TextFourierNet: Arbitrary-shaped Scene Text Detection Based on Fourier Contour Modeling
Fei Jiang, Zengfu Wang · 2022
Arbitrary-shaped scene text detection is a challenging task as the network can’t accurately learn the geometric layout of the texts such as curve shapes, various scales and random rotations. Most of the advanced methods address this problem through pixel-level classification. However, the classification-based approaches might lead to costly post-processing and fail to separate adjacent scene text instances. To tackle this issue, we propose a novel arbitrary-shaped scene text detector named TextFourierNet, which is a regression-based method. We model arbitrary-shaped scene text instances in polar coordinates and encode the complex geometric layouts into simple Fourier coefficients. TextFourierNet reconstructs text boundaries via an Inverse Fourier Transformation (IFT) and one NMS post-processing. To effectively separate adjacent texts and learn boundary information, we represent a text region as four sub-regions, which greatly improves the overall performance. Extensive experiments on two public datasets show that TextFourierNet is effective for arbitrary-shaped scene text detection.