L1 norm based double-prior learning for image denoising

Image denoising problem has attracted a large number of researchers. Generally speaking, there are two kinds of image priors considering the source of training sets, external priors and internal priors. The realistic image priors can be obtained from a large number of external example images or the corrupted internal image itself. However, external priors cannot give accurate image representations towards various corrupted images because the total number of example images is limited. While internal priors may bring too much noise along with useful information for denoising, which leads to unexpected denoising results. The most common assumption in denoising problem is that the image noise obeys Gaussian distribution, which is simple and ideal. If there are outliers in the corrupted images, Laplace distribution is more suitable to model the image noise. This paper proposes a denoising model towards image noise in Laplace distribution utilizing both external priors and internal priors. Gaussian Mixture Model (GMM) is used to model external priors and l1 norm is aimed to deal with outliers. Experiments on some publicly available databases show the performance of proposed method, resulting in denoised image of high quality.