Tensor Principal Component Analysis And Its Application In Recognition Of Image Sequences

Image sequences,such as video images,medical images and hyperspectral remote sensing images,belong to 3D tensor.Tensors are essentially multidimensional arrays,and they are multilinear generalization of matrices.Image sequences have not only become the most commonly used information carriers in human activities,but also have become a hot topic in the field of pattern recognition in recent years.Principal component analysis is the most commonly used method in feature extraction.Since image sequences belong to 3D tensor,the method of tensor principal component analysis is applied to the recognition of image sequence in this paper.However,in the recognition of image sequences,the existing method of tensor principal component analysis can not determine a suitable singular threshold in tensor model to determine feature preservation rate of image sequences.In the classification and recognition,the existing tensor classifier can only deal with two-dimensional data,and can not directly classify multidimensional features.The so-called tensor classifier can only be limited to two dimensions at present and it is essentially a two-dimensional tensor classifier.In order to solve the above-mentioned problems,this paper presents a study on the recognition of image sequences in tensor model.The specific contents are as follows.Above all,the traditional tensor principal component analysis method is unable to determine a suitable singular threshold in tensor model.That is to say,it can not find a balance between removing noise and retaining details in tensor model.This paper proposes a truncated tensor principal component analysis method based on TPCA,which is used to determine a proper threshold to filter the smaller singular values and preserve the larger singular values,so as to find a balance between removing noise and retaining details,and to complete feature extraction of image sequences.Then,in order to effectively improve the image recognition accuracy,taking into consideration that the image sequences after feature extraction are still the tensor model,this paper puts forward a three-dimensional support tensor machine of tensor mode,which is used to directly classify tensor data and avoids tensor data vectorization.The main work on the classification of tensor pattern is as follows: Firstly,in the 3DSTM algorithm,traditional support vector machine and two-dimensional support tensor machine are improved by the rule of tensor multiplication,and they are extended to the theoretical N dimension,directly handling the inputs of tensor model.Secondly,according to advantages and disadvantages of SVM and 3DSTM,combined with multiple rank,this paper proposes the design of multiple rank dimensionalsupport tensor machine classifier on the basis of 3DSTM classifier model,which makes the image sequences recognition rate higher.In this paper,two kinds of classification algorithms of tensor form TTPCA and 3DSTM,as well as TTPCA and MR3 DSTM are compared with the current prevalent two algorithms in experiments.The experimental results show that the proposed algorithm can improve the recognition accuracy and speed of image sequences.Besides,compared with the two algorithms of MR3 DSTM and 3DSTM,the recognition accuracy of MR3 DSTM is higher.