Tracking Based Multi-Orientation Scene Text Detection: A Unified Framework With Dynamic Programming.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Multi-orientation text detection in scene videos faces challenges like skew, low contrast, and motion.
• Conventional frame-by-frame methods offer limited performance for these complex scenarios.

Purpose of the Study:

• To develop a novel tracking-based method for multi-orientation scene text detection in videos.
• To enhance text detection accuracy by utilizing information from multiple frames within a unified framework.

Main Methods:

• A multi-information fusion approach is used for text detection within individual frames, identifying character candidates and text regions.
• Dynamic programming is employed to learn optimal tracking trajectories across consecutive frames, refining detection results.
• The method integrates detection, recognition, and prediction information for comprehensive analysis.

Main Results:

• The proposed system achieves state-of-the-art performance on public datasets for multi-orientation scene text detection.
• The tracking-based approach effectively addresses challenges posed by skew, low contrast, and motion in scene videos.

Conclusions:

• The novel tracking-based framework significantly improves multi-orientation scene text detection in videos.
• Integrating multi-frame information and dynamic programming offers a robust solution for complex video text analysis.