Multiscale Gabor Ternary Code for Face Recognition with Single Sample Per Class

In this paper, we propose an approach for handling expression and pose variations in face recognition with single sample per class. The local appearance based methods have been successfully applied to face recognition and achieved state-of-the-art performance. The Gabor Binary Code approach has the robust properties against facial expression, illumination, accessories and etc. In this paper we enhance the robustness of GBC in the form of Multi Scale Gabor Ternary Code (GTC) for pose variation with large rotation angle. Normally most of the local appearance based methods the facial features are extracted from several local regions and concatenated into an enhanced feature vector as a face descriptor. In this approach we divide the face into several (m×m) non overlapped parallelogram blocks instead of square or rectangle blocks as well as the wavelet transformed low frequency band of the face image is used to generate Gabor ternary code. The parallelogram blocks based comparison improves the performance of face recognition under perspective and expression variation. Experiments on Indian face dataset faces with large rotation angle up to 180θ and ORL datasets shows that the proposed approach outperforms GBC in the scenario of one training sample per person.