|
In recent years, scene text detection task has made great progress. However, most of works are devoted to improving the performance of detector and pay little attention to the practical applications. In this paper, we propose to exploit a mask branch to detect arbitrary-shaped text accurately, and compress the model to reduce computation and storage costs, achieving a balance between speed and accuracy. Specifically, we distill intermediate features and text proposal classification to transfer dark knowledge to student text detector. In the distillation process, we treat textual and Background features differently and decouple positive and negative text proposals. Experimental results on the ICDAR 2015 and ICDAR 2017 MLT datasets demonstrate the superiority of our lightweight scene text detector. |
|
Keywords:Deep Learning; Scene Text Detection; Arbitrary-shaped Text; Knowledge Distillation |
|