Zhu Yixing Du Jun

Abstract

Traditional text detection methods mostly focus on quadrangle text. In thisstudy we propose a novel method named sliding line point regression (SLPR) inorder to detect arbitrary-shape text in natural scene. SLPR regresses multiplepoints on the edge of text line and then utilizes these points to sketch theoutlines of the text. The proposed SLPR can be adapted to many object detectionarchitectures such as Faster R-CNN and R-FCN. Specifically, we first generatethe smallest rectangular box including the text with region proposal network(RPN), then isometrically regress the points on the edge of text by using thevertically and horizontally sliding lines. To make full use of information andreduce redundancy, we calculate x-coordinate or y-coordinate of target point bythe rectangular box position, and just regress the remaining y-coordinate orx-coordinate. Accordingly we can not only reduce the parameters of system, butalso restrain the points which will generate more regular polygon. Our approachachieved competitive results on traditional ICDAR2015 Incidental Scene Textbenchmark and curve text detection dataset CTW1500.