Scene text detection with fully convolutional neural networks

摘要

Text detection in scene image has become a hot topic in computer vision and artificial intelligence research, due to its wide range of applications and challenges. Most state-of-the-art methods for text detection based on deep learning rely on text bounding box regression. These methods can not well handle the case that if the scene text is curved. In this paper, we propose a new framework for arbitrarily oriented text detection in natural images based on fully convolutional neural networks. The main idea is to represent a text instance by two forms: text center block and word stroke region. These two elements are detected by two fully convolutional networks, respectively. Final detections are produced by the word region surrounding box algorithm. The proposed method does not need to regress the extant bounding box of the text instance, mainly because the predicted text block region itself implicitly contains position and orientation information. Besides, our method can well handle text in different languages, arbitrary orientations, curved shape and various fonts. To validate the effectiveness of the proposed method, we perform experiments on three public datasets: MSRA-TD500, USTB-SV1K and ICDAR2013, and compare it with other state-of-the-art methods. Experiment results demonstrate that the proposed method achieves competitive results. Based on VGG-16, our method achieves an F-measure of 78.84% on MSRA-TD500, 59.34% on USTB-SV1K, and 88.21% on ICDAR2013.