A real-time arbitrary-shape text detector

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Detecting arbitrary-shape text in natural scenes presents significant challenges.
• Existing methods often lack real-time capabilities or practical applicability.

Purpose of the Study:

• To develop an effective and real-time detector for arbitrary-shape text.
• To improve upon current limitations in arbitrary-shape text detection.

Main Methods:

• A YOLOv6-based architecture was adapted for text detection.
• Two additional branches were integrated into the YOLOv6 neck.
• The pixel aggregation (PA) algorithm was utilized for the detection head.

Main Results:

• The proposed method achieved high detection accuracy and real-time performance.
• On Total-Text (640x640), an F-measure of 84.1% at 291.8 FPS was recorded.
• On ICDAR2015 (896x896), an F-measure of 81.5% at 199.6 FPS was achieved.

Conclusions:

• The YOLOv6-based detector effectively addresses arbitrary-shape text detection challenges.
• The method demonstrates superior performance in both accuracy and speed compared to existing approaches.