Blind Image Deblurring via Bayesian Estimation Using Expected Loss

This paper introduces a new approach to single image blind deblurring via Bayesian estimation using expected loss, diverging from traditional maximum a posteriori (MAP) estimation methods that are limited by the delta kernel problem-a phenomenon where blur kernels are inaccurately estimated as delta functions due to the disparity of the number of pixels between the latent image and the kernel, leading to blurry latent images. We introduce the concept of robust intensity patch (RIP) values, which are median pixel values within local image patches, demonstrating remarkable stability through the blurring process. These RIP values are proposed as feasible substitutes for ground truth images, which are often unavailable in real-world scenarios due to the inherent difficulty of capturing both blurred and perfectly sharp versions of the same scene under identical conditions. We have developed a new loss function named robust intensity loss (RIL), designed to selectively penalize the latent image. This function aims to equalize the imbalance in the number of pixels between the latent image and the kernel. Through this approach, we have successfully redefined Bayesian estimation using expected loss as an optimization problem. Our contributions include the first application of Bayesian estimation using expected loss for single image deblurring, the introduction of RIP values, the development of the RIL function, and the integration with an advanced optimization scheme to significantly enhance deblurring accuracy. Our empirical results demonstrate the effectiveness of our method across various benchmark datasets, representing the way for future advancements in image deblurring.