High-quality Non-blind Image Deconvolution

The most annoying artifacts in image deconvolution are ringing and amplified noise. These artifacts can be reduced significantly by regularization using the Maximum a Posteriori (MAP) method that exploits not only the likelihood but also the image prior in image deconvolution. Although ringing and noise can be reduced significantly with strong regularization, image details are also reduced, so the deconvolved image is neither sharp nor clear. In this paper, we propose a non-blind image deconvolution method with adaptive regularization that can reduce ringing and noise more noticeable in a smooth region and preserve image details in a textured region. For adaptive regularization, after we make a quick estimate of the reference image that can indicate the strength of regularization, we perform regularization adaptively according to the local characteristics. Experimental results show that ringing and noise are suppressed significantly, while preserving image details effectively.