Abstract
As one of model-based methods for face recognition, a technique called Embedded HMM (Hidden Markov Model) was proposed, in which two or more HMMs are arranged spatially and connected in the fixed direction. This Embedded HMM can provide robust learning and recognition against changes of direction and positions of a face compared with the conventional methods based on PCA (Principle Component Analysis). However, when a face is rotated within an image plane, there is a possibility that face structure could not be expressed well by the Embedded HMM. Therefore, in this paper, we propose a novel Embedded HMM that has two kinds of structures with different direction. Furthermore, in order to improve its performance against partial hiding of a face, such as, existence of glasses, we propose a new recognition algorithm using plausible parts of face image.
