A method of reference model illumination mapping is proposed to deal with the problem of one sample per person. To enhance the classification information of single training sample, we extend virtual images generated from the given single training image. Using discrete wavelet transform (DWT), the low-frequency band is processed to map the illumination information from the reference model to create a virtual sample. Principle-component-analysis plus linear-discriminant-analysis (PCA+LDA) is performed on the virtual training set to extract features. Experiments are performed on the Yale B and extended Yale B facial image database. The results show that, compared with the PCA feature extraction with single sample, recognition rate is significantly higher using the proposed method.