A Perturbed Convolutional Layer and Gradient Normalization based Generative Adversarial Network

This paper develops Perturbed Convolutional layers and Penalty Gradient Normalization based Generative Adversarial Network (PCPGN-GAN) to address the issues of training instability caused by sharp gradient space and the discriminator’s memorization problem. The distinctive aspect of the perturbation convolutional layer lies in its departure from traditional convolutional layer. It accomplishes this by introducing random perturbations to the input tensor before conducting convolution operations. This modification significantly reduces the discriminator’s inclination toward memorization issue. Moreover, the introduction of the penalty gradient norm constraint is exclusive to the discriminator function. This constraint enhances its capacity while stabilizing gradient information, thereby improving training stability. Comparative evaluations across multiple datasets, including CIFAR-10, STL-10, Tiny ImageNet, CelebA-HQ and LSUN-Church, highlight the superior performance of the PCPGN-GAN. Experimental results indicate its considerable advantages, manifesting in higher image quality and surpassing existed methods in terms of Inception Score and Frechet Inception Distance.