LICS: Locating Inter-Character Spaces for Multilingual Scene Text Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Natural Language Processing

Background:

• Multilingual scene text detection faces challenges due to language-specific features and extensive data requirements.
• Traditional methods struggle with diverse character shapes, orientations, deformations, and occlusions in natural scenes.

Purpose of the Study:

• To develop a language-agnostic method for feasible multilingual scene text detection.
• To reduce the annotation burden for scene text detection in new languages.

Main Methods:

• Introduced Locating Inter-Character Spaces (LICS), detecting language-agnostic inter-character gaps.
• Employed a two-stage approach: training on synthetic data with gap annotations, followed by weakly supervised learning on real-world data with word-level labels.
• Introduced the Character-Labeled Street View Text (CSVT) dataset with standardized annotation principles.

Main Results:

• LICS demonstrated strong performance, especially on Asian scripts, on ICDAR and Total-Text benchmarks.
• Weakly supervised learning framework eliminated the need for character-level annotations in target languages.
• The CSVT dataset provides a valuable resource for multilingual scene-text analysis research.

Conclusions:

• LICS offers a robust and efficient solution for multilingual scene text detection.
• The proposed weakly supervised approach significantly reduces annotation costs.
• The CSVT dataset will advance research in multilingual scene-text understanding.