Inkjet classification based on a few letters

Abstract

Morphology-based classification of inkjet documents has the characteristics of low cost and high efficiency, but this method usually requires measurement and analysis of a large number of printed characters. This paper proposes a novel method for detecting the source of printed documents using a few printed letters. A dataset containing data pertaining to various inkjet printers, including 27 models of inkjets from HP, Canon, and Epson, and their printed documents were gathered. The specifications of the various brands and models of inkjets are summarised, and the characteristics of the microscopic appearance of the printheads are presented. Principal component analysis (PCA) of the variables was applied to describe the proximity between the specimens, and a two-dimensional kernel density estimation was used to describe the variation between and within printer brands and models. Then, specific cases were simulated by random sampling based on the collected inkjet dataset. Multivariate kernel density estimation was used to estimate the numerator and denominator probability distribution for computing the likelihood ratio (LR). The result of K-nearest neighbour analysis showed classification accuracy as high as 98%. The evaluation of the LR presented a significant result (EER=0, RMEP=0, RMED=0.07). This method helps to find a specific inkjet from even a few letters in the printed document for tactical purposes.

Keywords: Bayesian approach; Image features; Inkjet printing; Printers; Questioned document examination; Statistical analysis.