Multi-domain Image-to-image Translation Based On Generative Adversarial Networks

The popularity of generative adversarial networks(GANs)boosts research on image-toimage translation.There are numerous of interesting scenarios,such as transforming facial attributes,mapping rough sketches to real photos,and changing the seasons of scenery images,etc.,which have received much research attention from the computer vision community.Although image-to-image translation shows an unprecedented rise in terms of research attention,most of the researchers focus on two-domain adaptation scenarios.Therefore,how to freely convert an image between multiple domains has great research significance.In this thesis,we propose a new method to address the problems of multi-domain imageto-image translation.Different from the traditional models for image-to-image translation,the model proposed in this paper takes the image and its target domain information as input,and is supplemented with an auxiliary classifier,so that the image can be freely translated between multiple target domains flexibly and efficiently.Therefore,the proposed model no longer needs to retrain the same network in order to learn a new mapping,and hence reduces the training time greatly.Futhermore,the self-attention mechanism is introduced to the proposed model,aiming to generate images with high-quality details and obtain a background consistent with the original image.The self-attention mechanism can model the long-range dependencies among the feature maps at all positions,which are not limited to the local image regions.Simultaneously,we take the advantage of batch normalization to reduce reconstruction error and generate finegrained texture details.In particular,we adopt spectral normalization in the network to stabilize the training procedure of the model.In order to verify the effectiveness of the proposed model,a large number of experiments have been conducted on the model in a large public dataset,with both qualitative evaluation and quantitative evaluation.Experimental results show that compared with other methods,the proposed model has better performance and generates images in higher quality.Therefore,the model proposed in this paper can be well applied to the multi-domain image-to image translation task.