Local Geometry Feature Extraction for Face Sketch Recognition

In this paper, we present a new feature extraction approach for face sketch recognition. Based on the assumption that small image patches in the photo and sketch images form manifold with similar local geometry in two different image spaces, we propose local geometry feature extraction approach to reduce the modality gap between photos and sketches. Considering the impact of different scale to performance, we further propose multi-scale ensemble to improve the performance and avoid scale selection problem. Experimental results on CUFS database show that our method outperforms those state-of-the-art methods. Moreover, it also has a big advantage that it does not need any training procedure.