Deep feature representation for face sketch recognition

Abstract

Face sketch recognition aims at matching face sketch images to face photo images. The main challenge lies in modality discrepancy between face photo and sketch images. In this work, we propose a new facial sketch-to-photo recognition approach by adopting VGG-Face deep learning network, with which face images can be represented by compact and highly discriminative feature vectors. Different from existing VGG-Face based methods which directly match face sketches to photos, we firstly transform the gallery photos to sketches for decreasing the modality difference. Experimental results on multiple face photo-sketch datasets indicate the superiority of our method.