A Multiscale Method for HOG-Based Face Recognition

Image representation is an important process in image classification, and there are many different methods for representing images. HOG (Histograms of Oriented Gradients) is a popular one which has been used in many applications including face recognition, pedestrian detection and palmprint recognition. In this paper, a novel method is presented to improve HOG-based image classification by using the multiscale features of images. For each image, multiple HOG feature vectors are extracted under different spatial dimensions (or ’scales’). These ’multiscale’ feature vectors are then fused into a distance function to calculate the distance between two images. Experiments have been conducted on ORL face database, AR face database and FERET face database. Results show the use of multiscale HOG features has led to significant improvement in performance over the use of single scale HOG features. Results also show that the nearest neighbour classifier equipped with our distance function is comparable to the well-known and widely-used benchmark classifier.