Learning Deeply Supervised Scene Text Detectors from Scratch

Abstract

In this paper, we propose deeply supervised scene text detector (DSTD), a framework that can be learned from scratch. Our proposed method mainly addresses two problems. The first one is that state-of-the-art text detectors rely heavily on the off-the-shelf pre-trained models, which leads to several limitations including inflexibility and domain mismatch. The second problem is that unlike general objects, scene text usually appear in arbitrary orientations. Text detection using horizontal bounding boxes is inaccurate. In DSTD, we propose to regress rotated rectangles directly from horizontal default boxes to deal with multi-oriented text. Furthermore, we abandon the heavy pre-trained model from the SSD framework and incorporate dense layer-wise connections, enabling the network to be learned from scratch. The proposed method is evaluated on two public datasets, namely ICDAR2013 and ICDAR2015. Experimental results demonstrate its superiority over several state-of-the-art approaches.