Abstract: The problem is tackled that the existing conditional generation adversarial network takes fixed position labels as input,generates cell samples without sufficient diversity,and the accuracy of cell detection cannot be improved to meet the actual detection needs.An approach of automatic generation of blood cells is proposed based on the generation adversarial network.The detection method generates a cell sample from random noise,and then uses the label generator to generate the cell image labels from both the generated cell sample and the real cell sample.At the same time,a multifunctional discriminator is introduced.The gener-ated cell images,labels and real cell images are fed into the multi-function discriminator in a one-to-one correspondence for matching and authenticating the cell position.Experiments on the blood cell dataset in-dicate that the generation of cell samples from random noise not only increases the diversity and quality,but also increases the IOU of cell detection from 90.3%to 91.1%,and increases precision from 91.7%to 94.8%compared with the mainstream method LCGAN.