Abstract: In this paper a improved nearest neighbor（INN）algorithm is proposed, it consists of three steps:1.editing the primary training sample set;2.finding the minimal consistent subset of the edited subset with the help of the dirichlet tesselation （DT） of the edited subset;3.designing the nearest neighbor（NN） classifier using the minimal consistent subset.The storage reguirement for INN algorithm is much less than the storage requirement for NN algorithm and its performance is batter than the NN algorithm. An algorithm of computing the DT of point set in euclidean space is presented and proved. The INN algorithm is extended to some of non-euclidean metric space. In table 1., the results of Bayes classifier are compared with those of application INN algorithm to maha-lanobis distance, City-Block distance, Chebyshev distance and Euclidean distance As an example of application, the INN algorithm is applied, to recognize two categories of radar targets-tank and jeep, by means of computer simulation, the results are shown in Table2.