The Uncorrelated and Discriminant Colour Space for Facial Expression Recognition

Abstract

Recent research has shown improved performance by embedding the colour information in the process of facial expression recognition (FER). However, the RGB colour space may not always be the most desirable space for facial expression shown in face recognition. This paper addresses the problem of how to learn an optimum colour space for facial expression recognition based on the given training sample set. There are two typical learning colour spaces which have been used for face recognition. The uncorrelated colour space (UCS) decorrelates the three component images of RGB colour space using principal component analysis, and the discriminant colour space (DCS) creates three new component images by applying discriminant analysis. We will investigate these two colour spaces for facial expression recognition. First, colour face images are transformed into these colour spaces and represented by concatenating their component vectors. Secondly, facial expression recognition is achieved by utilizing Fisher Linear Discriminant (FLD). We test these colour spaces on Oulu-CASIA NIR&VIS facial expression database and CurtinFaces database in three ways: person-independent, person-dependent and crossing image sources. The results reveal that the uncorrelated colour space is more effective than RGB space in colour information representation for facial expression recognition, but the discriminant colour space fails to bear comparison with RGB space, which is significantly different from the case of face recognition.