In this paper a FIR nonlinear fuzzy filter for image processing, which is most effective in removal of mixed noise, is proposed. In general it's hard to distinguish noise and edges information. This ambiguity leads us to use fuzzy concepts. Fuzzy similarity is used here to suppress noise and preserve edges. Parameters of the membership function are optimized by genetic algorithm approach Since our problem here is stochastic, traditional optimization algorithms are of no use anymore. Simulation consists of some combination of Gaussian and salt and pepper noises on different classes of images. Results are compared with traditional median and Wiener filters both from subjective and objective points of view.