Printer Forensics Based on Identity Vectors of Image Texture Segmentation

Advances in the digital world are leading us to the development of digital forensic tools. The use of machine learning methods for source printer identification is one of the sub-fields of this area that is being developed. In this paper, a new method for extracting secondary features based on identity vector or i-vector to identify the print source is presented. In the proposed method, the classification process is accelerated only by extracting a low-dimension i-vector vector per page, without the use of optical character recognition (OCR) method, and by eliminating majority voting. Furthermore, the proposed method in extracting features is independent of the type and size of the font and the language of the text. Secondary features are obtained by splitting the document image into smaller patches and modeling the primary LBP features of the dark, border, and light areas in separate spaces. Modeling the primary features of different regions in separate total variability printer space makes it possible to extract class discriminator information from the remaining print texture in the bright area to increase classification accuracy. In this paper, the effect of using the texture of different regions and changing the patch dimensions using the SVM (Support Vector Machine) classifier through simulation has been carefully investigated. The simulation results show that only by refining the basic features of LBP we achieved 99.05% accuracy, which is more than the latest research in this field.