Page - 168 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Figure 1: IQM Comparison among GS vs LS vs GS+LS cessed imagesdecreases. Table 1displays IQM values for theCA. Patch Sizes PSNR VSNR UQI SSIM 3x3 Patch 64.80 32.78 0.99 0.99 5x5 Patch 51.95 32.23 0.99 0.99 7x7 Patch 47.10 47.58 0.99 0.99 9x9 Patch 42.86 40.89 0.98 0.99 Table1: IQM forCA. For UQI as well as SSIM we notice almost no qualitydegradationbytheCA,nomatterwhichpatch size is being used. For PSNR, CA is superior to all GS+LS variants and for almost all other settings ex- cept for extremely low smoothing strength. Finally, for VSNR, CA is again superior to all GS+LS vari- ants and for all other techniques but LS with patch- size 3 for low smoothing strength. Overall, the qual- ityobtainedwith theCAisverygood,andonlycom- parable to LS with patchsize 3, however, with all patchsizesconsidered. But how effective are the smoothing-based meth- ods in actually removing keypoints ? Contrasting to CA, in which all present keypoints are replaced by keypoint-freepatches, smoothingdoesnotguarantee that keypoints are actually removed. Fig. 2 illus- trates the percentage of original keypoints which are still present after smoothing for increasing smooth- ingstrength. 100 80 60 40 20 0 0 1 2 3 4 5 Strength of Smoothing GS LS(3x3) GS+LS(3x3) (a)PatchSize3×3 100 80 60 40 20 0 0 1 2 3 4 5 Strength of Smoothing GS LS(5x5) GS+LS(5x5) (b) PatchSize5×5 100 80 60 40 20 0 0 1 2 3 4 5 Strength of Smoothing GS LS(7x7) GS+LS(7x7) (c)PatchSize7×7 100 80 60 40 20 0 0 1 2 3 4 5 Strength of Smoothing GS LS(9x9) GS+LS(9x9) (d) PatchSize9×9 Figure 2: Share of retained keypoints: GS vs LS vs GS+LS For larger patch sizes, GS and LS perform al- most identically (which is clear considering the def- inition), while GS+LS is most effective in removing keypoints. For smaller patch sizes, GS is most ef- fective for high smoothing strength, while GS+LS is best for low smoothing strength. LS is not very ef- fectiveunder theseconditions. When applying techniques for keypoint removal, new keypoints are being created, e.g. at the edge of the patches in CA, LS, and GS+LS. This is not de- sired, as these new keypoints might match to exist- ing ones and thus aid the forensic analyst. Fig. 3 illustrates the creation of new, additional keypoints by showing the percentage of newly created ones. LS clearly introduces the lowest number of addi- tional keypoints, and if the size of the smoothing patch is increased then also the number of new key- points isalso increased. Thesmoothingstrengthalso plays a certain role: For weak smoothing, increasing the strength leads to more new keypoints, while af- ter reaching a peak, a further increase of smoothing strength decreases the number of newly created key- points. This effect is expected and most obvious for GS. In Table 2, the percentage of newly created key- points for CA is shown. Only LS with patchsize 3 givesbetter results, forallother techniqueswenotice higher percentages of newly created keypoints when comparingFig. 3 to thevalues inTable2. 168