Exploring the pattern of multi-scale directional second-order gradients for facial recognition

One of the major challenges in facial recognition is to find an efficient method to extract the necessary features from face images under various lighting conditions. To deal with this problem, a great deal of existing algorithms have focused on analyzing the first-order gradients of a face image. This method has proven to be mildly resistant to lighting variations. However, these algorithms fail to make adequate use of potential information obtained from high-order gradients. In this paper, we propose to explore such high-order gradients along with an algorithm framework that performs the following: (1)utilize not only the information obtained from a first-order derivative, but also make use of the information provided by a second-order derivative; (2)a directional second-order gradient operator that captures a face image's edges and orientations; (3)an image-processing technique developed to suppress the noise in computed gradients using a structure tensor to obtain the second-order gradient. Four operators; First-order Gradient Operator(FGO), Second-order Gradient Operator(SGO), Directional Second-order Gradient Operator( DSGO), and Multi-scale Directional Second-order Gradient Operator(Mul-DSGO) are taken into consideration. Extensive experimental results using Yale B and Extended Yale B databases show that SGO performs better than FGO, DSGO performs better than SGO, and Mul-DSGO performs better than DSGO for facial recognition under various lighting conditions.