PA-NAFNet: An improved nonlinear activation free network with pyramid attention for single image reflection removal

Abstract

Single Image Reflection Removal (SIRR) is an active topic in low-level vision, aiming to eliminate the influence of reflected objects or light sources on image quality. However, due to the ill-posed property of SIRR and the lack of large-scale real world reflection image datasets, existing methods degrade on real datasets and suffer from the problem of reflection residue. To address these issues, we propose an effective SIRR network called PA-NAFNet. It utilizes a non-linear activation-free network (NAFNet) as the baseline and incorporates a pyramid attention module to capture long-range pixel interactions. Additionally, during the training phase, color jittering technique is introduced to increase the diversity of the training dataset, thereby alleviating potential color distortion issues after reflection removal. Experimental results on multiple reflection removal benchmark tests demonstrate the effectiveness of PA-NAFNet. The relevant code is available on this link.