Deep Collaborative Model for Mix-Domain Applied to Photo-Sketch Synthesis

Abstract

The process of sketch synthesis from real face photos is of great importance in the area of face recognition and remains a challenging issue for law enforcement agencies. Due to the different characteristics between sketch and photo and limited training data, photo/sketch synthesis has become the topic of great concern for the research community. In addition, recent synthesis models are unable to generate a high-resolution realistic photo/sketch. To determine these issues, we propose a novel synthesis framework by employing the contrastive lost and generative loss in the form of collaborative loss. This collaborative loss discovers the coherent features between the photo and sketches and recovers the underlying structure which can be helpful to generate high-resolution photo-sketch synthesis. We learn a multi-domain mapping relationship from the sketch-photo mix domain by transferring high-level quality from insufficient phot-sketch training data. The resultant identity-preserved face sketches can be treated as hidden data which can be combined with insufficient original data to recover the deficiency or underlying structure. We perform the qualitative and quantitative experiments on the challenging publically available photo-sketch datasets and yield better performance compared to the existing state-of-the-art framework

Keywords

• photo-sketch synthesis
• face recognition
• contrastive loss
• collaborative loss
• generative loss
• multi-domain