Feature Selection using F-score Method for Offline Arabic Handwritten Fragment Identification

Abstract

Among the various means of document authentication, handwriting fragments represent one of the challenging tasks for automatic handwritten writer identification. Usually, the writer is represented by a set of features extracted for each handwritten fragment. However, feature components might be irrelevant due to their redundancy, affecting the pertinence of the feature vector. Hence, this paper proposes a feature-selection strategy based on a hybrid F-score method, performed genuinely for multi-class classification. The F-score is used conjointly to a classifier based-distance. According to a statistical analysis performed on F-score distribution during the training step, the number of pertinent feature components is deduced from the highest F-score value. The experimental evaluation performed on the well-known IFN/ENIT handwritten fragment dataset shows an improvement of the identification rate of 94.20% while reducing the size of the feature vector of 54%.