Boosting Feature Based Classifiers for Writer Identification

The identification of a writer of a handwriting image is very useful for applications in forensic and historic document analysis. Writer identification methods retrieve the closest image within a list of samples of different writers to a query sample. In automatic writer verification the system takes an automatic decision if two handwriting images were written by the same person. In recent years, several effective and powerful features were designed to capture and characterize writer individuality and been used in automatic writer identification and verification. A wide variety of classifiers were presented to work with such features presenting impressive results. Mostly, these classifiers assumed that all errors have the same cost and based on specific features set. In this paper, we analyze and improve some of these features and combine them by using boosting methodology which is error cost sensitive to instigate better classifiers. Results on the ICDAR2015 competition data set with KHATTT and ICDAR2011 competition databases, prove that the presented approach improves the accuracy.