Locality-constrained feature space learning for cross-resolution sketch-photo face recognition

Matching sketch facial images to mug-shot images have crucial significance in law enforcement and digital entertainment. Conventional methods always assume that both the sketch and photo face images have the same resolutions. However, in real criminal detection, the target facial sketches obtained by the artist usually have different resolutions against the source photos in the mug-shot database. In this paper, we propose a locality-constrained feature space learning (LCFSL) method to address the above cross-resolution sketch-photo facial images matching problem. The proposed LCFSL approach not only build bridge to associate cross-domain face images, but also can learn resolution robust representation features for cross-resolution sketch-photo face recognition purpose. After common feature space learning, we simply use nearest neighbor classifier to perform recognition based on the projected features obtained from sketch-photo faces with different resolutions. Experiments conducted on CUHK student database and AR database have shown the effectiveness and superiority of our method to some state-of-the-art face recognition approaches.