Source Printer Classification Using Printer Specific Local Texture Descriptor

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 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 1 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.

VIPPrint: Validating Synthetic Image Detection and Source Linking Methods on a Large Scale Dataset of Printed Documents

“…In the first set, which we call image texture descriptors, we used a set of common descriptors that are mainly used for image characterization. For a source attribution task, such descriptors can be useful to differentiate printers’ banding artifacts efficiently if the analyzed patterns do not change much, and therefore, they normally exhibit good performance in some printer source attribution tasks [ 18 , 20 , 22 , 34 ]. We considered four approaches in this set as follows.…”

“…The final classification result for an image was set to be the mode of the classifications obtained on single patches. This approach is commonly known as majority voting and was also validated in printed document forensics research [ 18 , 19 , 20 , 22 , 34 ]. To choose the patches, we applied the highest-energy procedure already described in Section 3 .…”

“…After the letters are classified, the source of a document can be found by verifying the most voted class among all the individual letters “I” classification. Several other techniques used a similar pipeline with few modifications, such as considering the statistics of gray-level co-occurrence matrices [ 4 , 5 , 6 , 11 ], Distance Transform [ 8 ], Discrete Cosine Transform [ 10 ], statistics of gray-level co-occurrence matrices together with residual noise and sub-bands of wavelet transform [ 12 , 13 , 14 , 17 ], deep neural networks [ 18 , 20 ], ad hoc texture descriptors [ 19 , 22 ] among others [ 7 , 9 , 15 , 16 , 21 ].…”

“…Solutions for printer attribution are mostly based on the analysis of the extrinsic artifacts contained in printed documents, with the most popular ones for laser printers being banding, jitter, and skewed jitters. The presence of these artifacts has been exploited by several works to identify the sources of printed texts [ 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 ], color images [ 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 ], or both [ 34 , 35 , 36 ]. Manipulation detection in printed documents has received some attention only recently [ 37 ] and mainly refers to unveiling post-processing operations that could alter the semantic meanings of the images.…”

Ensembling Shallow Siamese Neural Network Architectures for Printed Documents Verification in Data-Scarcity Scenarios

“…As an answer to the above needs, several works have focused on present Printer Source Linking solutions using Computer Vision and Machine Learning to pinpoint the ownership of printed texts [4]- [23], color images [24]- [34] or both [35]- [37]. In particular, the approaches based on Convolutional Neural Networks (CNNs) and Deep Learning (DL), in general, [19], [21], [37] have allowed significant progress in this research area, in special due to their ability of learning from the data itself, artefacts specific to given printers.…”