Image Style Transfer Based on Feature Contrast in Cyclic Generative Adversarial Networks

Unsupervised image-to-image translation is a task of learning the transformation between source and target domain images without paired training data. But Image stylist transfer is another challenging task and face the problem of content loss and model collapse. In order to solve these problems, this paper presents a novel local feature contrast method that can keep the content of images. The image encoder learns richer representations by extracting multi-layer deep features from the images using a feature extractor. Meanwhile, we introduce a local feature contrast loss for guiding the content generation favorable features learning. Experimental results show that proposed approach gives better [FID and KID] scores in comparison to the previous works, which eventually improves image generation ability.