Research On Methods Of Image Feature Extraction Based On Tensor Subspace

Feature extraction has been a key problem in pattern recognition, success or not offeature extraction will affect the result of classification. Its main task is to seek the mosteffective features from all the source features and achieve the purpose of dimensionreduction. Because the computational complexity of traditional one-dimensional featureextraction methods is large and also those methods destroy the intrinsic structure oforiginal images, this paper has a research on methods of image feature extraction basedon tensor subspace.Firstly, we propose a feature extraction method based on symmetric informationmatrix. It uses a parameter to control the proportion of the original and diagonal images,in this way, we not only get the symmetric image information matrix, but also make upfor the shortcomings that diagonal images destroy the structure of original images. Wemake full use of the complementarity of the original and diagonal images.Secondly, in this paper we improve the two-dimensional linear discriminantanalysis algorithm and propose a new algorithm called Two-Dimensional ShiftDiscriminant Analysis based on Maximum Margin Criterion (2DSDA-MMC). In thisalgorithm, we add the covariance of different rows to2DSDA, reduce the number offeature coefficient and get the optimal projection subspace through maximum margincriterion. After that, we extend2DSDA-MMC to the bilateral form in a cascadedmanner. It considers the covariance information of different rows and coloumssimultaneously, pays attention to the local structure of the images and improves therecognition rate further. Furthermore, the idea is used into Locality PreservingProjection (LPP). It retains the local structure information between samples andincreases the local information between the pixels, so another steady and effectivealgorithm is obtained.Finally, we extend Transformer Discriminant Analysis based on matrix to higherorder tensor space. This method can deal with the high order data without transformingthe data into vectors. It reserves the structure information of original data and reducesthe computational complexity. Experiment results verify the validity of this method.