Research On Semi-supervised Image Classification Based On Collaborative Training And Active Learning

With the advent of the "big data" era,all kinds of data are exploding and in these data,multimedia data represented by images occupy a huge proportion.Therefore how to extract effective and usable information from these huge multimedia data become a widely studied issue.Since the 1970 s,the field of image retrieval has gradually attracted people's attention,but most of the data are just simple data,which requires manual labeling and classification and consumes a lot of manpower and material resources.This greatly limits the development of image retrieval technology.In order to improve this dilemma,automatic image classification technology attracts extensive attention.Support vector machine(SVM)and deep learning are widely used in image classification technology,and have achieved good research results.However,there is a problem with both methods: the small sample problem,that is,the classification performance of the algorithm is very limited when the labeled sample size is small.To address this problem,this paper considers introducing semi-supervised ideas into image classification research,and proposes a semi-supervised image classification algorithm based on active learning and collaborative training.This paper first proposes a semi-supervised image classification model based on label propagation and active learning.This model uses SVM as the base classifier,uses the output results of SVM to select appropriate samples,and uses label propagation to perform label prediction and then joins the active learning process to optimize the classification performance of SVM.Then,this paper proposes a semi-supervised ladder network model based on a convolutional neural network.The convolutional neural network is introduced into a trapezoidal network to better extract image features,speed up model training,and optimize the classification effect.Finally,the adaptive weighted fusion strategy is used to select appropriate samplesfor collaborative training of the above two models,which further improves the classification performance of the classifier.Experiments show that when the number of labeled samples is small,the classifier can effectively use the information of unlabeled samples,which has better classification performance than some existing semi-supervised classification methods.