TextMountain: Accurate scene text detection via instance segmentation

In this paper, we propose a novel scene text detection method named TextMountain. The key idea of TextMountain is making full use border-center information. Different from previous works that treat center-border as binary classification problem, predict probability (TCBP) and center-direction (TCD). TCBP just like mountain whose top center foot border. mountaintop can separate instances which cannot be easily achieved using semantic segmentation map its rising direction plan road to for each pixel on at the group stage. TCD helps learning better. Our label rules will not lead ambiguous problem with transformation angle, so proposed robust multi-oriented also handle well curved text. inference stage, needs search path process efficiently completed in parallel, yielding efficiency our compared others. experiments MLT, ICDAR2015, RCTW-17 SCUT-CTW1500 datasets demonstrate achieves better or comparable performance terms both accuracy efficiency. It worth mentioning an F-measure 76.85% MLT outperforms methods by large margin. Code made available.