Iris and Periocular Recognition Using Mobile Phones Work done with Asem Othman Arun Ross Associate Professor Michigan State University rossarun@cse.msu.edu http://www.cse.msu.edu/~rossarun 2 Biometrics in Mobile Phones Here, the term “mobile” is referred to modern phones, tablets and similar smart devices 3 Ocular Biometrics in a Smartphone Performing ocular recognition in a smartphone environment Ocular images acquired using a smartphone camera Texture of dark-colored irises not easily resolved in color images Images acquired in various types of lighting conditions Resolution and view-point of images can change Images can be blurred or occluded with glasses 4 Work Elsewhere Smart Sensors Limited has ported their iris SDK into the Android OS AOPTIX STRATUS MX Some smartphone devices use NIR accessories to perform iris recognition AOptix Stratus MX is an iPhone hardware peripheral equipped with an iris sensor EyeVerify uses scleral patterns to authenticate a mobile phone user EYE VERIFY 5 MICHE Database Color images acquired using iPhone5, Galaxy S4, Tablet 2 Iris Periocular iPhone 5 Samsung Galaxy S4 Samsung Galaxy Tablet 2 De Marsico et al, “FIRME: Face and Iris Recognition for Mobile Engagement,” IMAVIS 2014 6 Images from iPhone 5 ###= subject number, IN = indoor, OU = outdoor, F =Front cam, R = Rear Cam, RI = right eye 001_IP5_IN_R_RI_01_1 001_IP5_IN_F_RI_01_3 001_IP5_OU_R_RI_01_1 001_IP5_OU_F_RI_01_1 7 Sensor Positioning 8 Illumination 9 Other Challenges 10 Periocular Biometrics Periocular recognition focuses on the region in the vicinity of the eye This is on a face database PERIOCULAR FACE Partial face image Periocular region segmentation for partial faces Park, Jillela, Ross, Jain “Periocular Biometrics in the Visible Spectrum,” TIFS 2011 11 Block Diagram of Analysis Methods to enhance images Methods to extract robust features from images Different distance measures for matching 12 Illumination Normalization Multi Scale Retinex (MSR) N RMSR n {log I ( x, y) log[ Fn ( x, y ) * I ( x, y)]} n 1 Multi Scale Self Quotient (MSQ) N RMSQ nT ( I ( x, y ) /( Fn ( x, y) * I ( x, y))) n 1 Homomorphic Normalization (HOM) f(x,y) ln DFT H(u,v) (DFT)-1 exp Wavelet-based Normalization (WA) Apply histogram equalization to detailed wavelet coefficients g(x,y) 13 Feature Extraction Segment and align the periocular region Tessellate image into overlapping patches Apply texture descriptor to each patch Compute histogram of texture descriptors for each patch Concatenate all histograms to obtain a fixedlength feature vector 14 Alignment Alignment based on eye corners 15 Texture Descriptors Used Local Binary Pattern (LBP) Local Phase Quantization (LPQ) Binarized Statistical Image Features (BSIF) Leung-Malik Filter (LMF) Histogram of Oriented Gradients (HOG) 16 Leung-Malik (LM) Filter Bank The LM filter bank has a mix of edge, bar and spot filters at multiple scales and orientations T. Leung and J. Malik, “Representing and recognizing the visual appearance of materials using three-dimensional textons,” IJCV 2001 17 Distance Metrics Used Euclidean distance (EUC) Chi-squared distance (CHI-SQ) Cosine measure (COS) Earth Mover's Distance (EMD) Histogram Intersection distance (HI) 18 Experiment Results Metric: Verification rate at 1% FAR BSI LMF LPQ F 51 % 55 % 73 % 50% 65% 80% 61 % 70 % 75 % HO G 53% 65% 75% Before Alignment (HOM) After Alignment (MSR) Without Glasses and In-Focus (MSR) 19 Summary Iris biometric did not perform well on this database Ocular biometric was more competitive Best performance: Multi Scale Retinex (MSR) in conjunction with HOG, BSIF, or LPQ Using only in-focus images without glasses further improves performance 20 Future Work Accurate alignment Quality-based image enhancement Subspace analysis of texture descriptors Cross-camera evaluation Porting code into smartphone platform Iris and Periocular Recognition Using Mobile Phones Work done with Asem Othman Arun Ross Associate Professor Michigan State University rossarun@cse.msu.edu http://www.cse.msu.edu/~rossarun