We apply a Bayesian method for inferring an optimal basis to the problem of
finding efficient image codes for natural scenes. The basis functions lear
ned by the algorithm are oriented and localized in both space and frequency
, bearing a resemblance to two-dimensional Gabor functions, and increasing
the number of basis functions results in a greater sampling density in posi
tion, orientation, and scale. These properties also resemble the spatial re
ceptive fields of neurons :in the primary visual cortex of mammals, suggest
ing that the receptive-field structure of these neurons can be accounted fo
r by a general efficient coding principle. The probabilistic framework prov
ides a method for comparing the coding efficiency of different bases object
ively by calculating their probability given the observed data or by measur
ing the entropy of the basis function coefficients. The learned bases are s
hown to have better coding efficiency than traditional Fourier and wavelet
bases. This framework also provides a Bayesian solution to the problems of
image denoising and filling in of missing pixels. We demonstrate that the r
esults obtained by applying the learned bases to these problems are improve
d over those obtained with traditional techniques. (C) 1999 Optical Society
of America [S0740-3232(99)03107-5].