Seite - 38 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

Input image Convolution Convolution Max- pooling Convolution Convolution Fully connected layer NN Max- pooling 32 feature maps NN 32 feature maps NN 22 NN  32 feature maps 44 NN  64 feature maps 22 NN  64 feature maps 22 NN  64 feature maps 256 neurons 2 output neurons Figure1: The architectureof theproposedConvNetmodel. Figure2: Examplesofannotated tattoo images. eighth layer consists of two neurons with the Softmax activation function, corresponding to the two output classes. Dropout, with thedropout ratio set to0.5, is applied to the fullyconnected layer. We implemented thedescribed network inPython,usingTheano [2,3] andKeras 2 libraries. 4. Experiments Given the relatively modest volume of work on tattoo detection, there are no readily available tattoo detection datasets. Recently, a dataset called Tatt-C has been published [19], but it cannot be freely downloaded. Hence, to facilitate the development and testing of our method we have assembled our owndataset3 bycollectingandmanually labeling890tattooimagesfromtheImageNetdatabase[22]. Each of the collected images contains one or more tattoos. We annotated each tattoo using a series of connected line segments. Example annotated images from our dataset are shown in Fig. 2. We attempted tocloselycapture theoutlineofeach tattoo,whichcanbeachallenging task,as tattooscan havehighly irregular edges. 2https://github.com/fchollet/keras, accessed March 2016. 3The dataset is availableat http://www.fer.unizg.hr/demsi/databases and code/tattoo dataset. 38