Writer identification and gender classification framework for forensic handwriting analysis in Devanagari script

Abstract

Over the recent years, the field of forensic handwriting analysis has undergone significant advancements facilitated by the application of machine learning models leading to noteworthy progress in writer identification and gender classification systems. The framework for writer identification and gender classification using handwritten Devanagari scripts outlines the complete workflow, covering data collection, pre-processing, and the application of machine learning models: SVM, K-NN, decision tree, random forest, XGBoost, AdaBoost, 5-layer CNN, VGG16, ResNet34, and ResNet50. In the experiments, the SVM model demonstrates performance on noised data with an accuracy of 54.83%, while the XGBoost model excels on denoised data, achieving an accuracy of 61.44% for writer identification. For gender classification, the random forest emerges as the most effective, attaining the highest accuracy of 76.85% with noised data, slightly improving to 79.08% with denoised data. In the evaluation deep learning models: 5-layer CNN, VGG16, ResNet34, and ResNet50. The ResNet34 model emerges as the most effective, achieving an accuracy of 93.52% in writer identification and 91.98% in gender classification on denoised datasets. Notably, this model maintains high accuracy, reaching 92.36% even with a noised dataset. It outperforms other models in gender classification, achieving an impressive accuracy of 91.98% on a dataset that has undergone denoising. The performance of two hybrid classification methods is evaluated. The accuracy achieved with parallel hybrid classification model (PHCM) is moderately satisfactory at 66.84%, the incorporation of deep learning models leads to a significant improvement, reaching an impressive accuracy of up to 93.78% in writer identification. For gender classification, PHCM achieves substantial accuracy levels, peaking at 83.53% and 92.67% with deep learning models.