<p>Historical document images are often degraded by various types of noise, which hinders their readability and preservation. Accurately classifying these degradations is crucial, as each noise type requires specific preprocessing for effective restoration. The objective of this work is to develop an automated and reliable classification system capable of identifying six degradation categories: paper damage, show-through, spots, faint text, low contrast, and clean images. Our methodology combines deep learning with ensemble learning. Four convolutional neural networks—VGG16, MobileNetV2, InceptionV3, and ShuffleNetV2—were individually fine-tuned on a curated dataset. Their predictions were then integrated using a stacking ensemble approach, with Logistic Regression selected as the meta-classifier after extensive experimentation. To further enhance robustness, handcrafted global and local features were incorporated into the ensemble, complementing the information captured by deep models. The results show that while each model had distinct strengths, the stacking framework significantly outperformed them individually. Logistic Regression achieved the best performance across all metrics, including accuracy, precision, recall, F1-score, specificity, and ROC-AUC. The addition of handcrafted features led to further improvements, particularly for difficult degradation categories. Comparative analysis with a recent state-of-the-art method demonstrated that our model generalizes better, outperforming it in most noise scenarios despite not being trained on the same degradation types. These findings highlight the potential of stacking-based classification to support more targeted and effective document restoration workflows, thereby contributing to the preservation and accessibility of cultural and historical heritage.</p>