Shape awareness and structure-preserving network for arbitrary shape text detection

摘要

Recently, scene text detection has witnessed rapid advancement. However, there still exits two limitations: (1) boundary information is processed with color, texture information together inside a deep CNN, this however may not be ideal as they have different type of information relevant for adjacent text discrimination; (2) previous methods are lack of text structure preservation, which prevents network to accurately localization when enlarging receptive fields. In this paper, we propose two modules named Gate Convolution Module (GCM) and Tree Filter Module (TFM) respectively. GCM is a separate processing branch which leverages text shape information to split the close text instances. TFM models long-range dependencies while preserving the text details by exploiting the structural property of minimal spanning tree. Benefiting from two modules, our method effectively separates the text instances which are close to each other, while preserving detailed text structure. Extensive experiments on four standard text benchmarks (ICDAR2015, MSRA-TD500, CTW1500 and Total-Text) demonstrate that our method achieves the excellent performance.