Research On Image Recognition With Semi-supervised Classification Algorithm Based On Generative Adversarial Networks

In recent years,it has achieved great improvement based on semi-supervised image classification and saved a lot of labor for manual labeling,especially in the generative adversarial network applications and semi-supervised classification.The semi-supervised generative adversarial network proposed by Salimans combines the discriminator and the classifier into a network.For labeled data,the discriminator can classify it with gradient descent like a classifier,however,the discriminator can only distinguish its true or false for unlabeled data,and can not truly classify the unlabeled data;and the image discriminator that generates better quality for generating data only discriminates its authenticity.It is not allowed to participate in the classification task and thus waste resources.Therefore,in view of the above shortcomings,two kinds of semi-supervised image classification networks based on generating confrontation networks are proposed.The main research contents are as follows:First,the traditional semi-supervised generative adversarial network(Improved GAN)has a problem that the classification accuracy is too low due to too few labeled data,and there is a problem of difficulty in training and slow convergence in the training process.Aiming at this problem,an improved GAN with Shannon Entropy(SE-Improved GAN)based on the information entropy loss function is proposed.By using the information entropy loss for the Improved GAN loss function,a large amount of unlabeled data is utilized.Information entropy provides feedback to the classifier to enhance classifier classification ability and classification convergence speed.Secondly,in order to improve the classification accuracy by using ACGAN to generate high quality image samples,this paper proposes a semi-supervised ACGAN model-SAC-GAN(Semi-supervised Auxiliary Classifier-GAN,SAC-GAN),because ACGAN is a kind of The supervised generation model,for the labeled and generated data,because there are corresponding labels,so the data can be mapped to the corresponding labels through the classifier,thereby improving the classifier's classification ability,but adding the pair to ACGAN Label data classification,through the Shannon entropy to increase the penalty item,you can classify the labeled data,the unlabeled data,and the generated data to achieve a better classification effect.The two semi-supervised classification models proposed in this paper have completed the image classification experiments in the standard handwriting recognition database MNIST,the street number database SVHN and the natural image database CIFAR-10,and compared with other semi-supervised models.The algorithm of the proposed model can obtain better classification performance.