RRFormer: A new transformer-based method for ultra high-definition reflection removal

Deep learning-based methods have achieved significant success in single-image reflection removal (SIRR). However, most of these methods focus on high/standard definition images while ignoring higher-resolution images such as Ultra-High-Definition (UHD) images. With the increasing prevalence of UHD images captured by modern devices, we aim to address the problem of UHD SIRR. Specifically, we first synthesize two large-scale UHD datasets at 4K and 8K resolutions. To our knowledge, these are the first large-scale UHD datasets for SIRR. We then conduct a comprehensive evaluation of twelve state-of-the-art SIRR methods on the proposed datasets. Moreover, we propose a transformer-based architecture named RRFormer for reflection removal, which comprises three modules: the Preprocessing Embedding Module, the Self-attention Feature Extraction Module, and the Multi-scale Spatial Feature Extraction Module. These modules extract hypercolumn features, global and partial attention features, and multi-scale spatial features, respectively. To ensure effective training, we utilize a combined loss function consisting of pixel loss, feature loss, and adversarial loss. Experimental results demonstrate that RRFormer achieves SOTA performance on both existing non-UHD benchmarks and our proposed UHD datasets.