Comparative study of subspace-based techniques in the task of partially occluded reconstruction of faces

Abstract

Facial recognition systems in controlled environments have presented satisfactory identification results. However, we can not make the same assertion when the collection environment is uncontrolled. The factors responsible for these low recognition rates are variations in illumination, pose, expression and occlusion, which introduce intraclass variations and degrade recognition performance. Compared with problems of pose, illumination and expression, the problem related to occlusion is relatively little studied in the area. In the literature there are some techniques based on subspace with initiatives to reconstruct the partly occluded face. However, there is no study showing the pros and cons of each variation. The objective of this work is to investigate the different existing techniques based on subspace, and with this to present the pros and cons of each technique. In this paper, the Wavelet transform was used to extract a set of characteristics of face images. According to the results we can see that the Fast Recursive PCA, Recursive and GPCA strategies achieved better performance, in terms of recognition rate, after evaluation with the Extreme Learning Machine classifier.