Abstract
This paper proposes a texture analysis of the printed document based on Local Binary Pattern (LBP) descriptor for the application of printer identification. The LBP provides a statistical description of the pixels’ gray level differences within their neighborhoods. The occurrence histogram of local binary patterns is able to capture the document’s texture modifications by the distortion during the printing-and-scanning process, such as halftoning, geometric distortion, and mechanical defects. The most frequently appeared local binary patterns represent bright or dark flat regions. Furthermore, Gou et al. proposed an approach based on the combination of three different types of statistical features for scanner identification.We deconstruct their approach in order to evaluate the effectiveness of each type of features for printer identification. Our proposed LBP descriptor based model provides an excellent identification rate at approximately 99.4%, with a low variance. These results were achieved by Support Vector Machine (SVM) classification via n-fold cross validation and leave one out. They exceed any of the results obtained using the features, employed by the Gou et al. approach either singularly or in combination. Our experiments were conducted on 350 printed images, as well as 350 printed text documents, by a set of similar printers, two of which were exactly identical. The proposed model remains robust against common image processing, including averaging filtering, median filtering, sharpening, rotation, resizing, and JPEG compression.