Image Classification Based On Semi-supervised Generative Adversarial Network

Image classification is one of the important research hot spots in computer vision and pattern recognition.According to whether training samples are labeled,the classification algorithm can be divided into supervised learning and unsupervised learning.The cost of labels in supervised learning is very large,while the unsupervised learning has inferior classification performance.Therefore,semi-supervised learning,which could obtain superior classification performance with little labeled samples,has become a research hot spot in current classification algorithms.As a new type of deep learning algorithm,generative adversarial network(GAN)are widely applied to the image generation,image restoration and image classification field.At present,the existing semi-supervised classification algorithms can not learn the relationship between labeled samples and unlabeled samples well,and the semi-supervised classification accuracy needs to be improved.Therefore,in case of the above insufficiency,this paper proposes two image classification using semi-supervised generation adversarial network models.The main research contents are as follows:First,the training process of regular GAN is not stable enough and GAN is prone to mode collapse.The robustness of the features extracted by the discriminator is poor.Aiming at this problem,this paper proposes a piecewise loss weighted generative adversarial network(PL-GAN).On the one hand,generator adopt different forms of loss function in different training stages.To achieve piecewise training purpose throughout the training,the generator combine two different forms of loss functions by time parameters;On the other hand,generator can be introduced the feature-level loss,which the mean square error between the real sample and the generated sample,and then weighted against the regular GAN generator's adversarial loss,which can effectively improve the regular GAN training instability and solve the mode collapse problems,result in the features extracted discriminator are more robust.Secondly,it is difficult to extract the task-related robust features for the traditional discriminator's loss strategy and structural framework,and a feature recalibration generative adversarial network(FR-GAN)is proposed.On the one hand,based on the existing semi-supervised GAN,the discriminator introduces the unsupervised mean square error loss regularization term in different model states.On the other hand,discriminator penalized different predictions for the same input to guide the direction of the feature calibration.In order to increase the nonlinear fitting ability of the network,a SENet module is added to the structure of the traditional discriminator.As result,the weighted vector of each input feature map learned by SENet,are weighted with the input feature map to achieve recalibration of features.When some labels are wrong in proportion,the classification accuracy of the FR-GAN does not change much,that is,it improves the fault tolerance of the network.The experimental results of the MNIST,CIFAR-10,SVHN,and STL-10 datasets show that the proposed algorithm can obtain better classification performance compared with the traditional semi-supervised classification algorithm.