A Hierarchical Scene Text Detector Concerning Hard Examples

Abstract

Nowadays, the scene text detection approaches based on deep learning have already achieved promising performances. However, due to the insufficient use of the intermediate feature layer information of the neural network, and the learning of the indistinguishable scale and hard examples, it is difficult to improve the accuracy and the recall. Therefore, based on the Fully Convolutional Networks (FCNs), we propose a hierarchical scene text detector concerning hard examples. Specifically, our proposed detector HST-DHE directly predicts text lines of arbitrary orientations in full images without pre-defined anchor boxes. And the feature pyramid is constructed to predict the features at different levels to make full use of the feature information in the middle layers. At the same time, the loss function is redesigned to make the network focus on hard examples and further improve the precision of text detection. Experiments on standard datasets including ICDAR 2013, ICDAR 2015 and MSRA-TD500 demonstrate that the proposed algorithm significantly outperforms state-of-the-art detectors. According to the results, our proposed method has stronger robustness in multi-scale and multi-direction natural scene images.