Texture-preserving deconvolution via image decomposition

Abstract

The image with rich textures can be decomposed into the sum of a geometric part and a textural part. Inspired by this fact, we propose an efficient texture-preserving image deconvolution algorithm based on image decomposition. Our algorithm restores the geometric part and textural part, respectively, by incorporating \(L_0\) gradient minimization and a wave atoms-based Wiener shrinkage filter. The \(L_0\)-based gradient minimization method could globally locate important edges, main structures. The wave atoms transform offers a better representation of images containing oscillatory patterns and textures than other known transforms. Our method contains three steps for restoring texture images. First, we propose an image deconvolution method based on \(L_0\) gradient minimization to restore geometric part of the image with minimal loss of image detail components. Next, we use a Wiener shrinkage filter in the wave atom domain to attenuate the leaked colored noise and extract fine details. Finally, we obtain the estimated image by adding the two image parts together. We compare our deconvolution algorithm with other competitive deconvolution techniques in terms of ISNR, SSIM, and visual quality.