A performance comparison of over-complete dictionaries for online writer identification

Biometrics is the field that explores ways to assess an individual’s unique physical and behavioral characteristics and utilize them for identification or authorization purposes. Handwriting is a significant behavioral biometric that is commonly used. The primary objective of writer identification systems is to determine the authorship of a handwritten document by comparing it with a database of existing handwritten samples. In this paper, we aim to evaluate and compare the performance of different over-complete dictionaries for the task of online writer identification. Specifically, we focus on generating writer descriptors using various over-complete dictionaries and analyze the identification rates achieved with these descriptors. Additionally, we investigate the computation time required by different dictionary learning algorithms within the writer identification system framework. To conduct our experiments, we utilize the publicly available IAM Online handwriting database. We have focussed on four different dictionary learning algorithms for the over complete dictionary generation viz. KSVD, MOD, HAAR exemplar and SPAMS. The results obtained from our study reveal that the choice of dictionary learning algorithms significantly influences the performance of online writer identification systems.