Enhanced Script Identification for Multilingual Document Analysis: A Dataset and Machine Learning Approach for Kashmiri and English

Objectives: The main purpose of this work is to develop an automatic script identification system for multilingual documents by proposing a benchmark dataset of printed Kashmiri (Perso-Arabic) and English scripts and systematically evaluating HOG-based machine-learning models. The study aims to support multilingual document analysis and enhance the pre-processing step of Optical Character Recognition (OCR) systems. Method: This study has created a dataset of 4,423 printed text images from different publicly available academic sources, pre-processed the images by resizing and binarization, and extracted Histogram of Oriented Gradients (HOG) descriptors to represent script characteristics. Four machine learning classifiers – Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Naïve Bayes (NB), and Random Forest (RF) were trained, and different evaluation metrics were used to evaluate their performance. Findings: The proposed HOG-SVM framework achieved an overall classification accuracy of 99.0%, demonstrating superior performance compared with KNN, Naïve Bayes, and Random Forest classifiers, and also produced the fewest misclassifications, with 8 English-to-Kashmiri errors and only 4 Kashmiri-to-English errors, demonstrating its effectiveness for Kashmiri–English script identification. Novelty: This work introduces a printed Kashmiri-English script benchmark dataset and a detailed baseline assessment of various machine learning classifiers in a low-resource scenario for script identification. The proposed optimized HOG-SVM-based framework is a reliable benchmark for further studies on Kashmiri script recognition and can serve as a foundation for developing a multilingual OCR system for low-resource languages. Keywords: Kashmiri Language, HOG-ML, Perso-Arabic Script, Printed Text Dataset, Multilingual Document Analysis, Script Identification