Noise in fundus images seriously affects the diagnosing of lesions. On the basis of prior knowledge about fundus images, four multiscale geometric transforms are adopted for de-nosing, which are respectively Ridgelet, Ridgelet combined with Wiener filer, Wavelet,Contourlet. The Local Mean and Local Standard Deviation algorithm is given based on Gaussian wave extraction, which is used to estimate Signal-to-Noise Ratio(SNR) of the processed images and give an objective quantitative evaluation to treatment effect of the above de-noising algorithms. The results show that the image processed by Ridgelet combined with Wiener filer is the most clearest, and its SNR improves the most obviously, about 5.04 times compared with the original image. The results of objective quantitative evaluation are in accordance with subjective visual feeling.