Enhanced Text Detection in Natural Scenes using Advanced Machine Learning Techniques

Abstract

Text detection in natural scenes remains a fundamental challenge in computer vision, impacting applications from mobile navigation to document digitization. Traditional methods struggle with varying text orientations, complex backgrounds, and inconsistent lighting, while recent deep-learning approaches face computational efficiency challenges. This paper presents a novel hybrid machine-learning framework that combines traditional computer vision with advanced machine learning to achieve robust text detection. The framework integrates optimized preprocessing techniques, feature extraction methods, including Histogram Oriented Gradients (HOG) and Maximally Stable Extremal Regions (MSER), and a lightweight convolutional neural network for improved accuracy and efficiency. Experimental evaluation on benchmark datasets demonstrates superior performance, achieving 98% precision, 97.5% recall, and 97.8% F1-score, while maintaining real-time processing capabilities at 45 fps. The framework significantly outperforms existing methods in handling diverse text scenarios, establishing a new standard for natural scene text detection. This research contributes to the advancement of text detection technology and offers practical applications in augmented reality, autonomous navigation, and document processing systems.