In order to extract the impact feature from rolling bearing fault vibration signal, which is significant for bearing fault diagnosis, a signal denoising method based on SVD (Singular Value Decomposition) of S transform time-frequency spectrum is proposed. S transform is a means of signal time-frequency representation and particularly suitable for processing the non-stationary signal with impact feature. During SVD denoising, the target data matrix is composed of S transform spectrum coefficients. The position of the threshold singular value, be less than or equal to which the singular value will be set zero, can be determined by the last peak index of the peaks swarm in singular value difference spectrum. Finally, inverse S transform of the data matrix resulted from SVD denoising is made to reconstruct the time domain impact feature. The simulation results show that the proposed method can successfully extract the periodic impact feature from low SNR signal. In the processing of the rolling bearing fault vibration signals, this method is able to obtain the impact feature frequency, which can be used to diagnosis bearing relevant faults effectively.