NSPSO-GAN: Generative Adversarial Networks Based on Particle Swarm Optimization with Natural Selection for Stable Image Generation

Although generative adversarial networks (GAN) has demonstrated remarkable potential in image generation, there are still problems such as mode collapse or difficulty in training convergence that limit the stable generation of images. In this paper, we proposes a novel approach, called NSPSO-GAN, to enhance the training efficiency of GANs by optimizing weight initialization in the generator. The methodology employs a nature selection-based particle swarm optimization (NSPSO) to determine optimal initial parameters before commencing the adversarial training process. Compared to traditional random weight initialization, this strategy demonstrates to improve convergence speed while simultaneously achieve superior image generation quality, as quantified by Fréchet inception distance (FID), peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) in MNIST, CIFAR10 and Celeb. The experimental results indicate that NSPSO-based parameter initialization effectively addresses common GAN training challenges, maintaining the stable image generation.