Copy-move forgery detection (CMFD) has been widely researched in the last few years. Approximately 20 feature sets and processing variants have been proposed. While some features sets are very robust towards rotation and scaling, the filtering of the found matches typically scales poorly under some parameters, for instance increasing image size or highly similar image content.

In our paper "On Rotational Invariance in Copy-Move Forgery Detection", presented at WIFS 2010, we addressed the filtering of matched block pairs. We designed explicitly a natural extension of the popular "same shift vector" filtering approach (proposed as early as [1, 2]). Our method, called "Same Affine Transform Selection" (SATS), preserves the properties of the same shift vector filtering under affine geometric transformations.

Thus, it is possible to detect rotated copy-move forgeries (and in theory also scaled copy-moved regions, although due to space limitations not tested in this paper). At the same time, this approach is very robust towards outliers. We tested it on large images, where a large number of false positive matches in the feature space are expected. However, SATS removes almost all of these false positive matches.

The code for SATS is integrated in our full CMFD framework (which you can find here). The implementation of SATS is in the class 'FastSats'.

If you want to use the code, please cite one of our CMFD papers, either "On Rotation Invariance in Copy-Move Forgery Detection"; Vincent Christlein, Christian Riess, Elli Angelopoulou. In: Workshop on Information Forensics and Security (WIFS) 2010, Dec. 2010, Seattle, USA or our paper on comparing different CMFD methods.

Note that the presented code is research quality, we take no liability for it. However, we are glad if you have suggestions for improvements.