Local Region Partitioning for Disguised Face Recognition Using Non-negative Sparse Coding

In this paper, three initializing methods for the Non-negative Sparse Coding are proposed for the disguised face recognition task in two scenarios: sunglasses or scarves. They aim to overcome previous sparse coding methods' difficulty, which is the requirement for a comprehensive training set. This means spending much more effort for collecting images and matching, which is not practical in many real world applications. To build a training set from a limited database containing one neutral facial images per person, a number of training images are derived from one image in the database using one of the three following partitioning methods: (1) grid-based partitioning, (2) horizontal partitioning and (3) geometric partitioning. Experiment results will show that these initialization methods facilitate Non-negative sparse coding algorithm to converge much faster compared to previous methods. Furthermore, trained features are more localized and more distinct. This leads to faster recognition time with comparable recognition results.