<p>Nonlocal self-similarity (NSS) property of natural image has been widely used as an effective prior in the image denoising task. However, most of the existing NSS-based denoising models exploit the NSS prior within a square window, resulting in slow denoising processes and suboptimal results. In order to solve the shortcomings of the existing method of building a denoising model by constructing a non-local self-similar group through square window search, a two-stage image denoising algorithm based on superpixel non-local groups is proposed. In the first stage, the noisy image is segmented into superpixels; then, similar patch groups are constructed within the superpixel regions of the image; after that, the singular value decomposition (SVD) and hard thresholding techniques are used to eliminate most of the noise; finally, the inverse SVD transform is applied to generate filter groups and the preliminary denoised result image is obtained by aggregation. In the second stage, similar patch groups are first obtained by a method similar to the first stage; then Wiener filtering is adopted to eliminate residual noise; finally, the final denoised result image is obtained by aggregation. Extensive denoising experiments on grayscale and color image datasets show that the proposed algorithm outperforms multiple image denoising algorithms in terms of both objective evaluation metrics and subjective visual perception.</p>