In this paper, we present a new Mask R-CNN based text detection approachwhich can robustly detect multi-oriented and curved text from natural sceneimages in a unified manner. To enhance the feature representation ability ofMask R-CNN for text detection tasks, we propose to use the Pyramid AttentionNetwork (PAN) as a new backbone network of Mask R-CNN. Experiments demonstratethat PAN can suppress false alarms caused by text-like backgrounds moreeffectively. Our proposed approach has achieved superior performance on bothmulti-oriented (ICDAR-2015, ICDAR-2017 MLT) and curved (SCUT-CTW1500) textdetection benchmark tasks by only using single-scale and single-model testing.