Research On The Generative Adversarial Models For Feature Learning And Image Enhancement

In deep learning,data samples such as natural images,audio signals and character symbols are considered to subject to special probability distributions.By modeling the probability distribution of these samples,the generative models directly generate samples which are similar to training examples.Researchers pay more attention to these generative models.Compared with the traditional generative models,the generative adversarial model does not depend on any prior assumptions.As a learning based sampling technique,it can synthesize variable kinds of data.In this thesis,we investigate the principle and application of the generative adversarial model.This thesis first introduces the application of the generative adversarial model in image enhancement and feature learning,then proves the ability of the model in fitting data distribution through theoretical analysis.Finally,we study the generative adversarial models in two image processing problems,that are image enhancement and image feature learning.In particular,two models are developed for image dehazing(image enhancement)and copy detection(feature learning),respectively.For image copy detection,we construct an adversarial network and a hashing network that compete with each other.Through the competition between the two networks,the hashing network generates robust and discriminative image features.On the other hand,following the idea of adversarial learning,we construct a discriminative network to distinguish between the generated image hash and the ideal hash sampled from the binomial distribution.This network guides the hashing network to generate the image features that follow the binomial distribution,so that the efficiency of copy detection can be improved.This thesis also proposes an attribute exchange and separation based image dehazing algorithm.The algorithm achieves dehazing by separating haze-relevant features from hazy image,and then these features are used to add haze effect on the clear images.To improve the quality of dehazed image,we use a discriminative network to distinguish between the real and generated images.In summary,this thesis applies the generative adversarial model in feature learning and image enhancement.We propose two generative adversarial network based image dehazing and copy detection algorithms.The proposed algorithms achieve better performance than the traditional ones.