[论文解读] EAST: An Efficient and Accurate Scene Text Detector

Xinyu Zhou, Cong Yao+5 more|arXiv (Cornell University)|Apr 11, 2017

Handwritten Text Recognition Techniques参考文献 58被引用 115

一句话总结

EAST 提出一种两阶段、全卷积文本检测器，直接从完整图像将文本区域预测为旋转矩形或四边形，达到最先进的准确率并具备高速度。

ABSTRACT

Previous approaches for scene text detection have already achieved promising performances across various benchmarks. However, they usually fall short when dealing with challenging scenarios, even when equipped with deep neural network models, because the overall performance is determined by the interplay of multiple stages and components in the pipelines. In this work, we propose a simple yet powerful pipeline that yields fast and accurate text detection in natural scenes. The pipeline directly predicts words or text lines of arbitrary orientations and quadrilateral shapes in full images, eliminating unnecessary intermediate steps (e.g., candidate aggregation and word partitioning), with a single neural network. The simplicity of our pipeline allows concentrating efforts on designing loss functions and neural network architecture. Experiments on standard datasets including ICDAR 2015, COCO-Text and MSRA-TD500 demonstrate that the proposed algorithm significantly outperforms state-of-the-art methods in terms of both accuracy and efficiency. On the ICDAR 2015 dataset, the proposed algorithm achieves an F-score of 0.7820 at 13.2fps at 720p resolution.

研究动机与目标

• Motivate a simple, end-to-end pipeline for scene text detection that avoids multiple intermediate steps.
• Directly predict word or text-line level regions in arbitrary orientations.
• Enable flexible geometry outputs (rotated boxes or quadrangles) with efficient processing.

提出的方法

• Use a lightweight Fully Convolutional Network to predict per-pixel text scores and geometry maps.
• Support two geometry representations: RBOX (axis-aligned box with rotation) and QUAD (quadrangle) with corresponding loss functions.
• Generate training labels by shrinking quadrilaterals for the score map and computing geometry targets per pixel.
• Train with a combination of score loss (balanced cross-entropy) and geometry loss (IoU-based for RBOX, scale-normalized smoothed-L1 for QUAD).
• Apply locality-aware NMS to merge nearby predictions efficiently in O(n) time in practice.

实验结果

研究问题

• RQ1 Can a two-stage FCN pipeline directly predict text regions without intermediate steps and still achieve state-of-the-art accuracy?
• RQ2 How do different geometry representations (RBOX vs QUAD) compare in accuracy and efficiency across diverse datasets?
• RQ3 What loss design and training strategies yield robust per-pixel text geometry predictions across scales?

主要发现

• Achieves strong accuracy and speed across ICDAR 2015, COCO-Text, and MSRA-TD500 benchmarks.
• On ICDAR 2015, F-score of 0.7820 at 13.2 FPS (720p); multi-scale F-score 0.8072.
• COCO-Text F-score of 0.3945; MSRA-TD500 F-score of 0.7608.
• Two-stage pipeline with end-to-end training outperforms prior methods in both accuracy and speed.
• Flexible geometry outputs: RBOX and QUAD yield competitive results with different base networks (PVANET, PVANET2x, VGG16).
• Locally aware NMS significantly reduces post-processing cost while maintaining accuracy.