Video Content Annotation Based On Multi-source Transfer Learning

Video content-based annotation can effectively classify and retrieve video,so it has been an important research topic.However,traditional machine learning method needs to manually label a large number of samples,which will consume a lot of manpower and material resources.Using the transfer learning method can well use the knowledge of the relevant domain to train the model,then complete the annotation of the video content.Because the source domain of single source domain transfer learning is relatively simple,the effect of transfer may be unsatisfactory due to the lack of correlation between the source domain and the target domain.Therefore,this thesis uses multi-source domain knowledge to train the target model.First,the Internet video knowledge and image knowledge are grouped.Then,the target classifier is learned by calculating the weight between the source domain and the target domain.Finally,the task of labeling the target video is completed.The main research work of this thesis is as follows:(1)For the negative transfer problem caused by missing annotation data for the target video,this thesis proposes a Multi-Source Adaptation(MSA).The method adds a small amount of labeled data to the target domain to calculate the weight of different source domain video groups according to the degree of correlation between the source domain Internet video group and the target domain user video.Then,the decision value of source domain video classifiers in the target domain video is calculated according to the weight,and finally the target classifier is trained according to the weight learning model based on the manifold regular term.The results of experiments show that the average label accuracy of MSA method on the Kodak database reaches 42.62%,which is 7.54%,4.59%,1.79%,and 9.99% higher than the CP-MDA,DAM,DSM and MDAHS methods.In the CCV database,it reaches 43.76% and has a relatively increase of 21.76%,17.38%,3.33% and 8.64%.(2)For the problem of insufficient extraction of source domain knowledge,this thesis proposes a Heterogeneous Compound Multi-Source Adaptation(HC-MSA).The method composes the source domain video group data and image group data into heterogeneous composite data sources.In order to reduce the mismatch of the data distribution between the source domain image group and the target domain,the maximum mean difference method is used to reduce the distance between the source domain image group and the target domain video.For each source domain group,learning the adaptive classifier and calculating its weight,then calculating the decision value of each source domain classifier in the target domain video according to the weight.and finally the multi-core learning and Laplaa manifold regular term are introduced to constrain the target classifier.The results of experiments show that the average label accuracy of HC-MSA method on the Kodak database reaches 47.32%,which is 14.69%,11.66%,6.41%,2.56% and 3.16% higher than the CP-MDA,DAM,DSM,MDA-HS and MSA methods.In the CCV database,it reaches 46.88% and has a relatively increase of 18.62%,14.17%,8.34%,3.47% and 2.87%.