Reversible Decoupling Network for Single Image Reflection Removal

TL;DR

RDNet addresses the ill-posed problem of single-image reflection removal by preserving information with a reversible encoder and decoupling transmission/reflection through a multi-column architecture. It introduces a transmission-rate-aware prompt generator to adapt feature processing via channel-wise priors $\alpha_R, \alpha_G, \alpha_B, \beta_R, \beta_G, \beta_B$, guided by a two-stage training regime with content and perceptual losses. The method achieves state-of-the-art PSNR/SSIM on multiple real-world benchmarks and wins the NTIRE 2025 in-the-wild challenge, demonstrating strong generalization to varied reflection patterns. Overall, RDNet advances practical reflection removal by combining invertible architectures with adaptive feature modulation for robust, cross-scale decoupling.

Abstract

Recent deep-learning-based approaches to single-image reflection removal have shown promising advances, primarily for two reasons: 1) the utilization of recognition-pretrained features as inputs, and 2) the design of dual-stream interaction networks. However, according to the Information Bottleneck principle, high-level semantic clues tend to be compressed or discarded during layer-by-layer propagation. Additionally, interactions in dual-stream networks follow a fixed pattern across different layers, limiting overall performance. To address these limitations, we propose a novel architecture called Reversible Decoupling Network (RDNet), which employs a reversible encoder to secure valuable information while flexibly decoupling transmission- and reflection-relevant features during the forward pass. Furthermore, we customize a transmission-rate-aware prompt generator to dynamically calibrate features, further boosting performance. Extensive experiments demonstrate the superiority of RDNet over existing SOTA methods on five widely-adopted benchmark datasets. RDNet achieves the best performance in the NTIRE 2025 Single Image Reflection Removal in the Wild Challenge in both fidelity and perceptual comparison. Our code is available at https://github.com/lime-j/RDNet

Figures (4)

• Figure 1: Quantitative comparison in PSNR between ours and previous state-of-the-art methods, where we achieve new records on all 5 datasets. Note that the scale of each axis is normalized by its second-best value. The best and second-best PSNR values are displayed for reference.
• Figure 2: Overall structure of our RDNet, the input is fed in the transmission-rate-aware prompt generator, pretrained hierarchy extractor, and the column embedding. The output of the prompt generator will be transferred into the column network. After interactions between the columns, each column uses a separate decoder to obtain a pair of image layers.
• Figure 3: Qualitative comparisons on real-world cases. Please zoom in for more details.
• Figure 4: Qualitative comparisons on samples from the Wild dataset. Please zoom in for more details.