DSFT-GAN: Dilated Spatial Feature Transform Generative Adversarial Network on Image Inpainting

초록

Generating a realistic and complete image from a complex corrupted image remains as a big challenge and always lead to unreliable and ambiguous structural prediction. In order to address these problems, we propose a framework of coarse and refinement phase that consists of two generators and two discriminator networks. Our framework, named as DSFT-GAN focuses on studying the relationship of affine transformation within the ground truth and coarse image by applying spatial feature transform onto the refinement network. The proposed method was evaluated on CelebA dataset with over 200K images.