Image filtering plays a very important role in remote sensed image processing, especially in relation to further discrimination of ground objects, image segmentation and classification processing. Anisotropic diffusion has been employed as an effective tool for image filtering. While a great deal of effort has been dedicated to the behavioral analysis of this technique, many researchers have discussed its extension to multispectral images. However, these efforts still left some difficulties unsolved. In order to overcome these drawbacks, an anisotropic diffusion nonlinear filtering algorithm is presented. In a first step, a local Gaussian variance selection is developed. The construction of an automatic and convergent strategy based on the nonlinear time-dependent cooling technique is then built for the gradient threshold. Finally, we established a new version of multispectral anisotropic diffusion incorporated with the above two improvements by recurring to the relationship between robust statistics and anisotropic diffusion. Experimental results on two satellite images are shown that the proposed algorithm not only effectively removes the impulsive noise caused by sensors, but also preferably preserves important edges and image quality.
展开▼
