Upscaling faces for recognition systems using trained filters

Face recognition systems which perform in unconstrained environment needs to cope with vary input resolutions. Low resolution input images are often quite problematic. This paper describes the improvement obtained on facial expression recognition and identity recognition when applying resolution upscaling methods based on trained filter techniques. In this paper, we explore the method proposed by Kondo et al. Kondo's methode has already been applied in display technology successfully, but hasn't been studied for recognition tasks before. In addition, we examine the influence on the recognition performance of the type of material exploited to train the upscaling filters, by selecting both generic videos and face images as samples. We then analyze how the size of the training set used in generating the upscaling filters influence the recognition rates. We show that adopting faces as training material produces upscaling filters capable of both higher recognition performance and much reduced requirements on the training size.