A Non Local Feature-Preserving Strategy for Image Denoising

In this paper, we propose a variational image denoising model by exploiting an adaptive featurepreserving strategy which is derived from the Non-local means (NL-means) denoising approach. The commonly used NL-means filter is not optimal for noisy images containing small features of interest since image noise always makes it difficult to estimate the correct coefficients for averaging, leading to over-smoothing and other artifacts. We address this problem by a non-local detail preserving constraint, which is performed by adding two terms in the Total variation (TV) model. One is a non local patch based regularization term that controls the amount of denoising to preserve textures, small details, or global information, the other is a new data fidelity term, which forces the gradients of desired image being close to the smoothed normal. The Euler-Lagrange equation is used to solve the problem. Experimental results show that the proposed method can alleviate the over-smoothing effect and other artifacts, while preserving the fine details.