Spatially adaptive Total Variation image denoising under salt and pepper noise

Automated selection of the regularization parameter for Total Variation restoration has shown to give very accurate reconstruction results. Most of the literature is devoted to the ℓ²-TV case (images corrupted with Gaussian noise), whereas for the ℓ¹-TV case (images corrupted with salt-and-pepper noise) there are only a couple of published algorithms. In this paper we present a computationally efficient algorithm for ℓ¹-TV denoising of grayscale and color images, which spatially adapts its regularization parameter. The proposed algorithm, which is based on the Iteratively Reweighted Norm algorithm, uses an adaptive median filter to initially estimate the outliers of the noisy (observed) image, and then proceeds to solve the ℓ¹-TV problem only for the noisy pixels while spatially adapts the regularization parameter based on local statistics. The experimental results show that the proposed method yields impressive results even when 90% of the image pixels are corrupted.