Restormer: Efficient Transformer for High-Resolution Image Restoration

Restormer (Restoration Transformer) is an efficient Transformer architecture for high-resolution image restoration. The authors note that CNNs learn generalizable image priors from large-scale data but suffer from limited receptive fields and an inability to adapt to input content, whereas Transformers overcome these limitations but have computational complexity that grows quadratically with spatial resolution. To make Transformers feasible for restoration, they introduce several key design changes in the core building blocks, the multi-head attention and feed-forward network, so the model captures long-range pixel interactions while still handling large images.

With these efficient designs, Restormer achieves state-of-the-art results across a broad range of image restoration tasks, including image deraining, single-image motion deblurring, defocus deblurring on both single-image and dual-pixel data, and image denoising covering Gaussian grayscale and color denoising as well as real image denoising. By reconciling the modeling power of self-attention with the practical demands of high-resolution inputs, the work made Transformers a viable and leading approach for low-level vision, and its source code and pretrained models were released publicly.