To improve accuracy of non-circular iris segmentation, an iris segmentation algorithm based on linear basis function model is proposed. The proposed algorithm regards iris segmentation problem as a machine learning problem which derives iris boundary curve from iris boundary points. First, boundary points are located by coarse segmentation. Second, a linear basis function model is constructed to derive boundary curve from boundary points, thus the segmentation is complete. Compared with traditional iris segmentation algorithms, the proposed method locates non-circular iris with high accuracy, with the expense of a few more time. Experiment results on the dataset of CASIA-3.0 IRIS-Interval reveal that the accuracy of proposed method is 99.92%, and the time cost is 1ms more, which is capable for state-of-the-art iris recognition systems.