Kinds of synthetic aperture radar (SAR) image filter methods are studied first, their faults being pointed out. Then a kind of SAR image filter method based on annealing and iterated filter is proposed, which describes image local property by correlated neighbourhood model. At first, a series of templates reflecting direction informatio are established and every template is present for a kind of neighbour structure. Then on the basis of SAR image statistical property, the MAP estimate of the real intensity under observation image values for every structure is got by Bayes formulation. Simulated annealing algorithm is used to get proper structure from the candidate structure according their probability, the estimate value corresponding to this structure is the estimate of the underlying cross-section. Simulated annealing algorithm makes sure that the best global result can be obtained when the temperature is reduced to zero.