Normalization Effects in Matching Pursuit Algorithm with Gabor Dictionaries

Abstract

The matching pursuit (MP) algorithm is a greedy method for signal decomposition used in video coding, data compression, and, particularly, analysis of EEG signals in various paradigms, including P300 and ER(D)S (motor imagery). An important issue for MP implementation is a correct treatment of normalization of atoms (functions) used in computations. Failing to account for normalization-related effects may affect both the numerical stability and the reliability of the algorithm. This paper describes these normalization effects, evaluates their impact on the algorithm’s performance, and describe the proper approach together with a ready-to-use implementation, available under a General Public Licence (GPL). Several performance optimizations used as a part of this implementation are also described.