Adversarial compact wrapping classifier learning for open set recognition

Abstract

Determining the limited and compact class acceptance region has become a key bottleneck in classifier design since Open Set Recognition (OSR) requires the classifier to successfully identify those of interest and reject other exceptions. The imperfect match between actual feature distribution of a class and its classification acceptance area is a critical barrier to performance enhancement in recent existing OSR algorithms. Therefore, we propose adversarial compact wrapping classifier learning for OSR, which achieves better feature representation and classification surface division through alternating learning of classifiers and compact wrapping points (CWPs). The core of the algorithm model lies in the confrontational game between CWP and the classifier. First, a new deep and compact hyperspherical crown classifier is designed, which expands neurons to achieve limited and compact classification surface division, improves the classification performance of known classes, and reduces the risk of open space. Then, the CWP algorithm is used to construct negative class sample points that tightly wrap the class feature distribution, and a joint optimization framework of CWP and classifier is constructed. The classifier improves the discrepancy between the class acceptance area of the classification and the actual class distribution area through adversarial learning with CWP, further controls the risk of unknown open spaces, and optimizes the learned feature distribution. It is worth noting that we have developed the CWP to allow concave feature distributions. Experimental results show that the proposed algorithm is significantly superior to recent comparison algorithms on various benchmark datasets and significantly improves OSR performance on the TinyImageNet dataset.