RECURRENT GLOBAL CONVOLUTIONAL NETWORK FOR SCENE TEXT DETECTION

Scene text detection is an important as well as challenging problem in computer vision. Text information plays an important role in scene understanding, image indexing, and indoor navigation. Deep neural networks are being widely used due to their capability to learn strong text features. However, the state-of-the-art scene text detection methods capture weak text features in the early layers with no scope to improve the captured features. In this paper we propose a novel recurrent architecture to improve the learnings of a feature map at a given time, by using global and local information from the same feature map at the previous time, for detecting occluded texts and long words. We design the recurrent text convolutional layer for seamless integration of recurrent architecture with local-global map. The experimental results on publicly available scene text datasets show the efficiency of our framework. We also create our own scene text dataset.