Abstract: The internal image denoising methods emphasize on low-rank, sparse, and other prior information, and rarely consider multi-scale property. The external image denoising methods make full use of prior information from natural image, but it is difficult to estimate the similar block group’s rank properly. Aiming at these problems, this paper proposes a multi-scale low-rank denoising algorithm which combines the internal and external priors together. It considers noise images’ information among different scales of noise images and statistical distribution of external clean images. In the pre-training stage, statistical distribution of external natural image data sets has been learned as the prior information. In the grouping stage, the external prior information will be adapted to guide the noise image grouping, and the low-rank matrices also can be constructed simultaneously. In the implementation stage, multi-scale prior and the generalized nuclear norm have been integrated, and the proposed multi-scale low-rank denoising method has been used to reconstruct the target images. The method has been tested on Set5, Set12, Kodak, McMaster, and other classical image datasets. Comparisons with several state-of-the-art denoising methods have been given. Experimental results show that the method has obvious improvement in objective evaluation indicators, for example, the peak signal-to-noise ratio is 0.2 dB better than the comparison method, and at the same time, it can effectively preserve image details and textures in subjective visual effects.