A Novel Deep Ensemble Framework for Online Signature Verification Using Temporal and Spatial Representation

Abstract

Although considerable improvements have been made in online signature verification (OSV) over the last decade, none of them take both temporal and spatial information into consideration, and thus there is still a room for boosting the performance. In this paper, we propose a novel ensemble based deep learning framework, which consists of a convolutional neural network model and our recently designed convolutional gated recurrent network (CGRN) for extracting spatial feature and temporal feature, respectively. However, it is not easy to combine these two types of features since temporal feature is two-dimensional with various length while the other is a fixed-length vector. In order to incorporate both types of representation, we firstly introduce cosine similarity for spatial feature to calculate the shape similarity and use dynamic time warping (DTW) for temporal feature alignment. Thereafter, the distance between reference signature and given signature is obtained by multiplying DTW distance and similarity score. In addition, we design a novel approach for DTW distance normalization, which significantly enhances the verification accuracy. Our method achieves new state-of-the-art result on DeepSignDB, and outperforms other existing OSV methods with at least 16.2% relative improvement in finger scenario.