<p>Preserving and analyzing historical documents is crucial for cultural heritage and academic research, yet many suffer from degradation due to poor storage conditions. This poses significant challenges to their digitization and subsequent processing. Binarization is a fundamental step in document analysis but becomes particularly difficult when dealing with severely degraded documents. This paper presents a novel lightweight machine learning-based approach for binarizing historical document images, combining the strengths of K-means clustering and a multi-layer perceptron (MLP) for pixel classification. K-means clustering is exploited for representative pixel sampling, while PCA-based feature compaction enables efficient learning from only a small fraction of image pixels. The MLP classifier then predicts foreground and background classes, producing high-quality binarization outputs. Unlike traditional thresholding methods, which struggle with complex degradation patterns, the proposed framework remains robust to noise, bleed-through, and uneven illumination. Experimental results on multiple DIBCO benchmark datasets demonstrate that our approach outperforms classical methods and achieves competitive performance compared to recent deep learning-based models, while requiring significantly lower computational cost.</p>