# Attention in Attention Network for Image Super-Resolution

**Authors:** Haoyu Chen, Jinjin Gu, Zhi Zhang  
**arXiv:** [2104.09497](https://arxiv.org/abs/2104.09497)  
**Date:** 2021-04-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2104.09497v3/

## Key points

- Single image super-resolution (SISR) has seen advancements due to convolutional neural networks
- Attention mechanisms are crucial for high-performance SISR models
- Not all attention modules are equally beneficial
- The authors propose an attention in attention network (A$^2$N) for more efficient and accurate SISR
- A$^2$N comprises a non-attention branch and a coupling attention branch, with a dynamic attention module generating weights for these two branches to suppress unwanted attention adjustments dynamically
- The weights change adaptively according to input features, allowing attention modules to specialize without penalties and thus greatly improve the capacity of the network with few parameters overhead.
- Experimental results show that A$^2$N outperforms state-of-the-art networks of similar sizes in terms of trade-off performance.
- The authors provide codes for their final model on GitHub which enables further research into this area.

## Summary

The field of single image super-resolution (SISR) has seen remarkable advancements in recent years, largely due to the use of convolutional neural networks. Attention mechanisms have been identified as crucial for high-performance SISR models, but their workings remain unclear. In this study, the authors attempt to quantify and visualize attention mechanisms in SISR and demonstrate that not all attention modules are equally beneficial. They propose an attention in attention network (A$^2$N) for more efficient and accurate SISR. A$^2$N comprises a non-attention branch and a coupling attention branch, with a dynamic attention module generating weights for these two branches to suppress unwanted attention adjustments dynamically. The weights change adaptively according to input features, allowing attention modules to specialize without penalties and thus greatly improve the capacity of the network with few parameters overhead. Experimental results show that A$^2$N outperforms state-of-the-art networks of similar sizes in terms of trade-off performance. The authors provide codes for their final model on GitHub which enables further research into this area. Overall, this study sheds light on the workings of attention mechanisms in SISR and proposes an effective approach for incorporating them into neural networks.