Adaptive diffusion and discriminant analysis for complexity control of time-frequency detectors

Achieving good performance with data-driven detectors requires matching their complexity to the amount of available training data. Receivers with a too large number of adjustable parameters often exhibit poor generalization performance whereas those characterized by an insufficient complexity cannot learn all the discriminant information carried by training samples. This paper deals with the complexity control of data-driven time-frequency detectors. Our approach is based on a locally adaptive filtering technique for time-frequency representations, called adaptive diffusion. It consists in using discriminant analysis to design the conductance function that controls the diffusion process. The resulting filtering scheme preserves discriminant information while acting on the complexity of the time-frequency detector. Simulation examples illustrate the efficiency of our approach.