Locating splicing forgery by adaptive-SVD noise estimation and vicinity noise descriptor

Abstract

Splicing forgery in digital images is a common form of photographic manipulation which composites two or more images into a single picture. Detection of splicing forgery remains a challenging task in image forensics. Base on the fact that images from different origins should have different amount of noise produced by image sensors or post processing by software, in this paper we proposed a novel detection method by analyzing noise discrepancy to expose and locate splicing forgery in digital images. To improve the accuracy of noise estimation, we proposed the Adaptive Singular Value Decomposition (Adaptive-SVD) to estimate the local noise. By combining local and global noise clues, we proposed the Vicinity Noise Descriptor to locate splicing forgery. Regional forensic information is inferred via machine learning method - Support Vector Machine (SVM). To evaluate the proposed method, we constructed splicing forgery databases which include various scenarios and different spliced objects with artificially added or camera-generated noise. Experimental results show that our proposed method is able to locate multi-objects spliced from different origins and comparing to state-of-the-art noise difference based methods, our method improves detection accuracy especially the value of precision, which significantly affects observers’ judgement.