Writer identification using texture features: A comparative study

摘要

A texture-based approach for writer identification of multiple scripts on a single platform is presented in this paper. Potential texture descriptors, namely Local Binary Pattern (LBP), Local Phase Quantization (LPQ), Discrete Wavelet Transform-based Local Extrema Pattern (DWT+LEP), Discrete Wavelet Transform-based Directional and Local Extrema Pattern (DWT+DLEP), Center Symmetric Local Binary Co-occurrence Pattern (CSLBCoP), and Local Tri-Directional Pattern (LTriDP) have been analyzed for identifying the writers. Comparative study for Latin, Arabic, and Devnagri databases was performed, with the Devnagri database contributed by us. The study shows high writer identification rates of 97.62% for IAM dataset using LBP features and Support Vector Machine (SVM) classifier, 95.60% for KHATT database using k-Nearest Neighbor (kNN), and 65.80% for Devnagri scripts using LPQ features and kNN classifier.