Removing Speckle Noise by Analysis Dictionary Learning

Speckle noise inherently exists in images acquired by coherent systems, for example, synthetic aperture radar (SAR) and sonar images. Removal of speckle noise is a challenging problem because the noise multiplies (rather than adds to) the original image and it does not follow a Gaussian distribution. In this paper, we focus on the speckle noise removal problem and propose a method using analysis dictionary learning. In our proposed method, the image recovery is addressed in the logarithmic transform domain, thereby converting the multiplicative model to an additive model. Our formulation consists of a data fidelity term derived from the distribution of the speckle noise and a regularization term using the learned analysis dictionary. Experimental results on synthetic speckled images and real SAR images demonstrate the promising performance of the proposed method.