Research On Semi-supervised Image Classification Based On Generative Adversarial Network

Image classification is a means of image processing method that classifies and recognizes images on the basis of the feature information extracted from them,and it is of great research value as a basic and key technology widely applied in the realm of computer vision.As the Internet technology making swift progress and the performance of computer hardware prominently bumping up,image data has emerged on a large scale,and the requirements for the accuracy of image recognition algorithms have been continuously improved.Therefore,traditional classification algorithms are gradually replaced by deep learning algorithms which can automatically learn advanced category features of images.However,there are much larger scale of unlabeled images than that of labeled images in practical application scenarios where it is difficult for deep neural networks to improve its performance with limited amount of label information.Considering the characteristics of semi-supervised image classification algorithms that it is able to learn the essential laws of images when there exist both labeled images in a small quantity and unlabeled ones in a large quantity,research on semi-supervised image classification based on generative adversarial network is conducted by this thesis.The main work is as follows:(1)In terms of the issue that the traditional generative adversarial network is difficult to characterize the importance of image features,a method of dual attention-based generative adversarial network for semi-supervised image classification is proposed.Firstly,the dual attention module is introduced into both the generator and the discriminator to extract the dependencies between local and global features as well as the dependencies between feature channels,so as to characterize the importance of task-related features,and then adaptively learn the features of images.Then,introduce spectral normalization into the generator to reduce gradient anomalies,thus enhancing the stability of model training.Finally,introduce the manifold regularization method into the discriminator loss to guide the direction of change in classification decisions.(2)Regarding the unmatch issue between the feature representation in the latent variable space of conventional generative adversarial network and the feature of real samples,a method of encoding transition-based generative adversarial network for semi-supervised image classification is proposed.Firstly,in order to increase the discriminative feature representation in the latent variable space,an encoder structure that maps the training samples from the sample space to the latent variable space to provide distribution information reference for the input noise variable of the generator.Then,the reparameterization method is adapted to describe the noise variable distribution,and the KL divergence is conducted to measure the difference between the input noise distribution and the latent variable distribution encoded from the real image to ensure the unity of the distribution of the input noise and the real images.Finally,in order to achieve the semantic matching between the images and the latent variables,the manifold agreement combination method is used to combine the image features and the latent variables,and the combinations are then input to the discriminator for authenticity discrimination and category classification.Experimental results on SVHN,CIFAR-10,and MNIST datasets indicate that the methods put forward in the thesis can make valid progress in enhancing the accuracy of semi-supervised image classification conducted by generative adversarial network.24 diagrams,22 tables as well as 116 references in total are included in this thesis.