Expression neutralization based algorithm for expression-invariant face recognition for single training sample problem
G. Tharshini, V. Shanmugarajah, H. Sathananthan, Jeya Mhiruna Thiruchelvam, Roshan Indika Godaliyadda, M. P. B. Ekanayake, Janaka Wijayakulasooriya · 2017
The deformations incurred on the face during facial expressions significantly degrade the performance of the face recognition systems. It is a critical issue in many applications such as in forensic and surveillance as it is easy to spoof or fake the system using facial expression. This paper addresses this problem by proposing a method that estimates the neutral images of the subject and extracts the robust neutralized face segments for face recognition. Further, it provides a solution to the limited training sample problem encountered in many applications. A linear discriminant analysis (LDA) based approach is proposed to encode the common expression deformations that can be observed on the neutral faces across different individuals. This information is extracted in the form of vectors using a set of generic subjects that contain neutral and expressive images. Those are then used to remove the expression of the unknown test subject and estimate some neutral components of it. Further, the statistics of neutralization errors that arise due to applying general expression information for expression removal are learned from the generic subjects. It is applied to find the discriminating features of the neutralized faces. Experiments on the Extended Cohn-Kanade (CK+) database show the effectiveness of the proposed method.