Study On Multi-classification Algorithms Using Support Tensor Machines Based On Optimal Projection

The research object of this article is the automatic target recognition using synthetic aperture radar (SAR) images. The original SAR images contain many background clutter and noise, which strongly influence the classification accuracy of radar targets. Therefore pre-processing of SAR images is firstly preformed to eliminate clutter and noise. The main steps include logarithmic transformation, adaptive threshold segmentation, morphological filter, geometric clustering processing, image enhancement, normalization processing and pose estimation.Then the introduction of tensor is helpful to solve the small sample size problem. TensorPCA and TensorLDA in the tensor subspace make samples reduce dimension, extract character. By comparing with vector, these methods reserve the structure information of objects, so as to be helpful to improve the learning performance. The support tensor machine (STM) available is based on the iterative method, which has a long runtime, and has a weak ability in generalization.Finally based on Fisher principle this paper applies optimal projection to multi-classification problem, and proposes an Optimal Projection algorithm for STM. Furthermore, in the proposed Optimal Projection algorithm we introduce two methods to determine the projection vector: In the first method, the training samples of all sub-classifiers are projected by the same projection vector; and in the second one, the projection vectors are calculated for every sub-classifier and then the samples are projected by the corresponding projection vector. Experimental results indicate that the proposed algorithm not only improves the training speed, but also makes the samples after projection have the best numerical separation. The first method has a higher recognition rate than the second one, which is due to the following reasons. After the samples are projected to the tensor subspace, samples belonging to different classes are separated, while samples belonging to the same class get close to one another. Finally, we use several multi-class strategies, such as One-versus-One, One-versus-All, and Directed Acyclic Graph, to classify the samples after projection.