Current existing sparse coding models employ the mean square of the error between the actual image and the reconstructed image to measure how well the code describes the image. Under the assumption that human visual perception is highly adapted for extracting structural information from a scene or a video, an alternative measure for information preservation assessment, based on the structural similarity, is introduced. After minimizing the cost function, the improved model attains a complete family of localized, oriented, and bandpass receptive fields, similar to those found in the primary visual cortex. The experimental results show that the improved sparse coding model is more consistent in human visual system.