Sparse discriminative region selection algorithm for face recognition

摘要

We introduce a sparse region selection problem and an effective solution algorithm where one seeks for a small subset of points in a two-dimensional plane (e.g., image) that are considered to be the most important. Its direct application is the face recognition in machine vision in which we aim to find a group of key pixels in a facial image that are the most salient in discriminating subjects from others. Sparseness plays a key role in enhancing the prediction performance since observed data often contain considerable amount of noise potentially. In addition to the sparseness constraint, the active features need to be spatially coherent so as to form meaningful contiguous areas, not just random scatters. We formulate the problem and approximate it as convex optimization with nonnegative L1 constraints, where we introduce an efficient solution method that modifies the gradient Lasso algorithm that was previously used for solving convex problems with L1 constraints. We demonstrate that the proposed approach not only yields superior prediction performance to the existing methods on several real-world benchmark face datasets, but also discovers regions around the key facial features such as eyes/eyebrows and nose/mouth that are widely believed to be important for face recognition.