Combining feature pyramid and CycleGAN for image generation

Image generation has always been a hot topic in computer vision community, which aims to learn the data distribution from a give image dataset and then generates new images obeying this distribution. Thanks to the rapid development of convolutional neural networks, breakthroughs have been made in the accuracy and speed of image generation. Currently, image generation is mainly based on the framework for generating countermeasures networks. However, limited by the quality of features, the generated images still have problems such as edge blurring, which restricts large-scale practical applications. In this paper, we propose an image generation algorithm that combines feature pyramid network (FPN) and CycleGAN. Specifically, the FPN containing two upsampling and feature fusion operations is added after the residual blocks and before the decoder of the CycleGAN generator, which can help the generator produce more detailed images and learn from features at different scales. Qualitative and quantitative experimental results show that the CycleGAN model with FPN outperforms the original CycleGAN model in terms of image generation.