Benchmarking Deep Learning Models for Telugu Document Image Classification

This study explores a hybrid machine learning and deep learning framework to classify digitized Telugu documents as either handwritten or machine-printed. The approach evaluates five models: three deep learning architectures AlexNet, ResNet50, and VGG16 and two traditional SVM classifiers using HOG and SURF feature extraction methods. The models are tested on a balanced dataset comprising 500 handwritten and 500 printed Telugu document images. Preprocessing steps such as median filtering and histogram equalization are employed to enhance image quality before feature extraction. Among all the models, VGG16 delivers the best performance, achieving a classification accuracy of 99.5%, surpassing ResNet50 (93.5%), AlexNet (92.5%), and SVM variants (90–92%). Additional evaluation using precision, recall, and F1-score confirms the robustness of VGG16. The results underscore the effectiveness of deep learning in script-specific document analysis and set a strong benchmark for future research in Telugu text classification.