SIFT Based Signature Classifier Prototype

Abstract

The need to distinguish the genuine signatures from forged signatures is important to identify the authenticity of the person and to verify the signature. In this research project a prototype is developed for automatically recognizing and classifying the student?s signatures in a Campus. A methodology is developed to crop the signature samples automatically from scanned student?s attendance sheets and to store it in the relevant class folder. Here, two sets of experiments were carried out, one is using Scale Invariant Feature Transform (SIFT) with k-means clustering algorithm and the other one is using Speeded Up Robust Features (SURF) with k-means clustering algorithm. In SIFT and SURF experiments the feature descriptors also known as the interesting points were extracted from the signature images. The k-means algorithm is used for the construction of visual codebook for the feature vectors, where the codebook is the set of centers of the learnt clusters. From the codebook the training and testing histograms were formed by assigning each feature of the training and testing signature images to the closest code word. From the initial histogram matching results it is found that SURF based experiments gives very promising results than SIFT based experiments. The optimum testing and training histograms obtained from SURF experiment is then fed to Support Vector Machine (SVM) for classification and prediction of the class category of the given signature image. The performance of the classifier is evaluated from the classification rate. The accuracy of the classifier is tested by mixing the forgery samples with the genuine testing samples by replacing the respective genuine samples in some known positions. The accuracy of the classifier is then verified by Human expert by evaluating the performance measurements such as True Positive Rate (TPR), False Positive Rate (FPR). From the SURF algorithm based experiments good cross validation accuracy (97 %) and classification accuracy (89.375 %) were obtained. After interchanging the training and testing samples these results were 92.25 % and 96.875 % respectively, this variation is discussed in the conclusion and the future work of this research project.