Impression transformation of a face based on discriminant analysis on separately coded representations of facial shape and texture

Abstract

This work describes an attempt to develop a real-life application for transforming the appearance of a person's face to give a more favorable impression. Face images are first represented as high-dimensional vectors by separating the shape and texture information, and variations in the appearance of face images are coded by applying principal component analysis to a set of various face images. The relationship between the coded representation and the impression of corresponding images is analyzed, and an impression transfer vector is defined by the Fisher's linear discriminant obtained from a number of training images previously categorized into two opposing classes for a given impression dimension. An image manipulation method for transforming impressions is proposed which shifts the projection of the input image in the parameter space by an arbitrary magnitude of the impression transfer vector. Experimental results for impression transformations of both age and gender dimensions showed that the method proposed for face image manipulation is an effective one.