On the use of Renyi Entropy of Time-Frequency Images for Change Detection

The aim of the paper is to estimate the Renyi entropy of time-frequency images, as descriptors of the information content and change detection purposes. The Renyi entropy is estimated by two approaches. Firstly, the image is properly normalized to estimate a probability density function. This approach is called Direct. Secondly, by using a statistical model based on probabilities, given by the histogram of the image. The approach is named Indirect. The estimation approaches are evaluated on artificially generated signals, commonly used in the field of communication engineering. Both estimations have no information about spatiality. The estimated entropies could be used as features extracted from time-frequency images. A change detection criterion is promoted based on cumulative sum function applied to the estimated entropies, followed by a double statistical expectation. The experiments show an optimum working point, to maximize the change detection criterion. The decomposition of the time-frequency image and, next, the computation of the Renyi entropy on sub-images or regions of interest, seems to be an interesting solution to follow.