Context-patch for difficult face recognition

Multiple research has shown the advantage of patch-based or local representation for face recognition. This paper builds on a novel way of putting the patches in context, using a foveated representation. While humans focus on local regions and move between them, they always see these regions in “context”. We hypothesize that using foveated context can improve performance of local region or patch based recognition techniques. The face images captured in uncontrolled environment suffer greatly due to blur, scale, resolution and illumination. In such situations, a facial patch by itself does not provide highly discriminative information. Correct patches may have higher intra-subject variation and incorrect patches may have lower inter-subject distance. To overcome this issue, we define a context-patch which is a face region that contains more information about a particular region and some contextual information about the rest of the face region. The low-resolution context is tolerant of intra-subject variations but still responds to many inter-subject differences. We build multi-class SVMs per context patch and fuse the normalized margins for classification. We show that by using the context-patch decision level fusion, the identification as well as verification performance of face recognition system can be greatly improved, especially in the case of highly degraded images. We conducted the experiments on the Remote Face Database and show the improvement over state of the art algorithms and the standalone patch fusion algorithm.