PRPN: Progressive region prediction network for natural scene text detection

Abstract

With the development of deep learning, scene text detection methods have made great progress in recent years. Most text detection methods are based on bounding box prediction with a 0-1 discrete distribution; thus, separating adjacent text instances is difficult. Direct prediction of the bounding box also renders difficult the detection various shapes of text, such as quadrangular text and curved text. In this work, we design a 2D progressive kernel for describing the progressive variety of text regions. It transforms the original ground truth (GT) of bounding boxes into the GT of a 0-1 progressive probability distribution. We also propose a novel progressive region prediction network (PRPN) with directional pooling for predicting the probability distributions of text regions. Then, a postprocessing algorithm is used to transform the probability distributions of the text regions into bounding box output for text detection. Experiments on standard datasets, including ICDAR 2013, ICDAR 2015, MSRA-TD500, and SCUT-CTW1500, demonstrate that the proposed method outperforms state-of-the-art methods in terms of accuracy and robustness. The method obtains an F-measure of 86.0% on ICDAR 2015 and 81.4% on SCUT-CTW1500. The code is available at https://github.com/xinyu-ch/ProgressiveTextDetection.