MSF-VQ is a kind of image features for face recognition. Firstly, vector quantization histogram feature is calculated with the predetermined codebook. Then, the histogram is extended through Markov stationary feature to get MSF-VQ features. MSF-VQ features for face recognition show high recognition rate. But it can still be improved on the determination of codebook and spatial information expression. According to these two points, this study puts forward an improved method. It first calculates the codebook based on the facial data set, so as to improve face resolution capability of vector quantization histogram, and then combines several MSF features calculated from different directions sampling, to increase spatial location information contained by MSF-VQ feature. Experimental results show that the improved MSF-VQ method has higher face recognition rate.