Page - 40 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

Patchsize 8×8 12×12 16×16 24×24 32×32 48×48 Falsenegatives 152 (8.03%) 229 (12.10%) 187 (9.88%) 213 (11.25%) 248 (13.10%) 290(15.32%) Falsepositives 593 (31.37%) 418 (22.12%) 444 (23.49%) 436 (23.07%) 337 (17.83%) 408(21.59%) Accuracy 0.8031 0.8290 0.8332 0.8283 0.8454 0.8155 Table1: Evaluationof thenetworkperformanceon differentpatchsizes geneousskinaround the tattooandsimplebackground. Wesee thatmany tattoopatchesarecorrectly detected,but therearealsosomemisclassifications. Inmoredifficultexampleswithmorebackground containing textured objects, the number of false positives rises. In the context of de-identification, this problem could be addressed by combining this detector with other stages of a de-identification pipeline, e.g. byeliminatingdetections outsideofcandidateperson locations. (a) the original images (b) labeled tattoopatches Figure4: Theoutputof thenetwork on full images. 5. Conclusion andoutlook Weaddressed thechallengingproblemof tattoodetectionforsoftbiometricde-identification. Instead of hand-crafting image features, we applied deep learning. We trained and evaluated a deep convo- lutional neural network using the dataset of positive and negative patches generated from a subset of ImageNet tattoo images annotated by hand. Our findings indicate that using a convolutional neural network toclassify small imagepatchescanbea reliableway todetect candidate tattoo regions inan image. Patchsizesshouldbekept small, up to32×32patches, inorder toobtainbestaccuracy,good foreground-backgroundsegmentationandminimize falsenegatives. In our future work, we plan to combine this method with other stages of a de-identification pipeline inorder to solve theproblemof falsepositives. Asourqualitativeanalysis shows that themajorityof false positives are in the surroundings rather than on the person, one possibility is to run the method only on the outputs of a person detector. We also plan to quantitatively evaluate the performance of our network on full tattoo images (as opposed to patches), and investigate whether this performance couldbe improvedbymerging thedetections intoblobs. 40