Detecting Oriented Text in Natural Images by Linking Segments

Most state-of-the-art text detection methods are specific to horizontal Latintext and are not fast enough for real-time applications. We introduce SegmentLinking (SegLink), an oriented text detection method. The main idea is todecompose text into two locally detectable elements, namely segments and links.A segment is an oriented box covering a part of a word or text line; A linkconnects two adjacent segments, indicating that they belong to the same word ortext line. Both elements are detected densely at multiple scales by anend-to-end trained, fully-convolutional neural network. Final detections areproduced by combining segments connected by links. Compared with previousmethods, SegLink improves along the dimensions of accuracy, speed, and ease oftraining. It achieves an f-measure of 75.0% on the standard ICDAR 2015Incidental (Challenge 4) benchmark, outperforming the previous best by a largemargin. It runs at over 20 FPS on 512x512 images. Moreover, withoutmodification, SegLink is able to detect long lines of non-Latin text, such asChinese.