Arbitrary-Oriented Scene Text Detection via Rotation Proposals

Abstract

This paper introduces a novel rotation-based framework for arbitrary-oriented\ntext detection in natural scene images. We present the Rotation Region Proposal\nNetworks (RRPN), which are designed to generate inclined proposals with text\norientation angle information. The angle information is then adapted for\nbounding box regression to make the proposals more accurately fit into the text\nregion in terms of the orientation. The Rotation Region-of-Interest (RRoI)\npooling layer is proposed to project arbitrary-oriented proposals to a feature\nmap for a text region classifier. The whole framework is built upon a\nregion-proposal-based architecture, which ensures the computational efficiency\nof the arbitrary-oriented text detection compared with previous text detection\nsystems. We conduct experiments using the rotation-based framework on three\nreal-world scene text detection datasets and demonstrate its superiority in\nterms of effectiveness and efficiency over previous approaches.\n

Affiliated Institutions

• Fudan University CN
• Shanghai Advanced Research Institute CN

Related Publications