Person Recognition Inspired By Human Visual System

Ms Mona Ragab Sayed Abdelgayed
Dr Sim Mong Cheng, Terence, Associate Professor, School of Computing
Dr Lim Joo Hwee, I2R

  03 Dec 2019 Tuesday, 02:30 PM to 04:00 PM

 Executive Classroom, COM2-04-02


Person recognition (PR) has been an attractive field to the computer vision community due to its various and important applications such as security, surveillance systems, etc. Due to the expertise of humans in recognizing persons in their daily life, their performance has been used as a benchmark for computational methods of PR. Our objective in this thesis is to inspire from the human visual system's (HVS) perceptual skills and mechanisms of PR, develop computational methods which improve the PR performance in machines.

We achieve our objective by addressing the following PR problems: First, the new head-body matching problem, which is defined as "given an image of a person's head, can we match his body image?". Inspired by the information processing pathway within humans' association cortex, as well as the human studies on head-body correlations, we establish the head-body correlation using features which can be extracted automatically from images and can be used for classification. We propose a dualpathway framework which computes head and body features independently, and utilize their correlation for head/body recognition. Second, unconstrained face recognition, in which a face may have different variations such as poses, illuminations, ages, expressions, etc. We conduct psycho-physics experiment to study humans' attention to unconstrained faces for face recognition. This attention is then utilized in a computational model to direct deep-CNNs during face recognition. Finally, face recognition with limited training data. We tackle this problem by inspiration from face caricaturing and anti-face adaptation effects in the HVS. By utilizing these effects, we introduce a new method for data augmentation which improves the face recognition performance in machines even if we use limited training data and primitive features such as PCA features.

Overall, our evaluation results show that leveraging the HVS perceptual skills and mechanisms of PR helps to improve the PR performance in machines.