Abstract: The microscopic images from leaf slices of Caragana stenophylla Pojark are inevitably corrupted by noise in obtaining, which will have a negative effect on its subsequent processes. The unpitched sound is usually treated as an additive Gaussian sound in order to denoise the natural images. It has not yet been known whether the unpitched sound and the additive Gaussian sound display the same distribution. The choice of the unpitched sound model will determine the estimation on its noise level and its subsequent choice of the denoising algorithm. BM3D (block-matching and three-dimensional filtering) is universally regarded as the best denoising algorithm, since it keeps the maximum information about texture, border, and other details of the image while denoising effectively. But the speed of BM3D is limited owing to too much calculation and its practical use is also limited because it needs the noise variation as a pretested input. SVD (singular valuable decomposition) is one of the best denoising algorithms but the noise of the image and its information can not be completely separated. The type of image, the level of the noise, and the change of color all affect the validity of the algorithm. This paper proposed a process, involving the estimation on the image type, the noise level and the algorithm of denoising, to solve the problems mentioned above and to denoise the microscopic images from leaf slices of Caragana stenophylla Pojark. The noise type was recognized as Gaussian noise by the rebuilding and fitting of the smooth histogram domain. The values, abstracted from the smooth histogram domain of 15 randomly selected microscopic images from leaf slices of Caragana stenophylla Pojark, were fitted with the Gaussian function, with the R2 value of 0.996 and the RMSE (root mean squared error) value of 0.144 6. And then the standard deviation of the noise was estimated via the estimation on noise level with the help of SVD domain image block. In the algorithm estimation, an initial estimative value was obtained from the SVD domain image block with the smallest entropy, and a more accurate value was got from its subsequent self-adapting correcting of the algorithm values. A two-step simulation experiment proved that this algorithm was much better in speed and accuracy in denoising than the other two estimation algorithms. When the standard deviation of noise was lower than 10, the proposed algorithm had the best accuracy; when the standard deviation of noise was higher than 10, this algorithm had the same accuracy with the estimation method of sliced image noise level, but had a better accuracy than SVD. This paper got a standard deviation of the image noise of between 2.5 and 4.0 for its estimation on the microscopic images from leaf slices of Caragana stenophylla Pojark. The image was denoised via GMF-BM3D, a joint algorithm by GMF (geometric mean filtering) and BM3D. GMF was used to denoise the background of images and then BM3D was applied to denoise the target area. Experiments proved the algorithm introduced in this paper was good at denoising and remaining detailed information of the image, such as border and texture, with a much greater speed. The average value of the BRISQUE (blind/referenceless image spatial quality evaluator) of the processed image was 10, equivalent to half of that of the original image. The algorithm also had an equivalent effect in denoising with the traditional BM3D, but eight ninths of time was saved. It was better than WT (wavelet threshold) denoising in BRISQUE, which was lowered by about 4. Therefore, the algorithm in this paper accurately denoises the microscopic images from leaf slices of Caragana stenophylla Pojark and provides a reliable technical support for its subsequent image processing.