Article Preview
Top1. Introduction
Digital images are nowadays used in the majority of the application fields in place of “old” analog images because of their easiness of usage, quality and above all manageability. These favourable issues bring anyway an intrinsic disadvantage: digital content can be simply manipulated by ordinary users for disparate purposes so that origin and authenticity of the digital content we are looking at is often very difficult to be assessed with a sufficient degree of certainty. Scientific instruments which allow to give answers to basic questions regarding image origin and image authenticity are needed (Chen, 2008). Both these issues are anyway connected and sometimes are investigated together. In particular, by focusing on assessing image origin, two are the main aspects to be studied: the first one is to understand which kind of device has generated that digital image (e.g., a scanner, a digital camera or it is computer-generated) (Lyu, 2005; Khanna, 2008) and the second one is to succeed in determining which kind of sensor has acquired that content (i.e., the specific camera or scanner, recognizing model and brand) (Chen, 2008; Khanna, 2007; Gou, 2007). The main idea behind this kind of researches is that each sensor leaves a sort of unique fingerprint on the digital content it acquires due to some intrinsic imperfections and/or due to the specific acquisition process. Various solutions have been proposed in literature among these the use of CFA (Color Filter Array) characteristics (Swaminathan, 2008) is quite well-know, nevertheless two seem to be the main followed approaches. The first one is based on the extraction, from images belonging to different categories (e.g., scanned images, photos, etc.), of some robust features which can be used to train a SVM (Support Vector Machine). When training is performed and whether features grant a good characterization, the system is able to classify the digital asset it is asked to check. The second approach is based on the computation of fingerprints of the different sensors (this is particularly used in sensor identification) through the analysis of a certain number of digital contents acquired by a device (e.g., images scanned by a particular scanner, photos taken by a camera and so on). Usually fingerprints are computed by means of the extraction of PRNU noise (Photo Response Non-Uniformity) (Chen, 2008; Mondaini, 2007) through a digital filtering operation; PRNU presence is induced by intrinsic disconformities in the manufacturing process of silicon CCD/CMOSs. After that the PRNU of the to-be-checked content is compared with the fingerprints and then it is classified. In this paper a new technique to distinguish which kind of device, a digital scanner or a digital camera, has acquired a specific image is proposed. Because of the structure of CCD set, the (PRNU) noise pattern, left over a digital image, will have a completely different distribution: in the scanner case it should show a mono-dimensional structure repeated row after row in the scanning direction, on the other hand, in the camera case, the noise pattern should present a bi-dimensional template. On the basis of this consideration we construct a 1-D signal and by resorting to a DFT analysis, which exploits the possible existence of a periodicity, understanding which has been the acquisition device. The paper lay-out is the following: Section 2 introduces a characterization of the sensor pattern noise and the periodicity is discussed, in Section 3 the proposed methodology is presented and Section 4 describes thresholds selection based on ROC curves. In Section 5 some experimental results are brought to support theoretical theses and conclusions are drawn in Section 6.