Anotace:
The automated classification of historical document scripts holds profound implications for historians, providing unprecedented insights into the contexts of ancient manuscripts. This study introduces a robust deep learning system integrating an intelligent feature selection method, elevating the script classification process. Our methodology, applied to the CLaMM dataset, involves preprocessing steps such as advanced denoising through non-local means and binarization using Canny edge detection. These steps, pivotal for image cleaning and segmentation, set the stage for subsequent in-depth analysis. To enhance feature detection, we employ the Harris corner detector, followed by a k-means clustering process to eliminate redundancy and outliers. This process facilitates the extraction of consistently sized patches, capturing distinctive features of various scripts in historical manuscripts. The dataset undergoes rigorous training using precise convolutional neural network (CNN) models, empowering our system to discern intricate patterns and features for informed decision-making during the classification process. Ultimately, for the definitive script classification of an entire document, we employ a majority voting mechanism on the patches. The results highlight the effectiveness of this comprehensive approach, with the system achieving an impressive accuracy rate of 89.2%. This underscores the system proficiency in accurately classifying historical document scripts, offering a reliable and efficient solution for historians and researchers. The robustness of our methodology positions it as a compelling tool for meticulous analysis of historical manuscripts, contributing significantly to the field of historical document research and preservation.