Abstract
Scene text recognition is a hot topic in the field of computer vision. Inspired by the success of the Single Shot Multibox Detection (SSD) on generic object detection, the architecture of SSD is implemented on scene text detection. SSD does not do well on text detection, because scene text as an object is usually smaller than a generic object and SSD cannot detect small objects well. Thus, the statistic analysis for scene text is made. Based on statistic characteristics of scene text, we propose a method named Text-SSD to detect scene text. Moreover, in order to boost the detection accuracy, multi-scale image are used to learn the multi-scale models. The voting based non-maximum suppressing is made for a candidate text region. The experimental results show that our method achieved the state-of-the-art performance on the benchmark dataset ICDAR2013 in the detection accuracy. Moreover, when using a single model, our method achieves the fastest speed compared with several latest text detection method based on deep neural network. Thus, experimental results demonstrate our method is efficient on scene text detection.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call