Computational complexity is a matter of great concern in real time face recognition systems. In this paper, four state hidden Markov model for face recognition has been presented whereby coefficients of feature vectors have been curtailed. Face images have been divided into a sequence of overlapping blocks. An observation sequence containing coefficients of eigen values and eigenvectors of these blocks have been used to train the model and each subject is associated with a separate hidden Markov model. The computational complexity of the proposed model has been minimized by employing discrete wavelet transform in the preprocessing stage. Furthermore, singular value decomposition has been employed on face images and a threshold singular value is determined empirically to reject or accept test images. Principal component analysis is used for feature extraction. Accepted test images are classified based on the majority vote criteria using different observation sequences of image features. Experimental findings on Yale and ORL databases in noisy such as Salt and Pepper and noise free environments reveal that the recognition accuracy of the proposed model is comparable to the existing techniques with reduced computational cost.