SAR despeckling based on clustering dictionary learning and sparse representation

Aiming at the speckle reduction of aperture radar synthetic (SAR) images, a method of SAR despeckling based on clustering dictionary learning and sparse representation is proposed. Based on the non-logarithmic model of the coherent speckle noise, the K-means clustering with the improved similarity measure and principal component analysis method, the dictionary atoms with structural clustering are obtained, which overcomes the effect of the non-Gaussian of the speckle noise. A sparse representation model combining clustering and sparsity under a unified framework is established. An iterative algorithm is proposed for solving the cost equation. Meanwhile, the point target protection measure is introduced into the algorithm to avoid the "over filtering" of the point target. Experimental results with SAR images from satellites and unmanned aerial vehicle show that compared with the existing SAR despeckling methods, the proposed method has a great improvement in both the visual effect and the objective evaluation indexes.