Iris recognition is a prospering biometric method, but some technical difficulties still exist. To get more representative iris features, features from space and frequency domain are extracted at the same time. Both variation fractal dimension and wavelet features are extracted to form the feature sequence. Multi-objective genetic algorithm is employed to optimize the features. Finally the selected features of different iris patterns are used to train iris classifiers. Furthermore, traditional SVM is modified as non-symmetrical support vector machine to satisfy the variant security requirements in real iris recognition applications. Experimental data shows that the new feature sequence represents the variation details in the iris patterns properly. Therefore the correctness is improved.