Error Encoded PDE-Textons for Face Recognition

The success of texture based face recognition systems confirms the important role of textural information in automated face recognition and retrieval. While discriminative modelling is considered as a powerful tool for modelling natural textural surfaces, it is not yet applied for face recognition and retrieval purposes. In addition, current approaches in discriminative modelling have so far ignored for the disparity between the representation and the actual images, and hence do not use this information for recognition purposes. In this work, we introduce the Error Encoded PDE-textons, constructed using both micro-and macro-patterns extracted using a set of linear and PDE-based filters. We use PDE-texton maps as a representation of the textural information in a given facial image. The residual image, the difference between the knowledge base and the given input image, is also used as this could provide information on possible errors in the discriminative representation. Face recognition is then performed by using features extracted from both the PDE-texton maps and the residual images. We compare the results of the proposed approach with those from four other different techniques. The results using only a single training image show that the proposed scheme has at least 9% reduced equal error rate compared to popular LBP-operators on the difficult AR dataset.