Abstract:In this paper, we propose a face perception model to mimic the biological mechanism of face perception and memory in human brain. We are mainly inspired by the fact that there are two functionally and neurologically distinct pathways after the early face perception and they both interact to process the changeable features of faces. Accordingly, our model consists of three perception parts, facial structure perception, facial expression perception and facial identity perception, which are all component-based. The structure perception has a feed-forward projection to the expression and identity perception, while the expression affects the identity perception with a modulation process. We embody the three parts referring to three bio-inspired computational models. For the facial structure perception, we utilize a cascaded-CNN (convolutional neural networks) approach to estimate the center locations of key facial components. For the facial expression perception, we propose a novel approach which exploits convolutional deep belief networks (CDBN) to spontaneously locate the places containing the most discriminative information and synchronously complete the feature learning and feature selection. For the facial identity perception, we propose an approach which adopts the hierarchical max-pooling (HMAX) model to encode notable characteristics of facial components and utilizes a new memory formation integrating the preliminary decision, expression modulation and final decision process. We evaluate our model through a series of experiments and the experimental results demonstrate its rationality and effectiveness.

A Biologically Inspired Model Mimicking the Memory and Two Distinct Pathways of Face Perception