Binarization of Degraded Document Images Using Deep Learning Techniques

Abstract

Binarization is a fundamental preprocessing step in image processing and analysis, particularly in the context of degraded historical document images. Over time, such documents often suffer from various forms of deterioration due to poor storage conditions and environmental factors, making reliable binarization essential for subsequent tasks like enhancement, recognition, and archiving. Traditional binarization techniques, widely referred to as thresholding methods, operate by selecting one or more global or local thresholds to separate text (foreground) from background. While these approaches have been extensively studied, they often fall short when dealing with complex degradations such as ink bleed-through, uneven illumination, or textured backgrounds. To address these limitations, this project proposes a deep learning-based binarization approach that eliminates the reliance on thresholding. Specifically, we employ Convolutional Neural Networks (CNNs) for pixel-level classification, leveraging their ability to learn discriminative features directly from data. CNNs have proven effective in capturing structural and contextual information, making them particularly suitable for the challenges posed by degraded document images. The proposed system was trained and evaluated on a diverse set of annotated historical documents. Experimental results demonstrate that our approach achieves competitive and often superior performance compared to classical and recent methods, as measured by evaluation metrics commonly used in international benchmarking competitions. These outcomes confirm the robustness and generalization ability of the proposed method, establishing its potential for real-world applications in document image analysis and digitization.