Off-Feature Information Incorporated Metric Learning for Face Recognition

摘要: Distance metric learning suppresses the intraclass variation while preserving inter-class between two feature vectors. However, these types of information are mixed in vectors that need to be separated based on from training data. The limited data may not able well separate and hence limits effectiveness learning. This letter proposes exploit off-feature help suppress For face recognition, some identity-independent such as pose, expression, occlusion is extracted source images utilized enhance performance distance In training, algorithm learns how incorporate features. similarity score a image pair determined by its space space. Extensive experiments demonstrate proposed incorporated helpful vectors, which visibly enhances existing algorithms.