MaskS R-CNN Text Detector

Abstract

Scene text detection and scene text recognition are important components of scene text recognition system. Scene text detection, the initial stage of scene text recognition, aims to find out text area in the picture. Recently the target detection method Mask R-CNN has been employed scene text detection and achieved good performance. In this paper, we set forth a model, MaskS R-CNN text detector, based on Mask R-CNN, which attempts to detect scene text. In this model, a network block of Mask Scoring R-CNN is introduced to learn the high quality of the predicted instance mask scores. The mask scoring mechanism correct the inconformity between mask quality and mask score, at the same time improves instance segmentation performance by attaching great importance to more accurate mask predictions. The method put forward in this paper can achieve multi-directional and multi-language natural scene text detection. Compared with some existing traditional location methods based on edge, color and texture and some location methods based on deep learning, it is a relatively innovative method.