Source Printer Classification using Printer Specific Local Texture Descriptor

TL;DR

This paper introduces a novel local texture descriptor for printer source classification that overcomes font dependency issues, achieving high accuracy across different fonts and outperforming existing methods in multimedia forensics.

Contribution

The work presents a new printer-specific local texture descriptor that effectively classifies printers regardless of font variations, addressing a key limitation of prior approaches.

Findings

Outperforms state-of-the-art algorithms on same-font datasets.

Achieves 100% accuracy in classifying multiple fonts with sufficient training data.

Improves cross-font classification accuracy and reduces confusion between similar printers.

Abstract

The knowledge of source printer can help in printed text document authentication, copyright ownership, and provide important clues about the author of a fraudulent document along with his/her potential means and motives. Development of automated systems for classifying printed documents based on their source printer, using image processing techniques, is gaining a lot of attention in multimedia forensics. Currently, state-of-the-art systems require that the font of letters present in test documents of unknown origin must be available in those used for training the classifier. In this work, we attempt to take the first step towards overcoming this limitation. Specifically, we introduce a novel printer specific local texture descriptor. The highlight of our technique is the use of encoding and regrouping strategy based on small linear-shaped structures composed of pixels having similar intensity and gradient. The results of experiments performed on two separate datasets show that: 1) on a publicly available dataset, the proposed method outperforms state-of-the-art algorithms for characters printed in the same font, and 2) on another dataset\footnote{Code and dataset will be made publicly available with published version of this paper.} having documents printed in four different fonts, the proposed method correctly classifies all test samples when sufficient training data is available in same font setup. In addition, it outperforms state-of-the-art methods for cross font experiments. Moreover, it reduces the confusion between the printers of same brand and model.