Semi-supervised And Tensor Learning For Multimedia Analysis

With the development of mobile Internet,massive user-generated multimedia has emerged, its high dimensions, complex information, no or weak label, and other charac-teristics have brought challenges to the multimedia analysis. In the vectorization process,the vector methods destroy the spatial structure of multimedia, and the high-dimensional vector leads to "curse of dimensionality". Since the tensor is the natural expression of images or videos, the tensor methods draw more attention in the image processing and video analysis. However, some tensor methods are used in the subspace learning or lower order space, which has many limitations in applications. In higher order tensor mod-el, some researchers introduce the CP (CANDECOMP / PARAFAC) decomposition in support vector regression and ridge regression, and obtain the support tensor regression and tensor ridge regression model. However, the non-uniqueness problem of tensor rank results in declined accuracy.In order to solve the non-uniqueness problem of tensor rank in CP decomposition,structured sparse is introduced in chapter ?, which produces different sparse effects on each factor matrix. Then add joint sparse constraint, so that all factor matrices have the same column sparsity, and the number of non-sparse columns is the tensor rank, so as to achieve the purpose of tensor rank selection. Due to the addition of two sparse con-straints, we can exploit the sparse information for each factor matrix, and then use the sparse information for tensor rank selection. In order to solve the structural information loss caused by Bag-of-Words (BoW) in the quantification process, we propose the tensor BoW in chapter ?. Experiments show that compared with other methods, the tensor BoW can more effectively express the contents of multimedia. The combination of the tensor al-gorithm and the tensor BoW has better results in this chapter. Although some researchers have proposed tensor logistic regression based on CP decomposition, the non-uniqueness problem of tensor rank still exists. In chapter ?, we get Tucker ridge regression algo-rithm by introducing Tucker decomposition in the tensor model. In chapter ?, we extend the logistic regression to their corresponding tensor model through the introduction of structured sparsity and Tucker decomposition. The tensor algorithms in chapter ? and IV effectively solve the non-uniqueness problem of tensor rank. Effectiveness is verified in different multimedia data sets.For supervised learning method, when the data dimension is too high and there are little labeled training data, "over-fitting" problem is prone to occur, thereby reducing the accuracy. Semi-supervised learning methods can use unlabeled data, so they are widely applied for real data. There are some videos with no label in the real world, and label-ing these videos is too costly. In order to handle video in the real world, we propose a semi-supervised transfer learning method in chapter ?. Firstly, establish a graph-based semi-supervised classifier based on video features, and we can utilize unlabeled video information to improve the accuracy. Secondly,we can establish a transfer learning mod-el by using the shared information between video key frame and images, so as to use relevant information of images to strengthen results. Finally, place the two classifiers in a unified framework for optimization, and get a more efficient video classifier. Exper-iments show that, the method proposed in chapter ? is more effective than that of the relevant algorithms. In order to effectively use the structure information of multimedi-a data and relevant information of unlabeled data, we combine the tensor learning and semi-supervised learning together in chapter ?, and propose the semi-supervised tensor learning model. Experiments demonstrate the effectiveness of this method. Finally, we present conclusions of the paper and future possible research directions.