On Feature Selection from Time-Frequency Images

The objective of the paper is to evaluate features of time-frequency images for change detection purposes. The features are obtained by extrapolation from time and frequency domain. Six features are considered from time domain, based on statistical moments. From frequency domain, the spectral flux and the spectral flatness are promoted. The feature selection process is based on the variance of the features, for the analyzed signal. The features with high variance are considered important and have assigned greater weights. Change detection is tested with a simple cusum (cumulative sum) criterion. The results obtained by computer-based simulations with real recorded signals reveal reliable discrimination of the change in the structure of the test signal. The design and the selection of the features in time-frequency domain is considered important and can be applied in various applications where time-frequency images are relevant.