Analysis of the effect of using non-composite multi-channel raw color images on face recognition accuracy with arbitrary large off-the-plane rotations

The need for a solution capable of identifying individuals from a distance, possibly wearing a disguise and in a crowded and busy environment is steadily increasing in many homeland security applications. The use of face recognition techniques has unique advantages when compared to other biometric methodologies - face images can be obtained from a distance and without the cooperation of the subject. A recent example demonstrating the growing need for such solution are the videos released by some terrorist organizations where the suspect's face is occluded with only the eyes being visible. Research has yielded several state-of-the-art algorithms but their accuracy is greatly lowered in the presence of various factors. Face recogition with arbitrary large face off-the-plane rotation remains one very challenging and open research problem. In order to build better face recognition algorithms, it is important to identify and analyze what and how factors affect the accuracy of said algorithms so these methods can be improved and made more robust. In our previous work we presented how factors such as image registration, number and type of training templates and the presence of varying amount of partial face information that can be used to improve the performance of three very popular widely used face recognition algorithms. In this paper, we examine how using multi-channel color images instead of their one-channel grayscale/color composite improves face recognition accuracy in the challenging case when face arbitrary large off-the-plane rotation is present. We demonstrate the improvement to face recognition accuracy through the very popular and widely used Eigenface-based, Fisherface-based and Direct Correlation-based algorithms. Our findings and experimental results with the data, show that when using frontal-facing images as external test images with frontal-facing images as training set images, the additional information contained in the multi-channel color images does not improve the composite gray-scale face recogntition accuracy. However, in the case where the algorithms are trained using only images with slight or considerable off-the-plane rotation, and externally tested either on frontal-facing images or images with arbitrary off-the-plane rotation, the information provided by the multi-channel color images boosts the face recognition accuracy for all three recognition algorithms.