Target recognition under complex background is the emphasis and difficulty of computer vision, and rotary objects is widely used in the military and manufacturing field. Rotary object recognition under complex background based on improved BP neural network is proposed in the dissertation, achieving the objectives of accurate extraction. Median filter is used to filter the image noise and an improved method of maximum class square error is used to compute the threshold of the image segmentation. The target recognition system based on improved BP neural network is established to recognize the rotary objects, and seven invariant moments of rotary objects serve as the input feature vector. The experiment results show that the image noise is reduced effectively and the image could be segmented exactly by the image preprocessing method putting forward in the dissertation, and the seven invariant moments is appropriate for the character of rotary objects, and the rotary object recognition system based on BP neural network achieves an excellent recognition result.