Multi-label Learning Based On Transfer Learning And Label Correlation

A sample has multiple labels at the same time in multi-label learning,such as a picture with both "ocean" and "sailboat" labels.However,in the existing multi-label learning method,most methods only consider the dependencies between the labels in the data set(local label correlation)or only the semantic similarity between the labels(global label correlation).In fact,in the multi-label learning,samples have multiple labels and both local and global label correlations may appear in real-world situation.In addition,we should not be limited to pairwise labels while ignoring the high-order label correlation.Therefore,we propose a novel and effective V-GLLCBN(V3-Global and Local Label Correlation Bayesian Network)multi-label learning method by transferring Inception V3 model combined with local and global label correlation network model(GLLCBN)in this paper.The main achievements of the research are as follows:(1)A model is proposed by using transfer learning to transfer the Inception V3 model that implements image classification into a multi-label image learning.As we know,the key point in solving the classification problem lies in the performance of the classification model,the Inception V3 model is a better performing deep neural network model,it is obtained by Google for large-scale image classification.Firstly,this paper uses the model transfer method to transfer the structure before the convolution module group of the Inception V3 model.Secondly,the convolution layer and linear structure models after constructing the bottleneck layer are customized by using the asymmetric structure instead of the symmetric structure.Then that,defining an output layer by analyzing attributes of the labels set in the data set.Finally,the transfer model is combined with a custom classification model to generate a novel multi-label image learning model,and the rapidity and effectiveness of the model is verified by experiments.(2)A novel and effective local and global label correlation multi-label learning model(GLLCBN)is proposed.For multi-label classification learning,label correlation is often overlooked,but label correlation has potential value.First of all,we analyze the pairwise labels in the label space of the data set to acquire the local correlation and obtains the dependence of pairwise labels.Secondly,we obtain the global label correlation and build a global correlation matrix by exploiting semanticsimilarity of labels.And then,a initial label correlation topology model was build by analyzing local and global label correlations.Next,using graph theory and probability theory to eliminate redundant dependency structure in the initial model,so as to get the optimal label dependent model called GLLCBN(model nodes represent individual labels,and edges represent the comprehensive correlation between labels).Last but not least,achieving multi-label learning by injecting the label of the label input source into the GLLBCN model one by one.(3)Through the combination of the proposed models in steps(1)and(2),this paper proposed V-GLLCBN method for implementing multi-label pictures,in which the(1)model is used as label input source of the(2)model.Based on the performance evaluation criteria of multi-label learning,the stability and effectiveness of the proposed method are verified by experimental design and results analysis.