Image Reflection Separation via Adaptive Residual Correction and Feature Interaction Enhancement
Loading...
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
The Eurographics Association
Abstract
Glass reflection superimposes images from both sides of the glass, resulting in severe image quality degradation that significantly impairs the performance of downstream tasks, such as object detection and image understanding. Therefore, it is essential to separate the transmission and reflection layers. However, due to lighting conditions and the material properties of glass, the relationship between the reflected and transmitted components often involves complex linear interactions, which limit the effectiveness of existing methods. Inspired by the observation that transmission components often dominate images with reflection in real-world scenes, we propose an image reflection separation method that integrates adaptive residual correction with feature interaction enhancement. Building upon a linear combination model enhanced with residual correction, we generalize the residual term based on the physical principles of light reflection and transmission. In order to ensure precise spatial alignment between the transparent and real images, We design an image registration mechanism and propose an Adaptive Hybrid Residual Loss, which significantly enhances the model's ability to perceive differences between the transmission and reflection layers, effectively balancing the complexity of linear mixture modeling with the diversity of real-world scenarios. To further highlight the interactive features between reflection and transmission, we incorporate a cross-dimensional attention mechanism into the dual-stream architecture designed for transmission-reflection processing. Extensive experiments and ablation studies show that our method achieves state-of-the-art performance on multiple real-world benchmark datasets, with an average PSNR improvement of 0.66 dB over the current best-performing model.
Description
CCS Concepts: Computing methodologies → Computational photography; Computer vision; Image manipulation
@inproceedings{10.2312:pg.20251275,
booktitle = {Pacific Graphics Conference Papers, Posters, and Demos},
editor = {Christie, Marc and Han, Ping-Hsuan and Lin, Shih-Syun and Pietroni, Nico and Schneider, Teseo and Tsai, Hsin-Ruey and Wang, Yu-Shuen and Zhang, Eugene},
title = {{Image Reflection Separation via Adaptive Residual Correction and Feature Interaction Enhancement}},
author = {Ke, Weijian and Mo, Yijun},
year = {2025},
publisher = {The Eurographics Association},
ISBN = {978-3-03868-295-0},
DOI = {10.2312/pg.20251275}
}