Research On Image Classification Algorithm Based On Semi-supervised Learning

Semi-supervised learning is indispensable in the field of artificial intelligence.Under the circumstance of lack of labeled data,lots of unlabeled data serve as auxiliary component,which effectively solve the problems such as the difficulty to obtain labeled data.Therefore,it has important application value in real-world scenario.However,in semi-supervised learning algorithms,it is not cost-effective to produce labeled data.Therefore,it may cause overfitting and effect to generalization performance of model.Moreover,the incorrectness of pseudo labels is inevitably,so it will do harm to the whole performance of model.Therefore,this paper is based on ensemble deep network,the semi-supervised classification algorithm based on weighted pseudo labeled data and mutual learning and the semisupervised classification algorithm based on decision fusion and consistency regularization are proposed.The main research is as follows:1.To prevent overfitting,a semi-supervised classification algorithm combined with disagreement-based decision fusion is proposed.First,the labeled data are enlarged by mixup data augmentation,which are used to initialize the model.Then the pseudo labels and their confidences are inferred by network predictions.Second,based on decision fusion strategy,the pseudo labeled data are divided into data with high confidence and data with low confidence.The data with low confidence are added to unlabeled dataset and are data with high confidence are added to labeled dataset.The weight of each sample is calculated accordingly.Last,based on mixup consistency regularization method,each classifier is trained with weight and consistency regularization.It can be seen from the results on MNIST,CIFAR10 and SVHN that the proposed algorithm prevents overfitting and perform well.Notably,the method has a higher accuracy than others recent years.2.As for the methods that are based on pseudo labels,the performance of model is limited in the early stage of model,which causes the incorrectness of pseudo labels.If they are not rectified,the performance of model may be damaged.To rectify incorrect pseudo labels,a new algorithm based on weighted pseudo labeled data and mutual learning is proposed.The model is built on a deep ensemble network.First,output smearing is employed to construct different training sets and perform model initialization.The pseudo labels of unlabeled data are inferred by network predictions.Second,based on selection and weighting strategies for pseudo labeled data,pseudo labeled data with high confidence are moved together with real labeled data.Accordingly,the model is retrained on the weighted pseudo labeled data.Last,a mutual learning strategy is applied to enhance the prediction consistency among classifiers.Furthermore,diversity fine tuning and mutual learning are performed alternately to determine the optimal balance between diversity and consistency,which consequently improves the accuracy of the pseudo label predictions.It can be seen from the experimental consequences on MNIST,CIFAR10 and SVHN that the algorithm rectifies the incorrect pseudo labels.Notably,it do better than other algorithms.