Enhancing Image Representation in Conditional Image Synthesis

Even though deep neural network-based conditional image synthesis has shown impressive advances in terms of image quality, they still fall short of dealing with domain-dependent global and local styles and distinct shape representations of synthesized images. To address this issue, we propose a novel GAN-based conditional image synthesis model that incorporates a conditional normalization layer called IAN for style and edge-weighted shape enhancing loss for shape. Comparative experiments and ablation studies on popular and different domain datasets show that the proposed model outperformed other popular image-to-image translation model for diverse image domains.