Image denoising plays an important role in various image-related applications. While serials of wavelet-based denoising schemes fit well to images with Gaussian (white) noise

few of them can handle images with various non-Gaussian noises effectively. This paper deals with the problem from the data mining approach. It treats noisy pixels in an image as isolating outliers that are discernible in color attributes from their neighbor pixels. Inspired by the idea from outlier detection analysis

it first maps the pixels of the image into a metric space and then introduces a distance among the pixels. By making use of the density function on the pixel data set

it formulates an analytical definition of the noisy point. Further

the paper discusses the properties of the non-noisy points and constructs a denoising algorithm. Results of experiments and real world applications show that this novel approach is effective both to Gaussian and non-Gaussian noise. The method can be implemented for mass image denoising with satisfactory efficiency and denoising quality.