Anisotropic diffusion for image filtering based on MDL and morphology

Summary form only given. Anisotropic diffusion (AD) has been employed as an effective approach for image filtering. A great deal of work has been dedicated to the improvements of this technique. However, these efforts still left some difficulties unsolved. The first one is to estimate accurately image gradients in noisy images for nonlinear diffusion implementation; the second one is to design adaptively a setting of the gradient threshold in the diffusion coefficient. In order to overcome these limitations, a nonlinear AD algorithm is presented. In a first step, a time-dependent local Gaussian scale selection based on minimal description length criterion (MDL) is developed, smoothing an image by convolving appropriately with the locally optimal Gaussian scale chosen to better recognize the image contour gradients. An adaptive strategy based on morphology is then built for the setting of gradient threshold according to local noise characteristics. Using the relationship between robust statistics and AD, we established the new version of the diffusion coefficient based on the above improvements. The proposed method not only effectively removes Gaussian noise and impulsive noise caused by sensors, but also preferably preserves important detailed edges and image quality. Experimental results are given to show that the proposed method has superiority capability over the previous AD schemes on visual judgment and quality statistical analysis.