Multi-Oriented and Multi-Lingual Scene Text Detection with Direct Regression.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Scene text detection is crucial for computer vision applications.
• Challenges include diverse text orientations, languages, and complex backgrounds.
• Existing Convolutional Neural Network (CNN) object detection frameworks require adaptation for scene text.

Purpose of the Study:

• To develop an effective scene text detection framework for multi-oriented and multi-lingual text.
• To address the key tasks of text region localization and boundary regression.
• To achieve state-of-the-art performance on diverse scene text datasets.

Main Methods:

• Proposed a framework based on Fully Convolutional Network (FCN) with a bi-task prediction module.
• The module performs pixel-wise classification (text vs. non-text) and regression of quadrilateral text boundaries.
• Utilized Non-Maximum Suppression (NMS) for word-level detection and a novel grouping method for line-level detection.

Main Results:

• Achieved state-of-the-art performance on multiple benchmarks, including multi-oriented and multi-lingual scene text datasets.
• Demonstrated the effectiveness of the proposed bi-task prediction module.
• Ablation experiments validated key factors for high-performance CNN-based scene text detection.

Conclusions:

• The proposed FCN-based framework with a bi-task module significantly advances scene text detection capabilities.
• The method effectively handles challenges posed by multi-oriented and multi-lingual text in complex scenes.
• This work provides a robust foundation for future research in scene text recognition and analysis.