation for Open Set Image Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Open set recognition (OSR) models must identify known classes and detect unknown samples not seen during training.
• Learning discriminative representations with high intra-class similarity and inter-class discrepancy is crucial for OSR.
• Trivial representation learning can cause models to incorrectly classify unknown samples as known ones without proper regularization.

Purpose of the Study:

• To propose Divergent Angular Representation (DivAR), a novel approach to address the limitations of current OSR models.
• To enhance the model's ability to discriminate between known and unknown classes.
• To improve the robustness and applicability of OSR models, including extending to one-class classification.

Main Methods:

• DivAR maximizes class discrimination using a specialized loss function.
• It boosts directional variation in representations across all samples to ensure separation between known and unknown classes.
• Self-supervision is integrated to enhance representation robustness and enable one-class classification (OCC).

Main Results:

• DivAR effectively separates known and unknown classes in the representation space.
• The method demonstrates improved performance on generic image datasets for both OSR and OCC tasks.
• DivAR integrates learning and inference into a single module, eliminating the need for extra inference machinery.

Conclusions:

• DivAR presents a plausible and effective solution for open set recognition and one-class classification.
• The proposed method overcomes the issue of trivial representation learning in OSR.
• DivAR offers a unified and efficient approach for both learning and inference in OSR tasks.