Reduction of feature statistics estimation error for small training sample size in off-line signature verification

Abstract

In this paper, we propose two methods to tackle the sparse data problem in off-line signature verification. The first one is to artificially generate additional training samples from the existing training set by an elastic matching technique. Feature statistics are estimated by using the expanded training set. The second approach applies regularization technique to the sample covariance matrix to overcome the problem of inverting an ill-conditioned covariance matrix and obtains stabilized estimation of feature statistics. Experimental results showed that both techniques are able to produce significantly improved verification accuracy when implemented with a set of peripheral features. © Springer-Verlag Berlin Heidelberg 2004.