Face Composite Sketch Recognition by BoVW-Based Discriminative Representations

Abstract

Face sketches are one of the main sources used for criminal investigation. In this paper, we propose a new approach for the recognition of facial composite sketches. We propose the use of discriminative representations as a way to bridge the modality gap between sketches and mug-shot photos. The intermediate representation is based on the bag-of-visual-words (BoVW) approach using dense SIFT features on multiple scales. Next, a discriminative representation is computed on top of the intermediate representation. Experimental results show that the discriminative representations outperform state-of-the-art approaches for this task in composite sketch datasets for both a close-set scenario as well as an open-set recognition scenario.