EAST (Zhou et al., CVPR 2017) collapsed the multi-stage pipelines of 2016 into one fully convolutional network. For every pixel of a down-sampled feature map it predicts whether the pixel lies inside text and, if so, the geometry of the word around it: either a rotated box (distances to the four edges plus an angle) or the four corners of a quadrilateral. Locality-aware non-maximum suppression merges the per-pixel guesses into final boxes.
Why it mattered
- One stage, no proposals. Compared with CTPN, there are no anchors, no recurrent layer and no line-building heuristics.
- Rotation. The RBOX output handles arbitrarily oriented words, which incidental scene text (ICDAR 2015) is full of.
- Speed. The abstract reports 13.2 FPS at 720p with an F-score of 0.782 on ICDAR 2015.
Limits
- Long text lines and very wide aspect ratios are hard because the receptive field of a pixel is limited; EAST tends to split long lines.
- Curved text is not representable by a rotated box or a quad.
- The word-level output needs a recognizer that copes with rotation.
Using it today
EAST is the detector in most “text detection with OpenCV” tutorials because OpenCV’s dnn module can load the frozen TensorFlow graph directly and wraps the post-processing in cv2.dnn.TextDetectionModel_EAST. Our OpenCV guide shows the complete script next to the DB equivalent.
Citation
@inproceedings{zhou2017east,
title = {EAST: An Efficient and Accurate Scene Text Detector},
author = {Zhou, Xinyu and Yao, Cong and Wen, He and Wang, Yuzhi and Zhou, Shuchang and He, Weiran and Liang, Jiajun},
booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2017}
}