DS3former: Dual-Stream Semantic Separation Transformer for Single Image Reflection Separation

Single image reflection separation is a challenging and ill-posed problem owing to diverse reflective surfaces and lighting conditions. This paper introduces DS3former, a dual-stream transformer network employing a semantic separation strategy to effectively distinguish between the transmission (T) and reflection (R) layers. We observe that within pre-trained deep semantic features of mixed images, individual channels exhibit varying affinities towards either the T or R layer, facilitating their differentiation. Based on this observation, we propose a novel semantic separation attention mechanism that adaptively extracts layer-specific features from different channels and performs inter-stream feature transfer and aggregation to enhance separation. To further improve performance at the semantic level, features from deeper decoder stages and external pre-trained models are integrated to guide the separation process in shallower encoder layers. Experimental results show that the proposed method outperforms state-of-the-art reflection separation methods in terms of quantitative metrics and visual quality.