It is well known that multiple observed signals are required for image denoising with ICA(Independent Component Analysis). In this paper, a method that multiple observations were generated by making the reduntant imformation of a single image sparse was presented. Firstly, made the only one noisy image to be sparse by using the dictionary compression algorithm of K-SVD (Kernel Singular Value Decomposition). Secondly, obtained the first-time denoised image by using the redundant information. Finally, made both the first-time denoised image and original noisy image as the multiple observations for ICA separation. It could be seen that the sparse image obtained by proposed method was more exact than that by using only a dictionary compression algorithm of Nlmeans (Non-Local means). The result obtained shows that when the Gauss white noise's standard deviation σ is in the range of 20-45, the proposed method is better than either K-SVD algorithm or NLmeans algorithm, and the denoised image's PSNR (peak signal to noise ratio) is 1.4 times larger than that of the original noisy image.