Open writer identification from handwritten text fragments using lite convolutional neural network

Usually, a writer identification system based on the convolutional neural network (CNN) is designed as a closed system, which is composed of many convolutional layers trained often on the entire document for achieving a high performance but requiring a high computation cost. This paper proposes an open writer identification system using a lite CNN composed of only four convolutional layers for extracting features from text fragments. The CNN is trained on a small subset of writers, and then, the resulting model is used for feature generation for new writers, without retraining, associated with the distance-based classifier. The proposed system is simple and easy to deploy for adding new writers without retraining. Extensive experiments performed on text fragments produced from the standard IFN/ENIT and IAM datasets show an encouraging performance against the state of the art of closed systems with an identification rate of 97.08% and 91.00%, respectively, despite few fragments used for writer identification.