Scene Text Detection Using a Modified SkipConnection Network

Scene text detection, unlike documentary textdetection, includes challenges such as a variety of text sizes,backgrounds, fonts, and the proximity of two or more words toeach other. In this paper, we first investigated the effect of skipconnections in a fully convolutional network on text detectionand found that adding SkipConv modules in the path of skipconnections improves the performance of the network for textdetection. We also proposed a post-processing operation on theoutput of the network, which can easily separate closely relatedtexts.

Keywords:

scene text detection , skip-connection , receptivefield