Face Photo Synthesis Via Intermediate Semantic Enhancement Generative Adversarial Network

Face sketch-photo synthesis is an important task in computer vision now. Recently, researchers have introduced face parsing to further improve the quality of synthesized face images. However, the semantic difference between face sketch parsing and photo parsing is usually ignored, leading to deformations and aliasing on synthesized face images. To solve these problems, we propose an intermediate face parsing to enhance the semantic information of the input face parsing. According to this intermediate face parsing, we propose an Intermediate Semantic Enhancement Generative Adversarial Network (ISEGAN) to generate high-quality realistic face photos. Furthermore, a Parsing Matching Loss (PM Loss) is proposed to encourage the intermediate face parsing to be more semantically accurate. Extensive comparison experiments demonstrate that our ISEGAN significantly out-performs the state-of-the-art methods.