Seite - 167 - in Joint Austrian Computer Vision and Robotics Workshop 2020

points, with size 3×3 to 7×7 pixels (denoted as GS+LS). Another strategy to remove SIFT keypoints is the collage attack (CA) [10], which substitutes an orig- inal image patch (patch containing a keypoint) with another patch (containing no keypoint) of the same size contained in a pre-computed patch dictionary. The new patch must not contain SIFT keypoints and should be as similar as possible to the original one according to some similarity criteria (e.g., [1] cre- atedadictionaryofabout120,000patchesandchose histogram intersection distance, widely used in im- age retrieval applications [22], as a patch similarity measure. The same approach is used in experiments of [3]. Removal with minimum distortion (RMD) [9] adaptively calculates a small image patch and adds it to theneighbourhoodof thekeypoint such that the overall operation results in a minimum least-square distortion in the keypoint neighbourhood under the condition that the keypoint is removed. Finally, the classification based attack (CLBA) presented by [1] arrangesGS+LS,CA,andRMDintoaniterativepro- cedure which first detects SIFT keypoints, classifies them into distinct classes, and subsequently applies one of the three individual removal techniques to the suited classes. For all these techniques, [2] suggested to remove onlyoneofthematchingkeypointsfromeachmatch- ing keypoints pair in case of preventing to detect copy move forgeries. There are also forensic tech- niques to counter those anti-forensic keypoint re- movalmethods (seee.g. [7,16]. AsGShassignificant impacton imagequality (as we shall see as well in the next section), also the combinationwithLS(i.e. GS+LS) isaffectedbythis quality impact. Therefore, we introduce a new tech- nique toremoveSIFTkeypointscalled localsmooth- ing (LS), andcompare thevariousperformance indi- cators toalreadyexisting(smoothing) techniques i.e. GS,GS+LS,andCA. 3.Experiments 3.1.ExperimentalSettings Withrespecttosoftwareandtools,wemainlyused Matlab 2014a [14] (on Windows 7 64bit) with some internal toolboxes(parallel toolboxfor fastcomputa- tion, image processing toolbox) and the external li- brary vl_feat [20] (the latter to smooth images and to compute SIFT keypoints; we have chosen Edge Thresh = 12 to control the number of keypoints used). For the computation of image quality metrics (IQM) (PSNR, VSNR, UQI, SSIM), we used the Ma- Trix MuX visual quality assessment package [15]. As experimental data, we used the first 100 images (i.e. from ucid00001.tif to ucid000100.tif) from the Uncompressed Image Database (UCID) [18] for ex- periments forkeypoint removalmethodsassessment. For the CA, we created a keypoint-free patch dictio- nary fromall imagesusingoverlappingpatches. For experiments with respect to detecting actual copy move attacks, we combined two datasets to re- sult in 100 images (50 actually forged images and 50 original images). Forged images are taken from a public dataset for assessing forensic techniques [5] (see Fig. 5 for examples), which contain sim- ple translated copies of objects/regions, while the “original” images are taken from the RAISE dataset [8] from theBUILDING PHOTO category (see Fig. 6). The latter data has been included to determine themethods’ robustnessagainst indicating falsepos- itives1. In keypoint removal for countering copy movedetection,weremovedonlyonekeypoint from eachmatchingpairofkeypoints as suggested. 3.2.ExperimentalResults In order to assess the quality of the image af- ter removing keypoints, we used different IQM, i.e. PSNR, SSIM, VSNR, and UQI. Fig. 1 com- pares three different techniques, i.e. GS, GS+LS, and LS. In GS+LS, an image is smoothed first glob- ally withσ= 1.3 as suggested in literature and af- terwards patches containing keypoints (of different sizes) are smoothed locally. In the plots, different smoothing strength (different σ values) is depicted on theXaxis, while theYaxis represents the output value for a specific imagequalitymeasure. Fig. 1 reveals that the quality of a locally smoothed (LS) image is better in comparison to the othertwosmoothingtechniques(i.e.GSandGS+LS) for all IQM. GS deteriorates image quality quickly for increasingsmoothingstrength. Also for thecom- binedmethodGS+LSthequalityisfoundtoberather low due to the impact of GS. The quality of the LS images is better because we are smoothing only the patchesaroundSIFTkeypointswhileotherpixelsare left untouched. As expected, when increasing the patch size in LS and GS+LS, the quality of the pro- 1Similar looking structures within an image may lead to an image incorrectly being classified as copy move forged image. 167