Multi-Oriented Scene Text Detection at Character Level

Recent scene text detection methods perform superior on benchmark datasets using deep-learning frameworks. In this paper, we re-implement the state-of-the-art text detection method, character region awareness for text detection (CRAFT), which can detect individual characters of scene text images. CRAFT is a character-based detection method with many advantages in detecting complex text by detecting character units and estimating the area between characters, capable of detecting texts of any shape. In the other words, we improve the detection performance of the baseline method, CRAFT, by some modifications in its architecture and proposing a training scheme that takes benefit of the advanced optimizer. The performance improvements of CRAFT are validated on three benchmark datasets: ICDAR2013, ICDAR2015, and COCO-Text. By applying the pre-trained models on COCO-Text, CRAFT shows that it cannot generalize without fine-tuning. We also improve the ICDAR2015 model and evaluate it on benchmark datasets. The evaluation results show improved precision performance compared to the original pre-trained model with fewer iterations and higher accuracy.