Two-stage single image reflection removal with reflection-aware guidance

Removing undesired reflection from an image captured through a glass surface is very challenging problem with many practical applications. For improving removal, cascaded deep models have been usually adopted to estimate the transmission in progressive manner. However, most existing methods are still limited exploiting result prior stage for guiding estimation. In this paper, we present novel two-stage network reflection-aware guidance (RAGNet) single removal (SIRR). To be specific, layer firstly estimated due that it generally much simpler and relatively easier estimate. Reflection-aware (RAG) module then elaborated better predicting layer. By incorporating feature maps observation, RAG can used (i) mitigate effect of (ii) generate mask soft partial convolution mitigating deviating linear combination hypothesis. A dedicated loss further presented reconciling contributions encoder decoder features. Experiments on five commonly datasets demonstrate quantitative qualitative superiority our RAGNet comparison state-of-the-art SIRR methods. The source code pre-trained model available at https://github.com/liyucs/RAGNet .