Arbitrarily Shaped Scene Text Detection with a Mask Tightness Text Detector.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Scene text detection is challenging due to arbitrary text shapes, fonts, and sizes.
• Detecting curved and irregularly shaped text remains a significant hurdle in the field.
• Existing datasets and methods struggle with the accurate detection of complex text geometries.

Purpose of the Study:

• To propose an improved method for detecting arbitrarily shaped scene text.
• To enhance the performance of text detection, particularly for curved and irregular text.
• To introduce a novel approach that addresses limitations in current scene text detection techniques.

Main Methods:

• Developed the Mask Tightness Text Detector (Mask TTD) incorporating tightness prior and text frontier learning.
• Enhanced pixel-wise mask prediction for more accurate text region localization.
• Integrated a polygonal boundary prediction branch for improved detection of arbitrarily shaped text.

Main Results:

• Achieved state-of-the-art performance on curved text datasets (CTW1500, Total-text, CUTE80) and common benchmarks (RCTW-17, MSRA-TD500, ICDAR 2015).
• Outperformed previous methods on CTW1500 by 16% at an IoU threshold of 0.8, demonstrating superior tight text detection.
• Showcased powerful generalization ability on the large Chinese dataset RCTW-17, achieving significant improvements in Average Precision and F-measure.

Conclusions:

• Mask TTD effectively addresses the challenge of detecting arbitrarily shaped scene text.
• The proposed method offers significant improvements in accuracy and robustness for complex text detection scenarios.
• Mask TTD demonstrates strong potential for real-world applications requiring precise scene text recognition.