Image Splicing Detection
We address the photomontage detection problem, with an assumption of simple cut-and-paste (splicing) without any post-processing such as matting or blending. Bicoherence, a third-order moment spectrum, is used for detecting the splicing discontinuity. A functional texture-decomposition method is used to improve the detection performance. We also provide an analytical model for image splicing and gives theoretical results for explaining why bicoherence can be used for detecting image splicing. The theory is verified by experiments.
Publications [bibtex]
"Blind Detection of Photomontage Using Higher Order Statistics" Tian-Tsong Ng, Shih-Fu Chang, Qibin Sun IEEE International Symposium on Circuits and Systems (ISCAS) May 2004. [Abstract][pdf][slides] "Blind Detection of Digital Photomontage using Higher Order Statistics" Tian-Tsong Ng, Shih-Fu Chang ADVENT Technical Report #201-2004-1 Columbia University June 2004. [Abstract][pdf]
Dataset [website]
Technical Illustration
Insufficiency in the Basic Bicoherence Features
The basic bicoherence magnitude and phase features achieve a low separability on the authentic and the spliced images, as shown in the left figure. Therefore, we incorporate the idea of authentic reference to improve the performance of the basic bicoherence features.
Splicing Detection on Idea of Authentic Reference
The idea is that if there exists an authentic reference to a spliced image, we can detect image splicing by looking at the difference between the spliced image and the authentic reference.
Texture Decomposition
We estimate the authentic reference by the structure component obtained from the functional texture decomposition proposed by Vese and Osher.
Bipolar Signal Model
We propose to explain the effectiveness of bicoherence in splicing detection using a bipolar signal model.
Columbia Image Splicing Detection Evaluation Dataset
We collected a dataset which consists of 933 authentic and 912 spliced grayscale image blocks of size 128x128 pixels. The image blocks are divided into five subcategories, as shown in the left figure.
Classification Performance
After incorporating new features derived from the authentic reference idea, the classification accuracy for the authentic and spliced images increases from 62% (for the basic bicoherence features) to 72%.