AN ADAPTIVE OPTIMAL-KERNEL TIME-FREQUENCY REPRESENTATION

Time-frequency representations with fixed windows or kernels figure pr ominently in many applications, but perform well only for limited clas ses of signals, Representations with signal-dependent kernels can over come this limitation, However, while they often perform well, most exi sting schemes are block-oriented techniques unsuitable for on-line imp lementation or for tracking signal components with characteristics tha t change with time, The time-frequency representation developed here, based on a signal-dependent radially Gaussian kernel that adapts over time, surmounts these difficulties, The method employs a short-time am biguity function both for kernel optimization and as an intermediate s tep in computing constant-time slices of the representation, Careful a lgorithm design provides reasonably efficient computation and allows o n-line implementation, Certain enhancements, such as cane-kernel const raints and approximate retention of marginals, are easily incorporated with little additional computation. While somewhat more expensive tha n fixed-kernel representations, this new technique often provides much better performance, Several examples illustrate its behavior on synth etic and real-world signals.