Noisy Blurred Image Restoration Based on Noise Variance Estimation in Blind Condition
Chong Yi · Sakura: Saitama United Cyber Repository of Academic Resources (Saitama University) · 2014
Many image restoration algorithms have been proposed over the last few decades. But in early work on image restoration, it was nearly always assumed that all the information required to restore an image is known previously. Unfortunately, this is not possible in most real-life situations. Thus, the technique of blind image restoration estimating both the true image and the blur from a degraded image has been researched. However, when taking noise into account, the estimation problem becomes more challenging. Estimating the blue and noise parameters simultaneously may cause a large estimation error. To minimize the estimation error, we propose a Maximum-Likelihood Estimation Algorithm based on noise variance estimation. This method estimates the noise variance only using the information from the known degraded image. This improves the estimation accuracy significantly. For better results for noise variance estimation from the degraded image, we propose a structure-based method. This method separates an image into blocks. Rejecting the edge included blocks, only the homogenous blocks are selected for the noise estimation process. Since image details can be better revealed by second-order operators, for better estimation, we further propose a difference eigenvalue edge indicator with threshold for more accurate block selection. After combining the noise estimation method, better estimation results of blur and noise are derived, leading to the better restoration results. Experiments show that the structure-based method is good for light noisy condition, while the difference eigenvalue based method is effective for fine texture images.