Two directional transform based sparse representation: A novel idea and method for sparse representation

Previous sparse representation (SR) methods are constructed on the assumption that the test sample can be approximately expressed by a linear combination of all original training samples. However, in most real-world applications samples are not subject to this assumption. Consequently, it is significant to explore a new way to improve SR. In this paper, we propose two directional transform based sparse representation (TDTBS) method. TDTBS can be viewed as a method that first maps the original training samples into a new dimension-invariant space and then generates sparse representation of the test sample in the new space. It seems that the devised two directional transforms enable the test sample to be better represented and classified.