Offline Writer Identification Using Textural Descriptors

Abstract

This thesis presents a novel approach for offline writer identification by applying textural descriptors to extract the unique writing style of individuals from static samples of handwriting. Offline writer identification is a challenging biometric problem due to the variability in handwriting caused by emotional, physical, and environmental factors, as well as missing dynamic writing details such as stroke order or pressure, to address these challenges, this manuscript investigates Run-Length-based feature extraction techniques, with a focus on Run-Length statistics. Our study begins by investigating traditional Run-Length (RL) descriptors, which are used to describe the distribution and repetition of binary patterns or runs in handwriting images. We enhance their discriminative ability with a new descriptor Z-Run-Length (Z-RL) that traverses images along a zigzag course to detect more subtle structural differences in handwriting. Furthermore, we evaluate a hybrid approach that combines baseline Run-Length features with the proposed Z-RL descriptor. The resulting feature vectors are classified using machine learning algorithms such as Support Vector Machines (SVM) and K-Nearest Neighbors (K-NN),

and are validated using robust evaluation protocols, including k-fold cross-validation and Leave- One-Out Cross-Validation (LOOCV).

Experimental results on the BFL dataset demonstrate that the proposed system achieves high accuracy and exhibits strong consistency across diverse handwriting samples.