In the field of visual data mining, Histogram of Oriented Gradients (HOG) and its variants have been widely used. The speed and ability to extract image features that are robust against many types of distortions such as scaling, orientation, affine and illumination that HOG offers have made it a popular choice for the task of detecting images in scenes for classification. However, the high dimensionality nature of HOG descriptors (features), usually in the order of multiple thousands of them per image, would require careful consideration in place to achieve accurate and timely categorization of objects within images. This work explores the possibility of processing HOG features as tensors, or multi-dimensional arrays. A direct result of that is tensor decomposition techniques such as canonical polyadic (CP) decomposition performed on the high-order HOG tensors as the mean for dimensionality reduction by filtering. This work focuses on the impact of this approach on both accuracy and efficiency, comparing it against the standard practice of processing HOG features. Validating with the Caltech-101 dataset, the results achieved with artificial neural network (ANN) classification indicate that the proposed method not only improves the overall system performance, it also achieves the edge in accuracy by a notable margin.