Evaluation on synthesized face sketches

Abstract

Face sketch synthesis takes an important role in law enforcement and digital entertainment. For law enforcement, it is usually considered as an intermediate procedure for sketch based face recognition. In this scenario, face recognition accuracy can be used to evaluate the performance of face sketch synthesis methods to some extent, but not directly. However, for other applications such as digital entertainment, visual quality is much more important to users. Then face recognition accuracy is inappropriate to evaluate the visual quality of synthesized sketches. This is because face recognition methods mainly depend on some discriminative information while the visual quality of synthesized sketches heavily relies on texture information. To this end, we construct a synthesized face sketch database for the purpose of image quality assessment (IQA). These synthesized sketches are generated by five face sketch synthesis approaches. In addition, eight representative full-reference IQA methods are evaluated on this database. Some insights can be found from our experiments, which have great significance to image transformation problems such as face synthesis and face hallucination.