Tensor Object Recognition Using Optimized Projection Based Support Tensor Machine

Classification is an important problem in artificial intelligence. In this work, a novel optimized projection based support tensor machine (OPSTM) has been presented for recognition of higher-order tensor objects. The proposed OPSTM can address the classification problem for high-order tensor object within the tensorized framework by reformulating the weights parameters as a tensor. By incorporating intra-class scatter matrix into objective function, the OPSTM can acquire the optimal projections which can gain not only maximal margin between the classes but also minimum class variance. Besides, the OPSTM employs different tensor decompositions for effective computation of inner-product, which can save computational time and storage space. To assess capability of the presented OPSTM, we employ seven third-order tensor datasets and three second-order tensor datasets to conduct experiments. The experimental results and Wilcoxon signed-ranks test demonstrate that, as for training speed and test accuracy, the proposed OPSTM is significantly superior to C-SVM and STM, specially for the third-order tensor classification tasks. Moreover, the proposed OPSTM is compared with the state-of-the-arts including deep learning methods and dictionary learning methods. The comparison result shows that the proposed OPSTM has its own advantages. The influence of rank values in different tensor decompositions is discussed and the range of optimal rank value is given, which can provide a good guidance for applying OPSTM in real-life tasks of tensor-based pattern recognition.