Military target recognition is a major research direction in the field of target detectionwhich is of great significance for early detection of enemy situation and accurate strike of the target.At presentthe mainstream military target recognition algorithms mainly use such unsupervised deep learning network models as YOLO and Fast RCNN.In these methodsthe manually-labeled datasets are inputted to the network and then processed to obtain the output.Howeverthe datasets have limited quantityunsatisfying accuracy and generalization ability.To solve these problemsthis paper proposes a military target recognition algorithm based on the unsupervised networkwhich effectively solves the problems of insufficient datasets and unsatisfying accuracy through Generative Adversarial Network (GAN).