On the basis of the compression sensing algorithm, this paper proposes to train a projection matrix in the process of dictionary learning algorithm, through which the method can obtain the sparse coefficient directly. The dictionary training process is based on the KSVD dictionary learning algorithm and is compared with the traditional L1 norm solving algorithm. It can be seen from the experiment that the method has more rapid and higher recognition rate than the traditional L1 algorithm using greedy method. The algorithm can solve the coefficient term directly through the matrix operation, while the latter is an NP problem, which needs to be solved by the iterative algorithm. For the large sample test, the proposed algorithm has better application space, and the time of saving will very noticeable.