Abstract

In this paper, we introduce a novel end-end framework for multi-orientedscene text detection from an instance-aware semantic segmentation perspective.We present Fused Text Segmentation Networks, which combine multi-level featuresduring the feature extracting as text instance may rely on finer featureexpression compared to general objects. It detects and segments the textinstance jointly and simultaneously, leveraging merits from both semanticsegmentation task and region proposal based object detection task. Notinvolving any extra pipelines, our approach surpasses the current state of theart on multi-oriented scene text detection benchmarks: ICDAR2015 IncidentalScene Text and MSRA-TD500 reaching Hmean 84.1% and 82.0% respectively. Morever,we report a baseline on total-text containing curved text which suggestseffectiveness of the proposed approach.