An Improved Face Synthesis Model for Two-Pathway Generative Adversarial Network

Abstract

Synthesizing photorealistic frontal face images from multiple-view profile face images has a wide range of applications in the field of face recognition. However, existing models still have some disadvantages such as high cost and high computational complexity. At present, the Two-Pathway Generative Adversarial Network (TP-GAN) is the state-of-the-art face synthesis model, which can perceive the global structure and local details at the same time. It solves the prier problems but has disadvantages such as training difficulty and lack of diversity of generated samples. Based on Wasserstein GAN with Gradient Penalty (WGAN-GP), this paper proposes a novel Two-Pathway Wasserstein GAN with Gradient Penalty (TPWGAN-GP) model to tackle these defects. TPWGAN-GP uses a gradient penalty method to satisfy the Lipschitz continuity condition, which solves the problems of difficulty in hyper-parameter adjustment and gradient explosion in the TP-GAN, making the convergence speed faster and the model more stable in training process. The generated samples are of higher quality, resulting in more photorealistic faces for recognition tasks.