Popular image generation based on popularity measures by generative adversarial networks

We study image-to-image translation and synthetic image generation. There is still no developed model to create popular synthetic images based on the user's opinion in the fashion industry. This paper uses a combination of generative adversarial networks (GAN), deep learning, and user's opinions to create popular images. Our proposed model consists of two modules; one is a popularity module that estimates the intrinsic popularity of images without considering the effects of non-visual factors. The second one is a translation module that converts unpopular images into popular ones. Our model also performs multi-dimensional translation and multi-domain translation. We use the ResNet50 neural network as the default deep neural network in which the last layer is replaced with a fully connected layer. We use a new dataset collected from Instagram to train our network. We evaluate the performance of the proposed method by FID, LPIPS scores, and popularity index in different scenarios. The results show that our proposed method shows at least 60% and 25% improvement in terms of FID and LPIPS in color-to-color image translation. These improvements confirm the proposed method's generated images' quality and diversity. The evaluations on the popularity score also confirms that the content-based translation is more effective than style-based translation in terms of popularity.