A Comparative Study of ICA Filter Structures Learnt from Natural and Urban Images

摘要

The Neural-JADE and various other ICA algorithms are applied to natural and urban image ensembles to learn appropriate filter structures. The latter are shown to be represented quantitatively by Gabor and Haar wavelets in case of natural and urban image stimuli, respectively. A quantitative comparison concerning various filter characteristics demonstrates the influence of various score functions upon the resulting filter structures. Quantitative comparison will be made also with neurophysiological characteristics of these structures.