This paper introduces a novel rotation-based framework for arbitrary-orientedtext detection in natural scene images. We present the Rotation Region ProposalNetworks (RRPN), which is designed to generate inclined proposals with textorientation angle information. The angle information is then adapted forbounding box regression to make the proposals more accurately fit into the textregion in orientation. The Rotation Region-of-Interest (RRoI) pooling layer isproposed to project arbitrary-oriented proposals to the feature map for a textregion classifier. The whole framework is built upon region proposal basedarchitecture, which ensures the computational efficiency of thearbitrary-oriented text detection comparing with previous text detectionsystems. We conduct experiments using the rotation-based framework on threereal-world scene text detection datasets, and demonstrate its superiority interms of effectiveness and efficiency over previous approaches.
展开▼
