Uma melhoria do algoritmo K-SVD com aplicações em reconhecimento facial

Abstract

Image representation is an essential issue regarding the problems related to image processing and understanding. In the last years, the sparse representation modeling for signals has been receiving a lot of attention due to its state-of-the-art performance in different tasks such as image denoising, image inpainting and classification. One of the important factors to its success is the ability to promote representations well adapted to the data which rised with the dictionary learning algorithm. The most well known of theses algorithms is the K-SVD. In this work we proposed the αK-SVD algorithm, an algorithm which tries to explore the search space of possible dictionaries better than the K-SVD. Moreover, we studied different ways of exploring the search space of dictionaries in order to understand its impact on the algorithm performance. All theses methods are evaluated based two factores: the ability of sparse representing a set of given signals; and the recognition perfomance on two public face recognition databases. The results showed that our approaches achieved better results than the K-SVD and LC-KSVD when the sparsity level is low, i.e., when the number of non-zero elements on the representation is small.