Research And Application Of Image Generation Based On Deep Learning

Generative image models are probabilistic models which model the distribution of natural images, and deep neural network can be regarded as a complicated function with high fitting capacity. We can use deep neural network to build generative model to es-timate the parameters of probability density function. Generative image models can be used for generating more exemplars of image, image restoration, transformation of differ-ent models of image or transformation of image and language, and predicting the future.Firstly, this paper introduces the definition of generative model in supervised learn-ing and unsupervised learning respectively, analyzes the reasons why we should learn generative models, and divides the prior works into several categories.Secondly, this paper introduces three deep generative models: deep belief net, vari-ational auto-encoder and generative adversarial network. Our methods are based on vari-ational auto-encoder.Thirdly, in order to learn interpretable representation to increase the controllability of image generation, this paper proposes a multitask variational auto-encoder to model the joint probability density function of face image and face contour to disentangle the lo-cation information of face images. We regard face attributes as some high-level variations of face images and specify a factored set of latent variables in variational auto-encoder to capture these variations utilizing the binary attribute labels. The proposed model can learn disentangled and interpretable representations of face images to control different interpretable parts of generated face images. Our model can also edit input face images to change some visual attributes.Lastly, this paper models face photo under condition of simple line drawing, and re-gards face visual attributes as a part of latent variables to control the face photo synthesis from simple line drawing. Experiments show that our model can synthesize detailed pho-torealistic face images with desired attributes. Regarding background and illumination as the style and human face as the content, we can also synthesize face photos with the target style of a style photo.