TextFuse: Fusing Deep Scene Text Detection Models for Enhanced Performance

Scene text detection is a prerequisite for scene text recognition. However, most existing approaches to scene text detection only utilize single source of deep features, which is vulnerable to changes of feature distributions. In this work, we propose a feature fusion-based scene text detection algorithm that can fully utilize the complementary advantages of different sources of deep features and improve the detection performance on scene text images. More importantly, a new result fusion-based scene text detection algorithm is also proposed that comprehensively integrates the results from different text detection algorithms (EAST, CRPN, RRPN, TextBoxes++). The proposed algorithms are validated on several benchmark scene text datasets, including ICDAR2013, ICDAR2015, MSRA-TD500, RCTW, and ShopSign, which demonstrate that our approaches can obtain significantly better recall and F-score results than the baselines.