EIGAN: ENHANCED INPAINTING GENERATIVE ADVERSARIAL NETWORK

Generating coherent texture in a repaired region, especially at the boundary, is one of the challenges in image inpainting. To maintain the coherence in the edge transitional region, we propose an efficient framework, enhanced inpainting generative adversarial network (EIGAN). EIGAN is composed of a multi-resolution intersection encoder, a dual decoder, and an adversarial patch discriminator. The encoder and the decoder are used to process multi-resolution semantic information in a parallel and cross-connected fashion. Specially, we utilize a soft-contextual attention module embedded into the decoder to capture contextual information so that the attention module can help the decoder to generate coherent texture and content. Moreover, to improve the details, we introduce an advanced patch discriminator with relativistic adversarial loss named adversarial patch discriminator. Experimental results show that our improvements can enhance the coherence and quality of the completed image, and enable EIGAN to outperform the state-of-the-art image inpainting methods.