According to the analysis of characteristics of the SVD and wavelet threshold de noising methods, the compound de noising method based on discrete stationary wavelet transform and optimized SVD was put forward. The effective rank of the matrix is optimized by the structural risk minimization principal of the statistical learning theory, solving the problem of choosing SVD de noising eigenvalue. Aiming to the shortage of the oscillation effects of traditional discrete binary wavelet transform in singularity and ignoring the noise influence of the approximation coefficients, the improved discrete stationary wavelet transform method was used to do threshold de noising after the SVD de noising. At last, the de noising experiments are carried out compared with the traditional discrete binary wavelet transform under different conditions. The validity and superiority of the method are proved by the experiment.