RBSFormer: Enhanced Transformer Network for Raw Image Super-Resolution

Siyuan Jiang, Senyan Xu, Xingfu Wang

Abstract: In smartphones and mobile camera devices, the Image Signal Processor(ISP) is applied to reconstruct the RAW image into a sRGB image for human reading by a series of signal modules. Due to the non-linear ISP transformation, it is complicated to model the degradation in the sRGB domain. Most existing super-resolution methods directly handle the sRGB image processed by the ISP, introducing more difficult degradation patterns. To address this challenge, we propose an enhanced transformer network named RBSFormer. Unlike other methods that operate on sRGB images, RBSFormer takes RAW images as input, thus avoiding the complex degradation introduced by ISP processing. We design two enhanced core components, i.e., Enhanced CrossCovairance Attention(EXCA) and Enhanced Gated Feedforward Network(EGFN), in the RBSFormer, and we further introduce data augmentation in the RAW domain and hybrid ensemble strategies to enhance our results. Experimental results demonstrate superior performance against the majority of methods both qualitatively and quantitatively. Our RBSFormer achieves 3rd place in terms of all the evaluation metrics both on the official validation and testing set with fewer parameters in the NTIRE 2024 challenge on Raw Image Super Resolution.