In order to detect small target among dense star background, an improved difference target detection method is proposed. Firstly, the registration between two consequence frames is acquired by use information of telescope’s encoder, position and direction. Then previous frame is dilated when the difference, which can reduce error caused by atmospheric disturbances and registration rounding. Secondly, after suppression of background star edge as much as possible, candidate target is obtained by segmenting, target clustering, and labeling of differenced image. Residual background star edges are removed by using the position of candidate targets to remove star edges which not get rid of in previous procedures. Finally, double stars are removed by information of the amount of registration. After a series of pretreatment, dense background star can be removed efficiently by this method. Experimental results show that candidate targets could be kept within 20 by this proposed algorithm, for 2 k×2 k, clustering star background image.