Reproduce “Omni Aggregation Networks for Lightweight Image Super-Resolution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, Jinfan Liu
Researchers from Shanghai Jiao Tong University and Huawei HiSilicon introduce Omni-SR, a lightweight Vision Transformer for image super-resolution that achieves state-of-the-art performance with high efficiency. Omni-SR obtains a PSNR of 26.95dB on Urban100 (×4 upscaling) with 792K parameters, consuming 28% fewer FLOPs than previous methods.
Key insights
- Simultaneously modeling interactions across both spatial and channel dimensions ("omni-axis") through covariance-based attention yields richer feature representations and improved optimization.
- Multi-scale feature aggregation (local, meso, global) effectively expands the receptive field in shallow networks, enabling the capture of diverse texture patterns essential for high-quality super-resolution.
- The proposed architectural components, Omni Self-Attention (OSA) and Omni-Scale Aggregation Group (OSAG), provide a compelling trade-off between performance and computational efficiency for lightweight Vision Transformers in image restoration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.