Multi-Layer Subspace Knowledge Transfer Framework for Open-Set Recognition

In several real-world pattern analysis applications, the methodologies usually have issues regarding the categories (classes), where some are not in involved in the training stage. Towards this open-world recognition scenario, there is an urgent need to disambiguate the samples from both known and unknown categories using the extra knowledge, since more of this information leads to lower open risk. Nevertheless, there is no explicit technique to extract the knowledge from the subspace (which tells the class membership of the data) for the open-set recognition problem. In addition, little investigation has been done to formalize the statistical decision boundary in the open environment theoretically. In this paper, we argue that subspace knowledge plays a positive effect on open-set recognition. We further statistically formulate a subspace knowledge transfer framework and demonstrate that the knowledge transfer can well incorporate with the open-set recognition. Specifically speaking, we proposed a Multi-layer Subspace Knowledge Transform (MSKT) framework for open-set recognition. In MSKT, we make it as an extendable framework which provides a progressive solution to the open-set recognition problem. Extensive experiments, comparisons with several state-of-the-art methods, and qualitative analysis are performed to show the superiority of the proposed MSKT method.