By learning property of human fingerprint in bioinformatics, we used SURF to extract frames' features and handle them by visual vocabulary and word frequency analysis. Videos can be represented uniquely in this way. Based on the theory of image sparse coding in mammal's visual system, we used standard library to train sparse dictionary which could encode SURF features of frames. By dealing with nonzero value only, we found that the cost of storage and computation was reduced. The experiment and simulation showed that optimized result could maintain the robust of original features and had good discrimination and accuracy.