Face sketch recognition is an important task for forensic investigations. In the last decades most of the sketches used by law enforcement agencies are composites made from real photos, using some specialized software. In this work we propose a new method for the automatic recognition of these composite sketches. We propose to use deep learning features extracted from facial components to represent the sketches. We decide to use intermediate layers from already trained deep models and we employ a metric learning approach, as an easier way to learn the differences between sketch and photo domains. The experimental evaluation conducted on two available databases, shows the superiority of the proposal with respect to the use of the original deep models, as well as to other state-of-the-art methods.
Local deep features for composite face sketch recognition / Mendez-Vazquez, H.; Becerra-Riera, F.; Morales-Gonzalez, A.; Lopez-Avila, L.; Tistarelli, M.. - (2019), pp. 1-6. (Intervento presentato al convegno 7th International Workshop on Biometrics and Forensics, IWBF 2019 tenutosi a mex nel 2019) [10.1109/IWBF.2019.8739212].
Local deep features for composite face sketch recognition
Mendez-Vazquez H.
;Tistarelli M.
2019-01-01
Abstract
Face sketch recognition is an important task for forensic investigations. In the last decades most of the sketches used by law enforcement agencies are composites made from real photos, using some specialized software. In this work we propose a new method for the automatic recognition of these composite sketches. We propose to use deep learning features extracted from facial components to represent the sketches. We decide to use intermediate layers from already trained deep models and we employ a metric learning approach, as an easier way to learn the differences between sketch and photo domains. The experimental evaluation conducted on two available databases, shows the superiority of the proposal with respect to the use of the original deep models, as well as to other state-of-the-art methods.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.