Non-greedy adaptive compressive imaging: A face recognition example

We present a non-greedy adaptive compressive measurement design for application to an M-class recognition task. Unlike a greedy strategy which sequentially optimizes the immediate performance conditioned on previous measurement, a non-greedy adaptive design determines the optimal measurement vector by maximizing the expected final performance. Gaussian class conditional densities are used to model the variety of object realization for each hypothesis. The simulation results demonstrate that non-greedy adaptive design significantly reduces the probability of recognition error from greedy adaptive and various static measurement designs by 22% and 33%, respectively.