The role of spectral entropy in signal analysis with applications to voice, still images, and video

Campbell[7] derived and defined a quantity called the coefficient rate of a random process that involves the process spectral entropy. This dissertation investigates the spectral entropy of various signals and methods of coefficient selection to represent the signals.; Spectral entropy is examined for autoregressive(AR) processes. The characteristics of the coefficient rate are investigated by comparing the coefficient rate with the entropy rate power.; The spectral entropy is represented as the integral of the log of the ratio over all possible distortion levels. The ratio is the slope of the AR power spectral density (PSD) over the slope of the uniform PSD for a given distortion. This ratio has information about the shape of the AR PSD, and the coefficient rate is interpreted as an indicator of a bandwidth in which most of the energy is contained.; The spectral entropy method for speech is developed and compared with the fixed number of coefficients method. Spectral entropy methods for linear prediction of both narrowband speech and wideband speech are proposed. The threshold to determine the number of coefficients to be selected in each component is calculated based on the output variance which results from selecting coefficients only from that component.; The spectral entropy method for video is developed and compared with the fixed number of coefficients method. The spectral entropy method is compared with the H.263 method under the constraint of same bit rate and same number of selected coefficients.; The spectral entropy method for images is developed and compared with JPEG, under the constraint of same bit rate. The spectral entropy method is compared with the embedded zerotree method under the constraint of same number of coefficients.