Face recognition with occlusion using Dynamic Image-to-Class Warping (DICW)

A novel approach Dynamic Image-to-Class Warping (DICW) is proposed to deal with partially occluded face recognition in this work. An image is partitioned into sub-patches, which are then concatenated in the raster scan order to form a sequence. A face consists of forehead, eyes, nose, mouth and chin in a natural order and this order does not change despite occlusion or small rotation. Thus, in this work, a face is represented by the aforementioned sequence which contains the order of facial features. Taking the order information into account, DICW computes the distance between a query face and an enrolled person by finding the optimal alignment between the query sequence and all sequences of that person along both time dimension and within-class dimension. Extensive experiments on public face databases with various types of occlusion have verified the effectiveness of the proposed method. In addition, our method, which considers the inherent structure of the face, performs with greater consistency than current methods when the number of enrolled images per person is limited. Our method does not require any training process and has a low computational cost, which makes it applicable for real-world FR applications.