Trimmed Non-Local Means Filtering for the Suppression of Mixed Noise in Color Images

In this paper, a novel approach to the problem of mixed Gaussian and impulsive noise reduction in color images is proposed. The elaborated denoising framework is based on the Non-Local Means technique, which has been shown to effectively suppress Gaussian noise. To overcome the inability of this filter to deal with impulsive distortions, a robust similarity measure between image patches, that is insensitive to the impact of impulsive distortions, was developed. First, the distorted pixels of the image are detected using the concept of a peer group. If the number of neighbors of a pixel at the center of a filtering window, whose Euclidean distance in the RGB color space does not exceed a threshold value, is smaller than a predefined value, then the considered pixel is treated as corrupted. Afterwards, the similarity between image patches are calculated in a standard way omitting the noisy pixels. In this way, only the average of distances between pairs of unaffected pixels are calculated. Thus, the closeness measure of the patches, which serves as the coefficient in weighted averaging of pixels in the processing block, is not affected by impulsive noise and is still able to cope with the Gaussian noise component. The proposed filtering design is capable of reducing even strong mixed noise while preserving image details and edges. Moreover, it is not computationally intensive, so it can be used in real-time image processing scenarios.