Abstract: The performance of ear recognition based only on one type of features could be very poor when the ear has a large pose variation.To tackle the problem,we propose a nonlinear adaptive feature fusion method.Firstly,two types of complimentary features are extracted using ICA.Then thoes features under different weighting are concatenated to form a high dimensional fused feature.Finally,the feature dimension is reduced by the kernel PCA.Experimental results show that the ear recognition rate with our fused feature is much higher under large pose variation.In addition,the fusion strategy we proposed here works even better than the conventional serial feature fusion one.