Noise-regression GAN for Image Inpainting and Multiple Generation
Guokai Fu, Chen Diao, Wanli Xue, Shengyong Chen · 2020
Image inpainting is a challengeing task for machine, recent years the emergency of deep learning and generative adversarial networks(GAN) shows that such task can besolved perfectly, however most approaches tend to generate only one plausible result for each input. in this paper, we propose a new model which can not only generate plausible image but also get multiple generation it seems that our method give creativity and imagination to machine. the model is based on GAN and convolutional neural network(CNN) by add noise to the input and build the mapping between ground truth and noise, our model eventually learns how to generate multiple plausible images, experiment on CelebA(a large dataset that contains 202599 images of celebrities) proves our model generates high quality results.