Artificial Neural Networks For Dictionary Selection in Adaptive Greedy Decomposition Algorithms With Reduced Complexity

In signal processing, adaptive greedy decomposition algorithms have been used to obtain sparse signal approximations. A representative example is the Matching Pursuit (MP) algorithm, it iteratively approximates signals through a weighted sum of waveforms selected from a redundant dictionary. These fundamental waveforms or atoms in the dictionary are usually produced by scaling, translating and modulating prototype functions, such as a Gaussian or an exponential. Each prototype function defines a sub-dictionary. Dictionaries comprising a variety of sub-dictionaries may provide higher sparsity levels, but demand higher computational costs. In this paper, we propose the Dictionary-Artificial-Neural-Network-Oriented (DANNO) MP algorithm to mitigate this dimensionality burden. Artificial Neural Networks (ANNs) are used to design a classifier that selects the best sub-dictionary to be employed at each MP iteration instead of using the complete dictionary. Simulation results show that the proposed algorithm can lead 2.5 times lower computational complexity than the MP for multiple sub-dictionaries scenarios, without compromising signal approximation performance.