DAN: A Segmentation-Free Document Attention Network for Handwritten Document Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Document Analysis

Background:

• Unconstrained handwritten text recognition is complex, typically requiring separate line segmentation and recognition steps.
• Existing methods rely heavily on accurate segmentation, which can be a bottleneck for performance.

Purpose of the Study:

• To propose a novel end-to-end, segmentation-free architecture for handwritten document recognition.
• To develop a system capable of recognizing text and logical layout tokens simultaneously.

Main Methods:

• The Document Attention Network (DAN) utilizes a Fully Convolutional Network (FCN) encoder for feature extraction.
• A stack of transformer decoder layers enables recurrent token-by-token prediction of characters and layout tags.
• The model is trained without segmentation labels, taking entire documents as input.

Main Results:

• Achieved competitive character error rates (CER) on the READ 2016 dataset at page (3.43%) and double-page (3.70%) levels.
• Attained a CER of 4.54% on the RIMES 2009 dataset at the page level.
• Demonstrated the effectiveness of a segmentation-free approach for handwritten text recognition.

Conclusions:

• The Document Attention Network offers a viable and competitive alternative to traditional segmentation-based methods.
• This end-to-end approach simplifies the handwritten text recognition pipeline.
• The released source code and pre-trained models facilitate further research and application.