Jun'ya Shimizu

Adaptive signal enhancement by a robust TLS estimation

An adaptive enhancement algorithm of signal disturbed by the impulse noise and the white Gaussian noise is proposed in this report. If the strong dependence between a signal enhancement and a needed-parameter estimation exists in an algorithm, it is difficult to reduce a large bias occurring in the first estimation. To weaken their relationship, the introduction of a total least squares (TLS) algorithm that estimates the model parameters directly from the disturbed signals is considered. However, the TLS estimation accuracy is dramatically deteriorated by the impulse noise, even if the TLS estimation accuracy is held to the non-Gaussian noise in some degree. Hence, we ensure the robustness of the algorithm by replacing a high amplitude signal with an estimated value based on a likelihood ratio test. Using these signals, we develop, an algorithm based on the TLS and the EM algorithm for enhancing the disturbed signal. We also show the effectiveness of the proposed algorithm through computer simulations.

An estimation of time-varying parameters using multi-AR lattice models in subbands

Analysis methods of time-varying signals have been used in many signal processing systems. In particular, the methods using a parametric model have been extensively studied. It is required to realize real-time processing and high estimation accuracy. In this report, we propose a new adaptive method to estimate time-varying AR parameters. The method uses multi resolution criteria based on the concept of multirate processing. We address several issues related to the conventional method using a unique least squares criterion. To overcome these issues, we develop an estimation method using AR lattice models in subbands; model orders and forgetting factors are also adaptively determined in each subband.<<ETX>>

Consideration on decimation factors in multirate adaptive filtering for a time-varying AR model

Multirate adaptive filters for the identification of time-varying systems are very important. However, a decimation factor to be used in the multirate adaptive filtering for the identification of time-varying system has not been considered quantitatively. We derive two conditions which limit the maximally decimated factor used in the time-varying system identification.