Image splicing detection approach based on low dimensional SVD features and kernel PCA / Zahra Moghaddasi

• Main Author: Zahra, Moghaddasi
• Format: Thesis
• Published: 2017
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://studentsrepo.um.edu.my/10078/1/Zahra_Moghaddasi.pdf; http://studentsrepo.um.edu.my/10078/2/Zahra_Moghaddasi_%E2%80%93_Thesis.pdf; http://studentsrepo.um.edu.my/10078/

Digital image forgery is becoming easier to perform because of the rapid development of various manipulation tools. Image splicing is one of the most common image forgery techniques. It is achieved simply by cutting a region from one or more images and pasting it, or them, into another image. This technique can cause inconsistencies in many features, such as an abnormally sharp transient at the splicing edges, and these inconsistencies are used to detect the forgery. To detect the spliced images several methods proposed utilizing the statistical features of the digital images. In this research, two efficient SVD-based feature extraction methods for image splicing detection are presented. In the first method, the natural Logarithm of inverse of each singular value is calculated. In the second method the concept of roughness measure is applied which is inversely proportional with condition number. Kernel Principal Component Analysis (PCA) is also applied as classifier feature selector to improve the classification process. And finally, support vector machine is used to distinguish between the authenticated and spliced images. The proposed methods are evaluated by applying three standard image datasets (DVMM v1, DVMM v2, and CASIA) in spatial and frequency domains. The first image dataset was the Columbia Image Splicing Detection Evaluation Dataset. This dataset contained 1845 gray-scale images (933 authentic images and 912 spliced images) in BMP format. The second image dataset is the Chinese Academy of Sciences, Institute of Automation (CASIA) with 1721 color images (800 authentic images and 921 spliced images). The third image dataset is DVMM v2, which contains 363 color images (183 authentic images and 180 spliced images). For the DVMM v1 image dataset, proposed method-1 shows an average accuracy of 98.78%. On the other hand, for CASIA image dataset, method-2 shows an average accuracy of 99.62%. Finally, with the DVMM v2 image dataset, both methods obtain an average accuracy of 100%, but in different color channels. These results outperform several current detection methods.