# SVTR: Scene Text Recognition with a Single Visual Model

**Authors:** Yongkun Du, Zhineng Chen, Caiyan Jia, Xiaoting Yin, Tianlun Zheng, Chenxia Li, Yuning Du, Yu-Gang Jiang  
**arXiv:** [2205.00159](https://arxiv.org/abs/2205.00159)  
**Date:** 2022-04-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2205.00159v2/

## Key points

- Single Visual model for Scene Text Recognition (SVTR) proposed
- Eliminates need for sequential modeling, more efficient than traditional hybrid architectures
- SVTR decomposes image text into small patches called character components
- Hierarchical stages of component-level mixing, merging, and combining
- Global and local mixing blocks capture inter-character and intra-character patterns
- Character recognition achieved through simple linear prediction
- Experimental results show effectiveness of SVTR in English and Chinese tasks
- SVTR-L model achieves competitive accuracy in English and outperforms existing methods in Chinese with faster processing speed
- Introduces SVTR-T as a smaller yet effective model with appealing speed during inference
- Success of single visual model relies on ability to extract discriminative character features
- Capturing intra-character local patterns (e.g. stroke-like features)
- Capturing inter-character long-term dependence (language analogous knowledge)
- Previous feature extractors inadequately modeled these properties, CNN backbones better at capturing local correlations than global dependencies
- Innovative approach to scene text recognition using single visual model within patchwise image tokenization framework
- Proposed SVTR method effectively captures discriminative character features through component level mixing
- Achieves competitive accuracy in both English and Chinese tasks while maintaining efficiency

## Summary

The existing study proposes a Single Visual model for Scene Text Recognition (SVTR) within the patch-wise image tokenization framework. This approach eliminates the need for sequential modeling, making it more efficient compared to traditional hybrid architectures. SVTR decomposes an image text into small patches called character components and performs hierarchical stages of component-level mixing, merging, and combining. It incorporates global and local mixing blocks to capture inter-character and intra-character patterns, enabling multi-grained character component perception. Character recognition is achieved through simple linear prediction. Experimental results on English and Chinese scene text recognition tasks demonstrate the effectiveness of SVTR. The SVTR-L (Large) model achieves highly competitive accuracy in English and outperforms existing methods by a large margin in Chinese while maintaining faster processing speed. Additionally, the study introduces SVTR-T (Tiny), which is a smaller yet effective model that offers appealing speed during inference. The authors argue that the success of the single visual model relies on its ability to extract discriminative character features. This includes capturing both intra-character local patterns, such as stroke-like features that distinguish characters at a fine-grained level, and inter-character long-term dependence that captures language analogous knowledge about characters from a complementary perspective. Previous feature extractors have not adequately modeled these properties, with CNN backbones being better at capturing local correlations rather than global dependencies. In conclusion, this study presents an innovative approach to scene text recognition using a single visual model within a patchwise image tokenization framework. The proposed SVTR method effectively captures discriminative character features through component level mixing and achieves competitive accuracy in both English and Chinese tasks while maintaining efficiency. The authors highlight the importance of modeling both intra character local patterns and inter character long term dependence for successful scene text recognition applications. The code for SVTR is publicly available at https://github.com/PaddlePaddle/PaddleOCR .