Improved spatially adaptive MDL denoising of images using normalized maximum likelihood density

This paper presents a new method for wavelet denoising using minimum description length (MDL) principle with normalized maximum likelihood density. Denoising is done by hard thresholding and a new spatially adaptive threshold which varies according to the estimated signal variance of each wavelet coefficient is derived using the MDL principle with normalized maximum likelihood density. As the normalized maximum likelihood code encodes the data with the shortest description length, smaller proportion of significant coefficients could be achieved after thresholding compared with simple MDL denoising. Thus better compression is obtained without detoriating the denoising performance measure (PSNR) compared to the MDL thresholding.