Reproduce “Single Image Reflection Removal Exploiting Misaligned Training Data and Network Enhancements” on OpenResearch

About this paper

Kaixuan Wei, Jiaolong Yang, Ying Fu, David Wipf, Hua Huang

Researchers from Beijing Institute of Technology and Microsoft Research developed the Enhanced Reflection Removal Network (ERRNet) to improve single image reflection removal by integrating context-encoding modules and introducing an alignment-invariant loss function for leveraging misaligned training data. ERRNet achieved state-of-the-art performance on several real-world datasets, including a PSNR of 22.89 and SSIM of 0.803 on the 'Real20' benchmark, while demonstrating effective use of unaligned data to enhance model robustness.

Key insights

• Embedding context-encoding modules (CWC and MSC) into the network architecture effectively leverages global and multi-scale spatial information, improving reflection separation.
• Deep features from pre-trained CNNs (specifically 'conv5_2' of VGG-19) exhibit robustness to spatial misalignments, allowing for the construction of an alignment-invariant loss function.
• Removing Batch Normalization layers in the generator and increasing network capacity can enhance performance in dense prediction tasks like reflection removal, which aligns with findings in other image-to-image translation problems.

Summary by alphaXiv.