A v-Twin Bounded Support Tensor Machine for Image Classification

Abstract

Support Vector Machine (SVM) is an effective tool for classification problems. With the advent of the information age, tensor data problems are common in pattern recognition field. However, SVMs may lead to structural information loss and the curse of dimensionality when encounter into tensor data. In this paper, we propose a novel tensor-based classifier called the v-Twin Bounded Tensor Machine $\left( {\nu {\text{-TBSTM}}} \right)$. It is an extension of $\nu {\text{-TBSVM}}$. $\nu {\text{-TBSVM}}$ solves two smaller Quadratic Programming Problems (QPPs) instead of a larger one, meanwhile, it adopts the structural risk minimization principle. Compared with existing SVMs, $\nu {\text{-TBSVM}}$ has certain advantages. $\nu {\text{-TBSTM}}$ inherits all the advantages of $\nu {\text{-TBSVM}}$, moreover, it utilizes the structural information of tensor data more directly and effectively, thus it gains better performance. The experimental results indicate the effectiveness and superiority of the new algorithm.