Adaptive Thresholding in Extracting Useful Information From Noisy Time- Frequency Distributions

This paper provides an analysis of the performance of an automatic method for extraction of useful information content from time-frequency distributions of nonstationary signals in dependence on the selected time-frequency method. The tested algorithm for the extraction of the signal components (useful information) from the noisy mixture is based on an initial segmentation of the time-frequency distribution which provides a fixed number of data classes. The normalized energies of the different classes are used as input to a statistical test which produces two outputs: “useful information” classes and “noise” classes, respectively. The quantity used as indicator of the class type, being the normalized energy of one class, is highly dependent on the time-frequency kernel filter. This paper reports the results of the proposed method applied to three well performing time-frequency methods, the Smoothed-Pseudo Wigner-Ville distribution, the Choi-Williams distribution, and the Modified-B distribution. The performance comparison attests the method's robustness for the different kernel filters, in various SNRs.