# SnowFormer: Context Interaction Transformer with Scale-awareness for Single Image Desnowing

**Authors:** Sixiang Chen, Tian Ye, Yun Liu, Erkang Chen  
**arXiv:** [2208.09703](https://arxiv.org/abs/2208.09703)  
**Date:** 2022-08-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2208.09703v3/

## Key points

- Snow degradations are a significant challenge for single image desnowing, requiring advanced image restoration techniques.
- SnowFormer is a novel transformer model developed by Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen to address snow degradation challenges.
- SnowFormer leverages efficient cross-attentions for local-global context interaction across patches, surpassing traditional methods that rely on local operators or vanilla transformers.
- Unlike recent Vision Transformers for image restoration, SnowFormer incorporates a multi-head cross-attention mechanism to establish effective local-global context interaction between scale-aware snow queries and local-patch embeddings.
- SnowFormer generates snow queries from aggregated scale-aware features using a query generator, resulting in superior restoration outcomes compared to existing desnowing methods and universal image restoration techniques.
- In evaluations on six synthetic and real-world datasets, SnowFormer outperformed state-of-the-art desnowing networks and prevalent universal image restoration transformers.
- The code for SnowFormer is publicly available at https://github.com/Ephemeral182/SnowFormer.

## Summary

, , , , 

Snow degradations pose a significant challenge for single image desnowing, requiring advanced image restoration techniques. Traditional methods have struggled to effectively address this issue, prompting the development of SnowFormer, a novel transformer model proposed by Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen. SnowFormer leverages efficient cross-attentions to facilitate local-global context interaction across patches, surpassing previous approaches that rely on local operators or vanilla transformers. Unlike recent image restoration Vision Transformers that utilize multi-head self-attention, SnowFormer incorporates a multi-head cross-attention mechanism. This enables the model to establish effective local-global context interaction between scale-aware snow queries and local-patch embeddings. Furthermore, SnowFormer generates snow queries from aggregated scale-aware features using a query generator. These features are rich in potential clean cues, resulting in superior restoration outcomes compared to existing desnowing methods and universal image restoration techniques. In evaluations on six synthetic and real-world datasets, SnowFormer outperformed state-of-the-art desnowing networks and prevalent universal image restoration transformers. The code for SnowFormer is publicly available at https://github.com/Ephemeral182/SnowFormer. This innovative approach not only addresses the complexities of snow degradations but also showcases the potential of transformer models in enhancing image restoration tasks.