Feature Selection for HOG Descriptor Based on Greedy Algorithm

In order to make an efficient face recognition algorithm with real time processing, we should design good feature extraction and classification methods by considering both low computational costs and high classification performance. Among various feature extraction methods, the histogram of oriented gradient (HOG) feature shows good classification performance to classify human faces. However, high-dimensional features such as HOG feature waste lot of memory and computational time. some parts of HOG features for occluded face regions have negative effects in classifying face images, especially occluded face images. Therefore, we should select variable HOG features not only to reduce the computational costs but also to enhance classification performance. In this paper, we applied the greedy algorithm to effectively select the good features within traditional HOG feature. In order to compare the proposed feature extraction with the conventional HOG feature, we fixed classification method such as compressive sensing technique for selected features. Experimental results show that the proposed feature extraction has better classification performance than the traditional HOG features for face datasets with partial occlusion and/or various illumination conditions.