Local area within a normal natural image can be thought as a stationary process. This can be modeled well using autoregressive models. In this paper, a set of autoregressive models will be learned from a collection of high-quality image patches. Out of these models, one will be selected adaptively and will be used to regularize the input image patches. In addition to the autoregressive models, a non-local self-similarity condition was proposed. The autoregressive models will exploit local correlation of individual image, but a natural will have many repetitive structures. These structures, which are basically redundant, are very much useful in image deblurring. The performance of these schemes is verified by applying to mage deblurring.
Volume 11 | 02-Special Issue
Pages: 1728-1736