Research On The Generative Models For Visual Feature Analysis

As the visual representation of outside world,image is one of the most common information carriers.Image processing has many applications in the fields of aerospace and industry test,etc.in which visual feature extraction plays a fundamental role and has attracted extensive research interests.Generative model can capture intrinsic characteristics of natural images by learning the probabilistic distribution of training set and is superior to hand-crafted feature extractors.Traditional probabilistic generative model fits the statistical distribution of data via an explicit function.While generative adversarial network implicitly maps a random noise vector to the space of data and is able to synthesize high quality images.Both types of models are of great implications in theoretical study and practice.This thesis first elaborates the probabilistic generative models,and the emphasis is placed on the product of experts model as well as the solutions.Since generative adversarial network is based on deep learning,this thesis also introduces these two theories detailedly as the basis of following research.For image copyright protection problem,this thesis proposes a robust content fingerprinting algorithm based on hierarchical generative model.The algorithm maps an input image to a short descriptor via a group of expert functions.Application-specific constraints are designed and added to the objective function to boost the discriminability and robustness of descriptor.The proposed algorithm utilizes mean pooling operation to remove the redundancy in feature map,and the pooling operation is cascaded with expert functions as a feature extractor.Sequentially learned extractors are stacked,each of which learns features from the output of the former layer.The algorithm achieves high accuracy in copy detection and outperforms other comparative algorithms.As for image retrieval problem,this thesis proposes a semantic hashing algorithm based on generative adversarial network,where the discriminator is formed by an autoencoder.The encoder is shared by the adversary stream as well as the stream for extracting semantic information and class label.The proposed hashing framework is trained in a supervised way.Since the generator can transform random noise into synthesize image which is extremely similar to the real one,it helps to augment training dataset.Experimental results on two datasets MNIST and CIFAR-10 demonstrate the excellence of proposed algorithm.In summary,this thesis investigates probabilistic generative model and generative adversarial network for feature engineering,and satisfactory results have been observed in both image copyright protection and retrieval.