In this paper, we evaluate performances of various feature extraction methods for iris pattern classification. Generally, an identification system using iris recognition consists of 4 stages: image acquisition, preprocessing, feature extraction and pattern matching. In this paper, we used the 2D bisection-based Hough transform and the radius histogram method for localizing the iris and used multilayer neural networks as a classifier. In order to further reduce the number of features, three linear feature extraction methods are evaluated. In particular, by using an efficient feature extraction algorithm, we explore the possibility to reduce the classification time and system complexity for a large iris data set. The tested feature extraction methods are the feature extraction based on decision boundary, canonical analysis, and principal component analysis. Experiments with 1831 iris images show that the feature extraction based on decision boundary and canonical analysis provide a favorable performance.