Learning sparse multiscale image representations

Phil Sallee and Bruno A Olshausen

Presented at
Neural Information Processing Systems, Vancouver, December 2002.


We describe a method for learning sparse multiscale image representations using a sparse prior distribution over the basis function coefficients. The prior consists of a mixture of a Gaussian and a Dirac delta function, and thus encourages coefficients to have exact zero values. Coefficients for an image are computed by sampling from the resulting posterior distribution with a Gibbs sampler. The learned basis is similar to the Steerable Pyramid basis, and yields slightly higher SNR for the same number of active coefficients. Denoising using the learned image model is demonstrated for some standard test images, with results that compare favorably with other denoising methods. 
Reprint (2.1M, ps) / Reprint (941k, pdf) / Poster (6.9M, pdf)