Image recognition and classification with HOG based on nonlinear support tensor machine

Spatial structure information is very important in image analysis algorithms. Traditional machine learning methods based on vectorization strategies often ignore the spatial information of the original data, resulting in low image recognition and classification accuracy. Different from the vector representation, the tensor representation can preserve spatial structure information in still images. Histogram of oriented gradient (HOG) is a feature descriptor that generates histograms by calculating the gradient amplitude and direction of the local area of the image. This paper proposes a new support tensor machine learning method based on tensor space, where the HOG method in the form of tensors is used to extract the features of the image data. Borrowing the broadcasting idea, we investigate a flexible and concise nonparametric tensor model to capture the nonlinear spatial information in HOG tensor. By the polynomial spline approximation to nonparametric functions and low-rank CP decomposition, a new spline support tensor classification algorithm is studied with the alternating iterative approach under the framework of original support vector machine. The experimental result shows the superior performance of the proposed model compared with the existing methods.