A multi-scale model named GNSA is proposed under the condition of taking account of approximate image information and detailed image information. The mapping relationship among different scale images is established by using a neural network with a hidden layer and 50 hidden units. This mapping relation is determined by training neural network with a back-propagation algorithm, which is utilized to estimate images at finer resolution from coarser versions. Similarity in brightness and similarity in contrast are used to gauge the degree of similarity between the estimated images and target images. The experiments show that two similarities between images estimated by GNSA model and target images are 89.907% and 96.196% respectively. The face recognition system based on the model has a good robustness for light illumination.