The Research Of Multi-instance Multi-label Based On Neural Network

With the high-speed development of information age,the amount of information data has grown exponentially,so there is an urgent need to live and work efficient management of mass data,how to accurately classify data has become a difficult problem to be overcome by scientists.To solve this problem,scholars have put forward the traditional supervised learning framework,multi-instance learning framework,multi-label learning framework and multi-instance multi-label learning framework.Since the real data contains complex information structures and multiple information,the multi-instance multi-label learning framework is one of the most suitable methods for solving big data management issue.In this paper,the multi-instance multi-label learning framework is applied to data classification,and the multi-instance multi-label classification of natural scene images is studied.The main research work and innovation of this paper is as follows:(1)In this paper,we study and analyze the classic multi-instance multi-label algorithm framework,which mainly includes SBN feature extraction method,and MIML-BOOST algorithm and MIML-SVM algorithm that ignores the correlation between instances and labels,and M~3MIML algorithm,MIML-KNN algorithm and MIMLRBF algorithm that considers the correlation between instances and labels.Then,combined with five algorithm evaluation indexes,we evaluate the classical natural scene classification algorithm through experiments to verify the correlation between instances and labels for MIML classification performance.(2)In this paper,a multi-instance multi-label classification algorithm,SCIB-NN(Sparse Coding Instance Based-Neural Network),is proposed.In order to realize the self-clustering description of the SBN sample package features,the traditional idea of artificial measurement clustering was replaced by the example based K-SVD sparse coding.Using the deep neural network structure instead of the classical classifier to solve the limitation of the artificial measurement clustering thought,the realization of the pattern self-matching of the label and the example;the small dataset coding effect is further optimized by introducing the sparse residuals.Based on the five algorithm indexes,the improved algorithm is evaluated with the classic multi-instance multi-label classification algorithm,and the effectiveness and generalization of the improved algorithm are verified.(3)Also in this paper,we proposed SSMIB-NN(Sparse Subspace Mapping Instance Based-Neural Network)MIML framework Based on convolutional neural network feature extraction and Sparse subspace mapping.In order to achieve the different size of the image characteristics,puts forward using convolution neural network to extract object instead of fixed area SBN feature extraction algorithm of feature extraction,dilated convolutions is applied at the same time for the optimization of feature extraction effect.In order to optimize the self-expression effect of the example feature,the sparse subspace mapping algorithm is introduced and the performance of the subspace mapping is proposed.In addition,the sparse residuals is used to reduce the influence of small dataset training,and the classification accuracy of classifier algorithm is improved.The effectiveness and generalization of two improved algorithms and classical algorithms are compared and evaluated by experiments.