Accurate Arbitrary-Shaped Scene Text Detection via Iterative Polynomial Parameter Regression

Jiahao Shi, Long Chen, Feng Su; Proceedings of the Asian Conference on Computer Vision (ACCV), 2020

A number of scene text in natural images have irregular shapes which often cause significant difficulties for a text detector. In this paper, we propose a robust scene text detection method based on a parameterized shape modeling and regression scheme for text with arbitrary shapes. The shape model geometrically depicts a text region with a polynomial centerline and a series of width cues to capture global shape characteristics (e.g. smoothness) and local shapes of the text respectively for accurate text localization, which differs from previous text region modeling schemes based on discrete boundary points or pixels. We further propose a text detection network PolyPRNet equipped with an iterative regression module for text's shape parameters, which effectively enhances the detection accuracy of arbitrary-shaped text. Our method achieves state-of-the-art text detection results on several standard benchmarks.