Single Image Reflection Separation via Collaborative Reverse Attention and Geometric Guidance

In Single Image Reflection Separation (SIRS), the objective is to recover the transmission and reflection layers from a single input image. Due to the highly ill-posed nature of the problem, achieving accurate layer separation remains challenging. In this work, we observe that the suppressed features in one layer' restoration process often contain useful information for restoring the other layer. Motivated by this, we propose a collaborative reverse attention (CoRA) mechanism that integrates the forward features of one stream with the reversed features of the other, enabling more effective cross-layer interaction. To enhance the utility of interactive features, we design three reversal functions and identify the optimal one for CoRA. Unlike prior methods, our approach reverses attention maps, facilitating more reliable feature exchange during interaction. Moreover, we find that depth estimation from the blended image predominantly captures transmission-layer geometry. Leveraging this observation, we incorporate geometric guidance from the estimated depth to assist reflection separation. Finally, CoRA and geometric guidance are embedded into a dual-stream Transformer architecture tailored for SIRS. Extensive experiments and ablation studies demonstrate the effectiveness of our approach, which achieves state-of-the-art performance across multiple real-world benchmark datasets.