Scene text detection via extremal region based double threshold convolutional network classification.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Text detection in natural images remains a challenging problem due to variations in scale, illumination, and background clutter.
• Existing methods often struggle with high recall rates and accurate classification of text candidates.

Purpose of the Study:

• To develop a robust text detection system for natural images.
• To improve the accuracy and recall rate of character candidate extraction and classification.
• To group detected characters into coherent text lines.

Main Methods:

• A saliency-enhanced MSER (Maximally Stable Extremal Regions) algorithm is used for character candidate extraction in a perception-based, illumination-invariant color space.
• A discriminative Convolutional Neural Network (CNN) is trained with multi-level information for classifying character candidates.
• Double threshold filtering and a recursive neighborhood search algorithm are employed for pruning non-text regions and tracking credible texts.
• Heuristic features are used to group characters into text lines.

Main Results:

• The proposed method achieves a high recall rate in character candidate extraction.
• The CNN classifier, leveraging multi-level information and double threshold filtering, effectively distinguishes text from non-text.
• The recursive search algorithm successfully prunes false positives.
• The final text line grouping demonstrates competitive performance on public datasets.

Conclusions:

• The presented text detection approach offers a robust and effective solution for natural images.
• The integration of saliency detection with MSER and a multi-level CNN classifier significantly enhances detection accuracy.
• The method demonstrates state-of-the-art performance compared to existing approaches on ICDAR 2011 and ICDAR 2013 datasets.