Learning to predict more accurate text instances for scene text detection

At present, multi-oriented text detection methods based on deep neural network have achieved promising performances various benchmarks. Nevertheless, there are still some difficulties for arbitrary shape detection, especially a simple and proper representation of instances. In this paper, pixel-based detector is proposed to facilitate the prediction instances with shapes in manner. Firstly, alleviate influence target vertex sorting achieve direct regression instances, starting-point-independent coordinates loss proposed. Furthermore, predict more accurate instance accuracy as an assistant task refine predicted under guidance IoU. To evaluate effectiveness our detector, extensive experiments been carried public benchmarks which contain We obtain 84.8% F-measure Total-Text benchmark. The results show that method can reach state-of-the-art performance.