Predicting segmentation errors in an iris recognition system

Abstract

Iris segmentation is defined as the isolation of the iris pattern in an eye image. A highly accurate segmented iris plays a key role in the overall performance of an iris recognition system, as shown in previous research. We present a fully automated method for classifying correctly and incorrectly segmented iris regions in eye images. In contrast with previous work where only iris boundary detection is considered (using a limited number of features), we introduce the following novelties which greatly enhance the performance of an iris recognition system. Firstly, we go beyond iris boundary detection and consider a more realistic and challenging task of complete segmentation which includes iris boundary detection and occlusion detection (due to eyelids and eyelashes). Secondly, an extended and rich feature set is investigated for this task. Thirdly, several non-linear learning algorithms are used to measure the prediction accuracy. Finally, we extend our model to iris videos, taking into account neighbouring frames for a better prediction. Both intrinsic and extrinsic evaluation are carried out to evaluate the performance of the proposed method. With these innovations, our method outperforms current state-ofthe-art techniques and presents a reliable approach to the task of classifying segmented iris images in an iris recognition system.