On-line Handwritten Signature Verification using Machine Learning Techniques with a Deep Learning Approach

Abstract: The problem to be solved in this project is to distinguish two signatures from each other, with help of machine learning techniques. The main technique used is the comparison between two signatures and classifying if they are written by the same person (match) or not (no-match). The binary classication problem is then tackled with a few alternatives to better understand it. First by a simple engineered feature, then by the machine learning techniques as logistic regression, multi-layer perceptron and nally a deep learning approach with a convolutional neural network. The evaluation method for the dierent algorithms was a plot of true positive rate (sensitivity) versus false positive rate (fall-out). The results of the alternative algorithms gave a dierent understanding of the problem. The engineered feature performed unexpectedly well. The logistic regression and multi-layer perceptron performed similarly. The main results from the nal model, which was a max-pooling, convolutional neural network, were a true positive rate of 96.7 % and a false positive rate of 0.6 %. The deep learning approach on the signature verication problem shows promising results but there is still room for improvement.