Lyu Pengyuan Yao Cong Wu Wenhao Yan Shuicheng Bai Xiang

Abstract

Previous deep learning based state-of-the-art scene text detection methodscan be roughly classified into two categories. The first category treats scenetext as a type of general objects and follows general object detection paradigmto localize scene text by regressing the text box locations, but troubled bythe arbitrary-orientation and large aspect ratios of scene text. The second onesegments text regions directly, but mostly needs complex post processing. Inthis paper, we present a method that combines the ideas of the two types ofmethods while avoiding their shortcomings. We propose to detect scene text bylocalizing corner points of text bounding boxes and segmenting text regions inrelative positions. In inference stage, candidate boxes are generated bysampling and grouping corner points, which are further scored by segmentationmaps and suppressed by NMS. Compared with previous methods, our method canhandle long oriented text naturally and doesn't need complex post processing.The experiments on ICDAR2013, ICDAR2015, MSRA-TD500, MLT and COCO-Textdemonstrate that the proposed algorithm achieves better or comparable resultsin both accuracy and efficiency. Based on VGG16, it achieves an F-measure of84.3% on ICDAR2015 and 81.5% on MSRA-TD500.