Seite - 35 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

TattooDetection for SoftBiometricDe-Identification Based on ConvolutionalNeuralNetworks∗ Tomislav Hrkac´1,KarlaBrkic´1, andZoranKalafatic´1 1FacultyofElectricalEngineeringandComputing UniversityofZagreb, Croatia {tomislav.hrkac, karla.brkic, zoran.kalafatic}@fer.hr Abstract Nowadays, video surveillance is ubiquitous, posing a potential privacy risk to law-abiding individu- als. Consequently, there is an increased interest in developing methods for de-identification, i.e. re- moving personally identifying features from publicly available or stored data. While most of related work focuses on de-identifying hard biometric identifiers such as faces, we address the problem of de-identification of soft biometric identifiers – tattoos. We propose a method for tattoo detection in unconstrained images, intended to serve as a first step for soft biometric de-identification. The method, based on a deep convolutional neural network, discriminates between tattoo and non-tattoo image patches, and it can be used to produce a mask of tattoo candidate regions. We contribute a dataset of manually labeled tattoos. Experimental evaluation on the contributed dataset indicates competitiveperformanceofour method andproves itsusefulness inade-identificationscenario. 1. Introduction Inthelastdecade,videosurveillancehasspreadtoalmostallaspectsofdailylife. Storingtherecorded surveillance data in its unprocessed form poses a privacy risk to law-abiding individuals, as their whereabouts and activities can be exposed without their consent. Privacy concerns are aggravated by thedevelopmentofvariousvideoretrieval techniques [17,26,16] thatenablesearchingforcontent in large volumes of video data, as well as by the development of techniques for person re-identification across different video sequences [1, 8]. In order to minimize privacy risks, many jurisdictions imple- mentstrict regulationsfor theprotectionofpersonaldata (seee.g. theDataProtectionDirectiveof the European Union1). For video sequences, protection of personal data entails obfuscating or removing personally identifying features of the recorded individuals, usually in a reversible fashion so that law enforcementcan access them ifnecessary. The process of removing personally identifying features from data is called de-identification. One of the most commonly used de-identification techniques, used in commercial systems such as Google StreetView, involvesdetectingandblurringthefacesofrecordedindividuals. However, thisapproach ignores soft biometric and non-biometric features like clothing, hair color, birthmarks or tattoos, that can be used as cues to identify the person [6, 20]. In this paper, we propose a method for detecting ∗Thisworkhasbeensupportedby theCroatianScienceFoundation,within theproject”De-identificationMethods for Soft andNon-Biometric Identifiers” (DeMSI, UIP-11-2013-1544). This support isgratefully acknowledged. 1http://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:31995L0046 35