A Two-Stage Method for Text Line Detection in Historical Documents

TL;DR

This paper introduces a two-stage deep learning method for detecting text lines in historical documents, effectively handling complex layouts and curved lines, with high accuracy and open-source tools.

Contribution

A novel two-stage approach combining ARU-Net pixel labeling with bottom-up clustering for robust text line detection in historical documents.

Findings

Outperforms state-of-the-art methods in complex layout detection

Achieves an F-value of 0.922 on the cBAD benchmark

Requires fewer manually annotated images for training

Abstract

This work presents a two-stage text line detection method for historical documents. Each detected text line is represented by its baseline. In a first stage, a deep neural network called ARU-Net labels pixels to belong to one of the three classes: baseline, separator or other. The separator class marks beginning and end of each text line. The ARU-Net is trainable from scratch with manageably few manually annotated example images (less than 50). This is achieved by utilizing data augmentation strategies. The network predictions are used as input for the second stage which performs a bottom-up clustering to build baselines. The developed method is capable of handling complex layouts as well as curved and arbitrarily oriented text lines. It substantially outperforms current state-of-the-art approaches. For example, for the complex track of the cBAD: ICDAR2017 Competition on Baseline Detection the F-value is increased from 0.859 to 0.922. The framework to train and run the ARU-Net is open source.