Abstract
M.Ing. (Electrical And Electronic Engineering)
Iris recognition systems have attracted much attention for their uniqueness, stability
and reliability. This recognition system is composed of four main modules,
namely, iris acquisition, iris segmentation, feature extraction and encoding and -
nally iris matching. However, performance of this system is a ected by poor image
quality. In this research, a novel iris image quality assessment method based on
character component is presented. This method is composed of two steps, individual
assessment of character quality parameters and fusion of estimated quality parameters
using Principal Component Analysis (PCA). The de ned quality parameters
considered in this research are entropy, sharpness, occlusion, dilation, area
ratio, contrast and blur. The designed technique was tested on three databases:
Chinese Academy of Science Institute of Automation (CASIA), University of Beira
Interior (UBIRIS) and Internal Collection (IC). Individual assessment of quality
parameters has shown that dilation, sharpness and blur have more in
uence on
the quality score than the other parameters. The images were classi ed into two
categories (good and bad) by human visual inspection. The e ect of the individual
parameters on each database is illustrated, with CASIA exhibiting higher quality
scores than the UBIRIS and IC databases. Support Vector Machine (SVM) and
Linear Discriminant Analysis (LDA) were used to evaluate the performance of the
proposed quality assessment algorithm. A k-fold cross validation technique was
employed to the classi ers to obtain unbiased results. Two performance measures
were used to rate the proposed algorithm, namely, Correct Rate (CR) and Area
Under the Curve (AUC). Both performance measures showed that SVM classi er
outperforms LDA in correctly classifying the quality of the images in all three
databases. The experimental results demonstrated that the proposed algorithm
is capable of detecting poor quality images as it yields an e ciency of over 84%
and 90% in CR and AUC respectively. The use of character component to assess
quality has been found to be su cient, though there is a need to develop a better
technique for standardization of quality. The results found using a SVM classi er
a rms the proposed algorithm is well-suited for quality assessment.