A Comparison of PCA and HOG for Feature Extraction and Classification of Human Faces

This paper presents a simple analysis and comparison of the efficacy of feature extraction and generalization achieved using Histogram of Oriented Gradients (HOG) and Principal Component Analysis (PCA). A Naı̈ve Bayes classifier was trained with a dataset of pre-classified human male and female face images. The classifier was then presented with a fresh set of faces and was given the task of classifying the images as male or female. Also attempted, was an experiment of testing using the same images used for training. It was observed that PCA was more adaptive than HOG, in identifying new faces. HOG was more accurate in identifying the same faces, provided there were sufficient training samples. In all, PCA was found to be a better feature extractor, since it did not require many training examples. It only mattered that a sufficient number of principal components were used to obtain good classification accuracy.