SVD Filter Based on Noise Singular Values Clustering

Aiming at determining the optical reconstruction rank of singular value matrix in SVD filter, a new method based on noise singular values clustering is proposed in this paper. Variable forward standard deviation (F-Std) has been defined and the mutation point in F-Std is used to separate signal singular values from noise singular values. Experiments have been simulated by some noisy signals using the new SVD filter and other two commonly used methods, the results and root mean square errors of these filters are shown that the new SVD filter is not only optimal or near to optimal and more effective than the two methods but also robust to the shape of the trajectory matrix even when the signal´s SNR varies from very high to -11.7dB.