A new texture descriptor for handwritten document writer identification

Abstract

Writer identification is a critical task in the realm of pattern classification, aimed at determining the authorship of a manuscript based on labeled handwriting sam- ples. This area has garnered considerable attention from researchers and has seen significant advancements in the last two decades, propelled by the inte- gration of novel computer vision and machine learning algorithms. Commonly, approaches within this field rely on calculating local texture descriptors of im- ages. In this work, we propose a novel local texture descriptor method, termed multi-points local binary patterns (MP-LBP), which is an enhancement of the traditional local binary patterns (LBP) descriptor. Our approach involves apply- ing the MP-LBP descriptor to patches surrounding Harris key points and aggre- gating the image descriptors into encoded vectors using the vector of locally ag- gregated descriptors (VLAD) encoding method. These vectors are subsequently classified by a ball tree classifier to associate the document with the most plau- sible writer. To assess the efficacy of our descriptor, we conducted evaluations on five publicly accessible handwritten databases: CVL, CERUG-EN, CERUG- CH, BFL, and IAM. The results of these tests provide insights into the perfor- mance of the MP-LBP descriptor in the context of writer identification.