Generating images with desired properties using the DiscoGAN model enhanced with repeated property construction

Abstract

The idea of image-to-image translation is to take advantage in certain areas such as adding the sharpness to images and improving the semantic segmentation. The most popular models for solving problems are generative adversarial network (GAN) [1] models such as DiscoGAN [2] and CycleGAN [3]. In training process, input images with no desired properties, and output images with the desired properties are fed into the generative model to train the model. After training, the model can synthesize the desired properties from the input images without those properties. However, in practical usage, an input image may be different from the training process because the input image may be the image with or without the desired properties. This research proposes the method of training the generative model by giving input images with and without desired properties in the same way as when the model is used. Our proposed model enhances DiscoGAN with repeated property construction to generate images with desired properties. The model can use unpaired data as the training data, which makes data preparation more efficiently and more comprehensive than paired data. The proposed model obtained approximately 8% better Frechet Inception Distance (FID) [4] score compared to the DiscoGAN model.