Image Classification based on t-Linear Combination of Tensor

A multidimensional linear algebra based image classification technique is proposed in this paper. The image on a particular class can be represented as a t-linear combination of images belong to same class whereas it is difficult to represent it using images in other classes. Motivated by this concept, the proposed method constructs tensors which provide approximate representation of each class during the training process. Lateral slices of these tensors are the images belongs to each class. In the next step, the strength of the representation coefficient is measured to determine the class where the image belongs. To achieve this, the proposed method computes a projection operator on to the set of subspaces throughout the learning process. Conventional method uses vectorization of images to form a data matrix. In contrast, the proposed method represents each class as a tensor for the effective utilization of structural information contained in the image samples. Feature extraction and classification is done in a single step. Moreover when compared to the predecessor, the proposed method requires a single projection operator which reduces the computational time to a large extent. The experimental results demonstrate the superiority of the proposed method over the existing method.