RefLap：Efficient and Scale-Robust Reflection Removal via Dual-Domain Feature Fusion

Abstract

Single image reflection removal (SIRR) remains a challenging problem due to the intricacies involved in separating layers with varying textures and intensities. While many recent methods have focused on maximizing perceptual quality or pushing performance benchmarks, their complexity and computational cost often hinder practical deployment. In this work, we propose a dual-branch reflection removal network within a Deep Laplacian Pyramid Network framework, which balances performance and efficiency through a structurally meaningful design. The frequency-domain branch, DWT-FFC, exploits Discrete Wavelet Transform and Fast Fourier Convolution inside a U-Net architecture to capture multi-scale frequency cues and suppress reflection patterns. While the spatial-domain branch, UHDM, uses pixel unshuffling, Residual Dense Blocks (RDB), and Scale Attention Modules (SAM) to improve the structural consistency of image restoration and restore fine details. For cross-domain integration to be robust, a hierarchical fusion strategy is proposed that adaptively transfers multi-scale residuals from the Laplacian-based DWT-FFC branch to guide the UHDM decoder through cross-scale attention. Various experimental results show that our method can eliminate reflections efficiently while holding onto sharp textures. Although our method does not outperform the latest state-of-the-art solutions in terms of quantitative metrics, we demonstrate that its structural simplicity, favorable model size, fast inference speed, and lower FLOPs make it a practical and efficient choice for lightweight reflection removal in real-world applications.