Eyelid Position Detection Method for Mobile Iris Recognition. Gleb Odinokikh FRC CSC RAS, Moscow

Size: px

Start display at page:

Download "Eyelid Position Detection Method for Mobile Iris Recognition. Gleb Odinokikh FRC CSC RAS, Moscow"

Logan Boyd
6 years ago
Views:

1 Eyelid Position Detection Method for Mobile Iris Recognition Gleb Odinokikh FRC CSC RAS, Moscow 1

2 Outline 1. Introduction Iris recognition with a mobile device 2. Problem statement Conventional eyelid detection approach Existing methods examples Drawbacks of existing methods 3. Proposed solution Idea description 4. Algorithm description 5. Experimental results Eyelid detection accuracy measurement Testing data information & results Summary 2

3 Introduction Iris recognition with a mobile device: Unconstrained environmental conditions User interaction difficulties Performance limitations Insufficient iris illumination Reflections from glasses High eyelid occlusion Iris image quality degradation Operation delays Recognition time increasing Processing time increasing High gaze angle Overexposure High pupil deviations User inconvenience Recognition errors Redundant power consumption (a) (b) (c) (d) Fig. 1 Image quality degradation: (a) gaze away, eyelid occlusion, (b) overexposure, (c, d) poor contrast, reflection from glass surface 3

Problem statement Conventional eyelid detection approach: Eyelid noise removal Iris image quality estimation Image Acquisition Eye Region Detection Pupil-iris Boundary Localization Iris-Sclera

4 Problem statement Conventional eyelid detection approach: Eyelid noise removal Iris image quality estimation Image Acquisition Eye Region Detection Pupil-iris Boundary Localization Iris-Sclera Boundary Localization Eyelid Boundary Localization Iris-sclera boundary localization Eyelid boundary localization Eyelash Detection Drawbacks of existing methods: Not robust in case of unconstrained conditions Most of them are computationally complex Performed after complex operations of iris center definition and irissclera boundary localization Iris Image Normalization Iris Feature Extraction Iris Pattern Matching Fig. 2 Iris recognition algorithm flowchart 4

5 Problem statement Existing methods examples: max a,k,h a k F d dh h y k 2 4a(x h) Author Preprocessing Localization Daugman Gaussian blur Parabolic IDO Wildes Sobel Hough transform y Vertex (k, h) Focal point (k, h) a Masek Sobel Line fitting (least squares) Kang & Park Sobel (modified) Parabolic IDO Xiangde et al. 1D peak shape filter Parabolic IDO Adam et al. Anisotropic diffusion Hough transform x Yang et al. Asymmetric Canny Parabola fitting (least squares) Kim, Cha at al. Histogram equalization Local minima search He et al. 1D rank filter Pre-established model fitting Fig. 3 Parabolic Integro-Differential Operator (Parabolic IDO) 5

6 Proposed solution Idea: Detection of eyelid position earlier: for definition of E u and E l points (see pic. below) right after pupil-iris boundary localization stage Use this information further for: eye opening condition estimation iris-sclera boundary localization algorithm parameters readjustment full eyelid boundary localization/refinement If eye isn t opened enough: proceed to the next frame immediately provide user with feedback like: Open eye fully Image Acquisition Eye Region Detection Pupil-iris Boundary Localization Eyelid Position Detection Is eye opened enough? no feedback to user yes E u Iris-Sclera Boundary Localization P c Eyelid Boundary Refinement E l Eyelash Detection Fig. 4 Eyelid position points E u upper eyelid position point, E l lower eyelid position point, P c pupil center point Fig. 5 Proposed flowchart modification 6

7 6xRp 6xRp Algorithm Description start Preprocessing stage image cropping image downscaling 4xRp image separation 2xRp E u upper eyelid image preprocessing lower eyelid image preprocessing edge map extraction using 1d n-peak gradient search du Detection stage dl vertical sliding window response calculation choosing the window with the max response getting of edge point y-coordinate closest to pupil center Input Frame (blurred) Division (relative to Pupil center) Edge map extraction Searching the Points E l defining the point as the upper eyelid position defining the point as the lower eyelid position End R p pupil radius Fig. 6 proposed algorithm structure 7

7 L range U range Algorithm Description Bitwise Operation

(Sliding Window) no yes no yes Getting N-peak Edge Map Fig.

8 7 L range U range Algorithm Description Bitwise Operation with pupil mask Image Separation Response Calculation (Sliding Window) no yes no yes Getting N-peak Edge Map Fig. 8 Choosing between two peaks rule Image rows Directional 2D Gabor Filtering Fig. 7 Proposed algorithm steps Final eyelid position Fig. 9 Final eyelid position selection from edge map 8

9 ±5% height = 100% ±10% ±15% Experimental Results Eyelid detection accuracy measurement: evaluation method admissible error rate: ξ adm 1..Ne = {5%, 10%, 15%} E(x, y M ) i E(x, y A ) i ±5% ±10% ±15% accepted (in range) error (not in range) Fig. 10 Different admissible error examples Fig. 11 Correct and incorrect eyelid position definition examples E j define a part of the images in single dataset are not accepted (found eyelid position isn t in admissible range): E j = 1 N i: E(x, y A) i E(x, y M ) i > ξ j adm height Then E j averaged for different datasets: AVG ξj adm = 100% N D N D (1.0 E j i ) where E(x, y A ) i - eyelid point detected by algorithm, E(x, y M ) i - eyelid point manually marked, N D - number of datasets used 9 i=1

10 Accuracy, AVG Accuracy, AVG Experimental Results Testing data information & results: 4 different datasets are used: MIR-Train, CASIA4-Thousand, CASIA3-Lamp and AOPTIX 500 images of each dataset have been cropped and manually marked by expert and used for testing % 10% 5% Admissible error rate %, ξadm % 10% 5% Admissible error rate %, ξadm Daugman Xiangde Adam Masek Proposed 2DGF+IDO Kang & Park Wildes Yang Kim He Daugman Xiangde Adam He Kang & Park Wildes Proposed 2DGF+IDO Fig. 12 Upper eyelid detection accuracy Fig. 13 Lower eyelid detection accuracy 10

11 Experimental Results Accuracy testing results for ξ adm = 5% Table 1. Upper eyelid detection accuracy (%) for ξ adm = 5% Dataset MIR CS4 CS3 APX AVG Daugman ,4 Wildes ,6 Masek ,6 Kang & Park ,0 Xiangde et al ,2 Adam et al ,2 Yang et al ,4 Kim, Cha et al ,0 He et al ,6 2DGF+IDO ,3 Proposed ,8 Table 2. Lower eyelid detection accuracy (%) for ξ adm = 5% Dataset MIR CS4 CS3 APX AVG Daugman ,8 Wildes ,3 Masek ,5 Kang & Park ,3 Xiangde et al ,8 Adam et al ,5 Yang et al ,5 Kim, Cha et al ,5 He et al ,3 2DGF+IDO ,8 Proposed ,8 Summary Proposed method: outperform all the existing methods by accuracy reliable robust on different datasets applicable for mobile applications, could be used for another purposes: gaze tracking, fatigue detection etc. allows to detect eyelid position on early stages saves processing time, allows to give user a feedback quickly processing time is about 1ms on Snapdragon 800 (2,26GHz), single core fast & simple 11

12 Q&A 12

13 Thank you. 13

A Fast and Accurate Eyelids and Eyelashes Detection Approach for Iris Segmentation

A Fast and Accurate Eyelids and Eyelashes Detection Approach for Iris Segmentation Walid Aydi, Lotfi Kamoun, Nouri Masmoudi Department of Electrical National Engineering School of Sfax Sfax University