Gaussian Process Classification as Metric Learning for Forensic Writer Identification

Abstract

In this paper, a statistical machine learning approach for constructing a metric separating unseen writer hands, is proposed. An unsupervised feature learning approach, based on dense contour descriptor sampling, was combined with a novel way of learning a general space for clustering writer hands, in a forensic setting. The metric learning inference was based on multiclass Gaussian process classification. Using the popular datasets IAM and CVL combined, the evaluation was performed on close to 1000 writer hands. This paper builds on earlier work from our group on building a system for estimating the production dates of medieval manuscripts, and act as a foundation for future use of writer identification techniques on our historical data.