End-to-End Handwritten Paragraph Text Recognition Using a Vertical Attention Network.

Area of Science:

• Computer Vision
• Machine Learning
• Natural Language Processing

Background:

• Handwritten text recognition (HTR) is a complex computer vision task.
• Traditional HTR systems use separate models for line segmentation and recognition, limiting performance.
• Paragraph-level recognition requires robust methods to handle unconstrained text.

Purpose of the Study:

• To propose a unified, end-to-end model for unconstrained handwritten paragraph text recognition.
• To improve the accuracy and efficiency of HTR systems by integrating segmentation and recognition.
• To achieve state-of-the-art performance on challenging paragraph-level HTR datasets.

Main Methods:

• Developed a unified end-to-end model employing hybrid attention for paragraph recognition.
• The model iteratively processes paragraph images line by line using an encoder-attention-decoder architecture.
• Implicit line segmentation is achieved through a recurrent attention module generating vertical weighted masks.

Main Results:

• Achieved state-of-the-art character error rates on three popular datasets: RIMES (1.91%), IAM (4.45%), and READ 2016 (3.59%).
• The unified model effectively handles unconstrained handwritten text at the paragraph level.
• Demonstrated superior performance compared to traditional two-model approaches.

Conclusions:

• The proposed hybrid attention model offers a more effective approach to handwritten paragraph recognition.
• End-to-end processing with implicit line segmentation significantly enhances recognition accuracy.
• The model's availability facilitates further research and application in HTR.