Effective Writer Identification from Text Fragments Using Lightweight Convolutional Autoencoder

Writer identification using handwritten text fragments is a challenging task due to the variability of handwriting styles and the limited amount of data available for training. This paper proposes a new open writer identification system based on a lightweight convolutional autoencoder (CAE) framework. The proposed system leverages the CAE’s ability to extract discriminative features from handwritten text fragments. Notably, the system is trained on a small subset of writers, making it an open system. A new metric, based on the number of correctly classified fragments, is introduced during the design step to enhance the robustness of extracted features, which are then fed to the distance-based classifier (DBC) for writer identification. Importantly, integrating new writers into the system does not require retraining the lightweight CAE. Experimental results demonstrate that the proposed system achieves competitive identification rates of 95.86% and 89.78% on both the IFN/ENIT and IAM datasets, respectively.