The NTIRE 2024 RAW Image Super-Resolution Challenge aimed to upscale RAW Bayer images by 2x, considering unknown degradations such as noise and blur. A total of 230 participants registered, and 45 submitted results during the challenge period. The performance of the top-5 submissions is reviewed:
Team Samsung MX,SRC-B proposed a two-stage network with a Focal Pixel Loss, achieving the best performance. Their solution focuses on recovering the image structure in the first stage and refining the details in the second stage.
Team XiaomiMMAI introduced a dual branch network based on HAT, adopting re-parameterization during training to fully exploit the potential of the method. They also proposed a step-by-step and task-by-task training approach for RAW Image Super-Resolution.
Team USTC604 proposed a transformer-based framework, RBSFormer, which excels in capturing long-range pixel interactions by applying self-attention across channels.
Team McMaster adopted a hybrid model integrating Swin-FSR with simple CNN layers, utilizing a designed degradation process to add noise and improve robustness.
Team NUDT RSR addressed the degradation pipeline, model design, and model supervision, achieving significant performance by fusing frequency domain and spatial domain information through Spatially-Adaptive Feature Modulation.
The proposed methods demonstrate the ability to increase the resolution and details of RAW images while reducing blurriness and noise, without introducing color artifacts. The challenge results suggest that RAW image super-resolution can be solved similarly to RAW denoising, but more realistic downsampling remains an open challenge.
他の言語に翻訳
原文コンテンツから
arxiv.org
深掘り質問