To remove mixed noise better, which is composed of impulse noise and Gaussian noise, a mixed noise removal method based on non-local means filter and Small Univalue Segment Assimilating Nucleus(SUSAN) operator was proposed. Firstly, according to the difference of shape features of univalue segment assimilating nucleus between the impulse noise points and conner points, the method extracts many characteristic points by SUSAN operator, which contains mainly impulse noise points and may also contain a small part of corner points. The characteristic points are sorted. The points whose frequency of occurrence ranks top two are impulse noise points. Then, the mean of non-impulse noise points in the pulse noise point′s neighborhood is calculated by the improved mean filter method to replace the impulse noise points′s gray value. Finally, for the image whose impulse noise has been filtered, the remaining Gaussian noise is denoised by non-local means filter method. A large number of denoising experimental results show that, compared with the method combining adaptive median with weighted mean, the method combining median filter with wavelet transform, and the method combining pulse coupled neural network with median filter, the proposed method has better subjective visual effect, and preserves the edges and details of image better, and it has great improvement in objective quantitative evaluation indicators such as peak signal to noise ratio. The proposed method shows obvious advantages in removing mixed noise.