Characters are the basic components of text. Accurate character detection plays an important role in text detection and recognition. Previous character detectors tackle characters as independent objects, without considering the meaningful context information among them. In this paper, we propose a new module named constrained relation module which utilizes both the geometric and contextual information to exploit the strong relationship between characters. With this module, we build a new network named constrained relation network for character detection and recognition. To the best of our knowledge it is the first work to utilize contextual information among texts for character detection in scene images. The module can improve the detection results by suppressing the confusing text-like regions and recalling the hard examples. Experiments on SynthText, ICDAR2013 and SCUT-PORU demonstrate the effectiveness of our method on both detection and recognition tasks.
展开▼
