EStarGAN: Enhanced StarGAN for Multi-Style High-Definition Image Translation

Among the major remaining challenges in image-to-image translation is the capacity to generate multi-domain High-Definition images. Recently, the StarGAN is proposed to solve the multi-domain image translation problem. However, the StarGAN focuses on the low-resolution facial image synthesis, leaving some high-frequency information and unsuitable for HD image translation tasks. To address the issue, we propose the Enhanced StarGAN (called EStarGAN), aiming to generate multi-style High-Definition images with fine-grained details simultaneously. Specifically, we propose the Channel-and-Spatial Residual-in-Residual (CSRIR) module to learn high-frequency information in the deep network more effectively. Furthermore, we propose a new reconstruction loss, which consists of mean squared error and Structural Similarity loss to enhance the quality of the generated images. Extensive experiments on real-world datasets prove that our EStarGAN excels baselines with respect to both subjective and objective evaluations.