A new algorithm for face recognition based on wavelet transform and hidden Markov model(HMM) is proposed. Three low frequency sub-band images are selected by applying three-level wavelet transform. The low frequency wavelet sub-trees are formed by arranging three low frequency images in order. The feature wavelet sub-tree branches, as observation vectors of HMM, are derived by using independent component analysis. A set of images representing different instances of the same person is used to train each HMM. The relationship between the dimensionality of observation vectors and the recognition rates is shown. The effect of the number of states and Gaussian kernels on the system performance is examined. The essence of HMM is described qualitatively. Experimental results on the ORL face dataset are compared with other published algorithms, and show that the proposed algorithm has a high recognition rate with a good perspective.