Measuring Human Perception to Improve Handwritten Document Transcription.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Automated transcription of modern handwriting is advanced, but historical documents present significant challenges.
• Deep neural networks (DNNs) are widely used for recognition tasks, but their error reduction can be improved.
• Psychophysical measurements offer insights into human visual perception that could benefit AI models.

Purpose of the Study:

• To incorporate psychophysical measurements of human visual perception into the loss function of DNNs for recognition tasks.
• To develop and assess a general enhancement strategy for deep learning-based document transcription systems.
• To improve the accuracy of handwritten document transcription, particularly for historical texts.

Main Methods:

• Developed a novel loss formulation integrating human visual perception data into DNN training.
• Applied this strategy to deep learning models for handwritten document transcription.
• Experimented on standard datasets (IAM, RIMES) and a new dataset of 9th-century Latin manuscripts.

Main Results:

• Demonstrated reliable performance improvements across three different network architectures on IAM and RIMES datasets.
• Successfully showed the feasibility of the approach on a novel dataset of historical Latin manuscripts.
• The proposed strategy effectively reduces errors in document transcription tasks.

Conclusions:

• Integrating human visual perception into DNN loss functions is a viable strategy for enhancing recognition tasks.
• The developed enhancement strategy offers a general solution applicable to various deep learning transcription systems.
• This research opens new avenues for improving the transcription of challenging historical documents.