An improved LSA model is proposed, and it is applied to face recognition. Firstly, the model is established to represent the map between the pixels residing at images of various resolutions; after that, parameters of the model are obtained from original image and its wavelet decomposition results to decide the map; finally, the identified map is used to estimate images at finer resolution from coarser versions. Parameters of the model about testing image and about training images are compared to classify the testing image. Experimental results show that images estimated by presented model are more similar to target images than those estimated by LSA model. The face recognition system based on the presented model is robust to illumination.