# Character Region Awareness for Text Detection

**Authors:** Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, Hwalsuk Lee  
**arXiv:** [1904.01941](https://arxiv.org/abs/1904.01941)  
**Date:** 2019-04-03  
**Summary page:** https://www.summarizepaper.com/en/paper/1904.01941v1/

## Key points

- Neural network-based methods have improved scene text detection
- Previous methods trained with rigid word-level bounding boxes have limitations in representing text regions in arbitrary shapes
- The proposed method, called Character Region Awareness for Text Detection, explores each character and affinity between characters to effectively detect text areas
- Lack of individual character level annotations is overcome by exploiting given character-level annotations for synthetic images and estimated character-level ground-truths for real images acquired by the learned interim model
- The network is trained with a newly proposed representation for affinity to estimate the relationship between characters
- Extensive experiments on six benchmarks showed that their character-level text detection significantly outperformed state-of-the-art detectors
- The proposed method guarantees high flexibility in detecting complicated scene text images such as arbitrarily-oriented, curved or deformed texts.

## Summary

The field of scene text detection has seen recent advancements with the emergence of neural network-based methods that have shown promising results. To address the limitations of previous methods trained with rigid word-level bounding boxes in representing text regions in an arbitrary shape, a team of researchers led by Youngmin Baek proposed a new scene text detection method called Character Region Awareness for Text Detection. This method explores each character and affinity between characters to effectively detect text areas. To overcome the lack of individual character level annotations, the proposed framework exploits both given character-level annotations for synthetic images and estimated character-level ground-truths for real images acquired by the learned interim model. The network is trained with a newly proposed representation for affinity to estimate the relationship between characters. The team conducted extensive experiments on six benchmarks, including TotalText and CTW-1500 datasets which contain highly curved texts in natural images. The results showed that their character-level text detection significantly outperformed state-of-the-art detectors. Additionally, their proposed method guarantees high flexibility in detecting complicated scene text images such as arbitrarily-oriented, curved or deformed texts. Overall, this research presents a novel approach to scene text detection that addresses some of the limitations of previous methods and demonstrates superior performance on various benchmarks.