BorderNet: An Efficient Border-Attention Text Detector
Recently, segmentation-based text detection methods are quite popular in the scene text detection field, because of their superiority for text instances with arbitrary shapes and extreme aspect ratios. However, the vast majority of the existing segmentation-based methods are difficult to detect curved and dense text instances due to principle of these methods. In this paper, we propose a novel text detection method named BorderNet. The key idea of BorderNet is making full use of border-center information to detect the curve and dense text. Furthermore, a efficient Multi-Scale Feature Enhancement Module is proposed to improve the scale and shape robustness by enhancing features of different scales adaptively. Our method outperforms SOTA on multiple datasets, achieving 89% accuracy on ICDAR2015 and 87.1% accuracy on Total-Text. What's more, we can maintain 84.5% accuracy on DAST1500.