Data Inpainting Based On Generative Models

With the rapid development of data science and machine learning,in recent years,both industrial community and academia have shown a prosperity in data analyzing and modeling.In the real world,however,unbiased complete data are often too expensive to obtain,while the most of the data come with bias,loss and noise.In this paper,we address the issue of modeling the image data with different types of lossy measurements,including random block,dropout loss.Traditionally,there are two approaches to address the issue: the texture based and the partial differential equation(PDE)approaches.The former adopts an inpaiting algorithm to recursively restore the pixels from the loss margin,which is proved to be the most effective methods.Whereas the approaches require to iterate through all pixels in the images,the information of the restored pixels can only come from the original image.On the contrary,the latter approachs formulate the problem as the constraint optimizations of the PDEs,as well as significantly preserve the smoothing area and high frequent edges.The major deficiency is that it requires a heavy computational cost to solve the optimization of PDEs,and the connectivity in the original image might be altered during the optimization.Both of the two types cannot generalize well when a large area in image is biased.Traditional approaches cannot create new contents on images.However,generative adversarial networks(GAN)create new contents via learning features with unbiased historical data.Especially,GANs have good performances on absolute clean data.We use these partially masked images to predict the contents of missing parts by the idea of AmbientGAN.That it says,the approach produces the new information of these images.There is a good performance on the task of super-resolution.The shortcoming of the AmbientGAN is fully depending on lossy measurements.Once the parameters of lossy measurements is different from the real parameters,the results of inpainting is far from expectations.The discriminator directly distinguishes the naturalness of generative images and component images,instead of using real label of missing parts.