Enhancements on a signature recognition problem

This paper presents an enhanced method of partitioning a dataset into clusters when dealing with the handwritten signature recognition problem. The goal of the present system is improving the performance of two previously developed systems. In the first version of our system we dealt with data extraction from signature images and obtained a recognition rate of 91.04% using the Naïve Bayes classifier and the feature selection method. In the second version of our system we performed a hierarchical partitioning of the dataset into clusters in order to obtain a faster and better classification and then we applied the Naïve Bayes classifier in order to determine the recognition rate. The best results were reported when partitioning the data into 7 and 8 clusters and the mean accuracy obtained was 91%. In the current version of our system, we identified the most appropriate model for each cluster by selecting the best performance obtained after applying 7 different classifiers on the clusters. In order to improve the recognition rate and reduce the time required to build a model, we applied the feature selection method on each of the clusters with the previously determined classifiers. We obtained an increased accuracy with 1.62%on a dataset with 14 instances/class and an increment of 3.24% on a dataset with 20 instances/class.