Irregular scene text detection via attention guided border labeling

Scene text detection plays an important role in many computer vision applications. With the help of recent deep learning techniques, multi-oriented text detection that was considered to be quite challenging has been solved to some extent. However, most existing methods still perform poorly for curved text detection, mainly due to the limitation of their text representations (e.g., horizontal boxes, rotated rectangles or quadrangles). To solve this problem, we propose a novel method to detect irregular scene texts based on instance-aware segmentation. The key idea is to design an attention guided semantic segmentation model to precisely label the weighted borders of text regions. Experiments conducted on several widely-used benchmarks demonstrate that our method achieves superior results on curved text datasets (i.e., with F-score 80.1% and 78.8% for the CTW1500 and Total-Text, respectively) and obtains comparable performance on multi-oriented text datasets compared to the state-of-the-art approaches.