# Restormer: Efficient Transformer for High-Resolution Image Restoration

**Authors:** Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang  
**arXiv:** [2111.09881](https://arxiv.org/abs/2111.09881)  
**Date:** 2021-11-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2111.09881v1/

## Key points

- Convolutional neural networks (CNNs) widely used in image restoration for learning generalizable image priors from large-scale data
- Emergence of Transformers showing significant performance gains on natural language and high-level vision tasks
- Restormer, an efficient Transformer model introduced by Syed Waqas Zamir's team, addresses limitations of CNNs and captures long-range pixel interactions in large images
- Achieves state-of-the-art results in various image restoration tasks including deraining, motion deblurring, defocus deblurring, and denoising
- Restormer's design modifications enable it to learn long-range dependencies while maintaining computational efficiency
- Focuses on developing an efficient Transformer model for handling high-resolution images in restoration tasks
- Overcomes computational bottlenecks associated with traditional Transformers through innovative design elements in multi-head self-attention mechanism
- Represents a significant advancement in high-resolution image restoration with potential as a valuable tool for researchers and practitioners
- Availability of source code and pre-trained models enhances accessibility for leveraging cutting-edge technology in enhancing visual quality

## Summary

In the field of image restoration, convolutional neural networks (CNNs) have been widely used due to their ability to learn generalizable image priors from large-scale data. However, recent advancements in neural architectures have led to the emergence of Transformers, which have shown significant performance gains on natural language and high-level vision tasks. While Transformers address some of the limitations of CNNs, such as limited receptive fields and inadaptability to input content, their computational complexity grows quadratically with spatial resolution, making them impractical for high-resolution image restoration tasks. To bridge this gap, a team of researchers led by Syed Waqas Zamir introduced an efficient Transformer model named Restormer. By implementing key design modifications in the building blocks like multi-head attention and feed-forward networks, Restormer is able to capture long-range pixel interactions while remaining applicable to large images. This novel approach enables Restormer to achieve state-of-the-art results in various image restoration tasks including deraining, single-image motion deblurring, defocus deblurring (single-image and dual-pixel data), and image denoising (Gaussian grayscale/color denoising and real image denoising). While other methods aim to reduce complexity by applying self-attention within local image regions using designs like Swin Transformer, these strategies limit context aggregation within local neighborhoods and may not be ideal for image restoration tasks. In contrast, Restormer's Transformer model can effectively learn long-range dependencies while maintaining computational efficiency. The proposed method focuses on developing an efficient Transformer model capable of handling high-resolution images for restoration tasks. By introducing innovative design elements into the multi-head self-attention mechanism, Restormer overcomes computational bottlenecks associated with traditional Transformers. The model's ability to capture long-range interactions and deliver superior performance across various image restoration challenges underscores its potential as a valuable tool in the field. Overall, Restormer represents a significant advancement in the realm of high-resolution image restoration through its efficient utilization of Transformer architecture and innovative design choices. The availability of source code and pre-trained models further enhances its accessibility for researchers and practitioners looking to leverage cutting-edge technology for enhancing visual quality in images.