PERSON RECOGNITION INSPIRED BY HUMAN VISUAL SYSTEM

Abstract

Our objective in this thesis is to inspire from humans’ visual system skills and mechanisms of person recognition, develop computational methods which improve the person recognition performance in machines. We achieve this objective by addressing three problems of person recognition. First, the new head-body matching problem, defined as follows: given an image of a person’s head, can we match his body image? We tackle this problem by utilizing the head-body correlations. Second, unconstrained face recognition, in which faces may have single/multiple variations such as pose, aging, illumination, expressions, etc. We tackle this problem by utilizing the human attention to faces to direct machines’ attention during face recognition. Finally, face recognition with limited training data. We tackle this problem by utilizing caricature faces and anti-faces in training. Overall, our results reveal that the person recognition by machines can be improved by learning from the human visual system's perceptual skills and mechanisms.