Please use this identifier to cite or link to this item: http://hdl.handle.net/10397/85974
DC FieldValueLanguage
dc.contributorDepartment of Electronic and Information Engineering-
dc.creatorChan, Lit Hung-
dc.identifier.urihttps://theses.lib.polyu.edu.hk/handle/200/7005-
dc.language.isoEnglish-
dc.titleDigital camera identification for forensics applications-
dc.typeThesis-
dcterms.abstractWith the large number of digital imaging devices nowadays, the use of digital photos in court keeps on increasing. One might want to authenticate the origin of the photo (i.e., whether it is downloaded or produced from a certain camera). Source camera identification thus becomes important in digital forensic applications. Camera identification can generally be classified into two types, namely source model identification and individual source camera identification. Both of them try to extract device signature and check whether that signature can be found in a given photo. But the former can only determine the brand and the model of the camera while the latter can uniquely identify each individual camera. The focus of this thesis is on the individual source camera identification. Existing source camera identification methods use a type of pattern noise called the photo response non-uniformities (PRNU) noise. It is caused by manufacturing imperfections and is presented in every image taken by the device. The PRNU is extracted through image denoising. In particular, it is obtained as the difference between the original image and its denoised version. One major problem is that the PRNU can be affected by image content. For example, the PRNU is completely absent in saturated area. Previous studies have found that the image content can seriously affect the identification accuracy. The objective of this study is to investigate ways to compensate for the scene content effect in PRNU estimation. A possible solution to deal with the image content is to use the correlation predictor. It tries to quantify the seriousness of the scene content effect on the pattern noise. We have performed a detailed study on the relation between the predicted correlation and the identification accuracy. Using 2D Gaussian modeling, the identification accuracy can be obtained for different image content as characterized by the predicted correlation. Using this result, a 2D classifier is proposed for individual source camera identification. The 2D classifier uses the predicted correlation as one of the features to quantify the scene content effect. It helps setting different correlation thresholds for different types of image content. Experimental result shows the identification accuracy increases by about 4% as compared to the traditional identification methods. The correlation predictor is able to characterize the image content effect in a block-based manner only. We extended the characterization to a pixel level. In particular, a non-linear regression model is used to formulate the scene content effect. Then, a confidence map is generated which indicates the reliability of each pixel in the PRNU estimation. By using the confidence map as a weighting function in correlation calculation, the scene content effect can be compensated. Experimental results show that the proposed confidence map is able to achieve an accurate camera identification result. As compared with state-of-the-art identification methods, our proposed method can achieve about 2%-5% and 4%-20% improvement in detection accuracy at JPEG quality factor of 90 and 70 respectively.-
dcterms.accessRightsopen access-
dcterms.educationLevelM.Phil.-
dcterms.extentxiv, 118 leaves : ill. (some col.) ; 30 cm.-
dcterms.issued2013-
dcterms.LCSHImage processing -- Digital techniques.-
dcterms.LCSHComputer crimes -- Investigation.-
dcterms.LCSHHong Kong Polytechnic University -- Dissertations-
Appears in Collections:Thesis
Show simple item record

Page views

44
Last Week
0
Last month
Citations as of Mar 24, 2024

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.