Historical Text Line Segmentation Using Deep Learning Algorithms: Mask-RCNN against U-Net Networks.

Area of Science:

• Computer Vision
• Document Image Analysis
• Artificial Intelligence

Background:

• Text line segmentation is crucial for text transcription algorithms.
• Current U-Net-based deep learning networks (ARU-Net, dhSegment, Doc-UFCN) require post-processing for instance segmentation.
• Mask R-CNN offers direct instance segmentation capabilities.

Purpose of the Study:

• To compare the performance of Mask R-CNN against U-Net-based networks for text segmentation in historical documents.
• To evaluate the impact of different segmentation networks on handwritten text recognition (HTR).

Main Methods:

• Direct comparison of Mask R-CNN with U-Net-based networks (ARU-Net, dhSegment, Doc-UFCN) on historical document datasets.
• Evaluation using standard line segmentation metrics.
• Assessment of Mask R-CNN's effectiveness with light mask processing.
• Comparison of HTR performance using outputs from different segmentation networks.

Main Results:

• Mask R-CNN demonstrated superior performance compared to ARU-Net, dhSegment, and Doc-UFCN in text line segmentation metrics.
• Performance evaluation benefits from light mask processing rather than solely relying on raw masks.
• Mask R-CNN integration resulted in enhanced handwritten text recognition (HTR) accuracy.

Conclusions:

• Mask R-CNN is more effective for direct text line instance segmentation in historical documents than U-Net-based architectures.
• Optimized mask processing improves segmentation evaluation.
• Employing Mask R-CNN for segmentation positively impacts downstream HTR tasks.