Refined Single Shot Text Detector with Cascaded Regression in Complex Environments

Inhalt:
In this paper, we propose a novel method call Refine Single Shot Text Detector (RSSTD) for arbitrary-oriented text detection in complex environments. The model consists two inter-connected modules called coarse detection module (CDM) and fine detection module (FDM). The CDM performs negative anchor filtering to provide better initialization for the FDM, in which the accurate text boxes are regressed based on the refined anchors. This results in a end -to-end trained model that essentially works in a coarse-to-fine manner. Meanwhile, we utilize a pixel IoU (PIoU) loss which is well-suited for oriented boxes. The proposed method is evaluated on two public datasets, namely ICDAR 2013 and 2015. Experimental results demonstrate its superiority over several state-of-the-art bounding box-based approaches.