A Novel Method for Curved Scene Text Detection

Scene text detection in the wild has gained momentum due to the rapid development of deep learning. Previous methods have excellent performances on detection horizontal and multi-oriented text, thereby the research focus gradually shift into curved text detection. To resolve this problem, we propose a novel curved text detection method by combining segmentation with regression in this paper. For each curved text instance, boundary points, text central region is predicted through a Fully Convolution Network with multi-layer feature fusion to locate text. Experiment results on two public benchmarks for curved scene text detection demonstrate that our method obtains accurate description of text boundary and achieves state-of-art performance.