Leveraging Deep Learning for Writer Identification in Handwritten Historic Arabic Documents

Project Details

• Student(s): Zeina Ayoubi
• Advisor(s): Dr. Joe Tekli
• Department: Electrical & Computer
• Academic Year(s): 2024-2025

Abstract

Writer identification is a fundamental task in digital document analysis. Traditional approaches rely on local descriptors such as SIFT, followed by clustering, CNN-based feature learning, and VLAD encoding to construct global document representations. While effective, these methods often struggle to capture subtle handwriting variations across writers. This paper introduces an enhanced writer identification solution that combines SIFT keypoint detection with ResNeSt, a Convolutional Neural Network (CNN) incorporating split-attention mechanisms for improved feature discrimination. Local descriptors extracted from image patches are further aggregated using RVLAD, a robust extension of VLAD that normalizes residuals and reduces sensitivity to outliers. Experiments conducted on two benchmark datasets — the Balamand Arabic historical dataset, and the ICDAR 2019 dataset —demonstrate the improved performance of the proposed method.