Cascading Blend Network for Image Inpainting

Image inpainting refers to filling in unknown regions with known knowledge, which is full flourish accompanied by the popularity and prosperity of deep convolutional networks. Current methods have excelled completing small-sized corruption or specifically masked images. However, for large-proportion corrupted images, most attention-based structure-based approaches, though reported state-of-the-art performance, fail reconstruct high-quality results due short consideration semantic relevance. To relieve above problem, this paper, we propose a novel image approach, namely cascading blend network (CBNet), strengthen capacity feature representation. As whole, introduce an adjacent transfer attention (ATA) module decoder, preserves contour structure reasonably from layer blends structure-texture information shadow layer. In coarse delicate manner, multi-scale contextual (MCB) block further designed felicitously assemble multi-stage information. addition, ensure high qualified hybrid information, extra supervision applied intermediate features through cascaded loss. Qualitative quantitative experiments on Paris StreetView, CelebA, Places2 datasets demonstrate superior performance our approach compared algorithms.