Text Detection in Natural Scene Images by Stroke Gabor Words

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Text detection in natural scenes is challenging due to complex backgrounds and diverse text appearances.
• Existing methods often struggle with variations in font, color, and scale.
• Understanding stroke components is crucial for robust text recognition.

Purpose of the Study:

• To propose a novel algorithm for accurate text region detection in natural scene images.
• To leverage stroke components and descriptive Gabor filters for enhanced text identification.
• To develop a method robust to variations in text patterns and background complexity.

Main Methods:

• Utilizing stroke components as fundamental units of text characters.
• Employing Gabor filters to analyze and describe stroke components.
• Applying K-means clustering to define Stroke Gabor Words (SGWs) for universal stroke description.
• Implementing a suitability measurement for Gabor filter confidence and image window analysis.
• Using heuristic layout analysis and SGW characteristic distributions for text/non-text classification.

Main Results:

• The algorithm successfully detects text regions in natural scene images.
• Demonstrated robustness against complex backgrounds and variations in text font, color, and scale.
• Achieved effective classification of text and non-text windows using SGW distributions.
• Validated performance on benchmark datasets.

Conclusions:

• The proposed algorithm offers a novel and effective approach for text detection in challenging natural scenes.
• The combination of stroke components and Gabor filters provides a powerful descriptor for text.
• The method shows significant potential for real-world applications requiring robust text recognition.