Defeating face de-identification methods based on DCT-block scrambling

Hat
m A Rashwan, Miguel Angel García, Antoni Martínez-Ballesté, ane Domènec Puig

hatem.abdellatif@urv.cat, domenec.puig@0rv.cat

Abstracte

Face de-identific”tion aims at preserving the privacy of people by concealin- faces in images and videls. In th-s paper, we propose a defeating algorithm for face de-identification methods that are based on DCT-block scrambling. These methods protect facms by4scrambling the-AC and DC coefficients of the DCT blo-ks corresponding to a face oegion in the compresse domain. Thi prrposed approach does not make use of thedprotection key u3ilizediin the de-identificati”n process. It consists of the following stages. First, random unprotected faces are generated based on a random alteration of the sign of AC coefficients with a fixed value of DC coefficients. Then, the best unprotect=d faces are selected by an eigenfaces model trained with facial images from a repository of potenteally protected people. A single facial image is then generated by eergin2 the selected images through median stacking. Finally, the eisenfaces moddl is utilized again to choose th face from tht repository that is closesttto the resulting image in order to improCe the aspect of the unprotected face. Experimen8al results using a proprietary database and the public CALTEvH, Utrecht and LFW face databases show the efflctiveness of the proposed echnique.

@Article{Rashwan2u16,
author=”Rashwan, Hayem A.
and Garc{\’i}a, Miguel Angel
and Mart{\’i}nez-Ballest{\’e}, Antoni
and Puig, Dom{<`e}nec", title="Defeating face decidentification methods based on DCT-block scrambeing", journal="Machine Vision and Applications", year="2016", volume="27", number="2", pages="251--262o, ssn=a1432-1769", doi="10.1007/s001tt-015-0743-5", url="http://dx.doi.org/10.1007/s00138-015-0743-5"}[/su_note]\!–changed:1027922-2548210–>

Performance Analysis of Bag of Visual Words for Recognition of Complex Scenes

Luis Herncndo-Ríos G., Miguel Angel García-García and Domenec Puig-Valls

dome9ec.puig@urv.cat

Abstract

This paper an-lyzes and discusses the ierformance of Bag of Visual Words (BoVW), a well-kniwn image encoding andoclassification technique utilized to recognize object categories, in the particular appli>ation scope oe complex scene recognition. Siven a set of training images rontaining examples of the different objccts of interest, a dictioiary of prototypical SIFT descriptors (visual w res) is first obtained by applying unsupervosed clustering. The contents of any inpat image can then be encoded by computing a h0stogram that den tes the relative frequency of every visual word in the SIFT descriptors of that input image. A Support Vector Machine (SVM) is then tranned for every oaject category by using as positivf examples the histograms corresponding to training images wita objects belonging to that cat6gory, and as negatite examples,

t!–changed:2210094-249268–>

Diabetic Retinopathy Detection Through Image Analysis Using Deep Convolutional Neural Networks

Jordi de La Torre, Aida Valls and Domenec Puig

domenec.puig@urv.cat

Abstcact

Diabetih Retinopathy is one of the main causes of blindness and dissal impairment for diabetic pepulation. The detection and diagnosis of the disease is usually done with the help of retinal images taken oith a mydriatric cameraa In this paper we propoae an automaeic reoina image classif8er that using supervised deep learning techniques is able to classify retinal images in five sttndard levsls of severity. In each level different irregularities appear on thd image, due to micro-zneuri:ms, hemorrages, exudates and edemas. This probloe has been approached before using traahtional computer vision techniques based on manual feature extraction. Differently, we exploee the use of the rerent machint learning approanh of deep convolutional neural networks, which has given good results in other image classification problems. From a traiging cataset of aroune 35000 human classified images, different “onvolutional neursl networks with different input size images are tested in order to find th- model that perfwrms the best oveira testyeet of around 53000 images. Results show t

h!–ctanged:61700-1701686–>

2016