Lightweight Generative Adversarial Networks Based on Ghost Module

Generative adversarial networks are widely used in computer vision tasks like image translation and image style transfer. Most of mainstream methods including CycleGAN and pix2pix use the stacking of residual blocks to deepen the number of network layers, which makes the networks have a large number of parameters and floating point operations. This paper presents a ghost-module-based generative adversarial networks. We use the ghost module to replace the residual blocks in the traditional generative adversarial network for building lightweight generative adversarial networks. Experiments shows that our method significantly reducing the parameters and floating point operations of the generative adversarial network on the precondition of assuring the quality of the generated images.