A Stepwise Frontal Face Synthesis Approach for Large Pose Non-frontal Facial Image

Frontal face synthesis plays an important role in many fields. The existing methods mainly synthesize frontal face based on the consistency assumption that non-frontal and frontal face manifolds are locally isometric. But the assumption couldn’t be held well when non-frontal faces have large variations. To solve this problem, we propose a stepwise frontal face synthesis approach for large pose non-frontal facial image. Considering that the consistency is desirable when the angle variations of different poses are small, we divide frontal face synthesis into multiple stepwise synthesis steps. In each step, the intermediate pose training sets between non-frontal and frontal training sets are used to synthesize intermediate pose faces. Furthermore, in each step, we utilize the geometric structure of target face space with small pose as constraint to represent the input face with larger pose. Experimental results demonstrate that the proposed method outperforms other state-of-the-art methods quantitatively and qualitatively.