A deep convolutional generative adversarial network algorithm based on improved Fisher's criterion (FDCGAN, deep convolutional generative adversarial network algorithm based on improved Fisher's criterion) is proposed to address the problem that the quality of generated images decreases dramatically when the training sample size is insufficient or the number of iterations decreases. The method adds a linear layer to the discriminative model, which extracts the category information. Fisher's constraint criterion is used in back propagation to combine label and category information. The weights are adjusted iteratively to minimize the error while keeping the intra-class distance small and inter-class distance large so that the weights can approach the optimal value more quickly. By comparing the experiments with the latest six different network models, the FDCGAN model achieves better results in all the FID metrics. In addition, the experimental results all achieve better results by applying the method to the current advanced models for generalization tests.