LGEGAN: A Lightweight Evolutionary Generative Adversarial Network with Statistic Global Information

Generative Adversarial Networks (GANs) have been applied in many fields. However, the existing GANs and their variants encounter many problems, including pattern collapse, training instability and falling into local optimal. Therefore, we construct a lightweight evolutionary generative adversarial network (LGEGAN) with statistic global information. To solve the problem that shallow convolutional neural networks are difficult to capture long-range feature dependence and the training process is prone to pattern collapse, LGEGAN is different from EGAN in that we add an improved self-attention mechanism to the generator network. To solve the problem of instability in the training process, we add spectral normalization to LGEGAN, which increases the stability of each generation of training process. Finally, to efficiently evolve individuals who adapt to the environment in a short time and solve the problem of falling into local optimal, different from EGAN, we construct a novel selection operator and apply it to the selection stage of LGEGAN's evolution of the generator. In the experiments, we evaluate the LGEGAN from four aspects: the quality and diversity of generated images, the collapse of patterns, the stability of training and the robustness of the architecture. The experimental results indicate that the performance of LGEGAN is better than that of EGAN, MOEGAN, SMOEGAN, LRGAN, ProbGAN and the other generative adversarial network models.