Face Recognition Using The Invariant Moment

Face detection and recognition technique is one of the most important and practical issue in the field of pattern recognition and computer vision. In the influence of lightness, orientation and expression of face images, disparities of different version of the same image are sometimes larger than another image. So the feature descriptor of face image is both invariant and stable under these influences. Invariant moments are shift-,rotation-,scale-,intensity-invariant and suitable for face image representation and recognition in this aspect.The traditional Pseudo-Jacobi(p=4,q=3)-Fourier moments(PJFM's) is defined in the polar coordinate system. At first, an improved algorithm of Pseudo-Jacobi-Fourier moments, which calculated directly in the Cartesian coordinate system, is proposed in this paper. The experimental results of using the improved PJFM's to describe and reconstruct the binary image show that the improved algorithm reduces not only the calculating complexity but also the rounding error of the coordinate transformation. The improved PJFM's also improves the calculation accuracy of Pseudo-Jacobi-Fourier moments and the recognition rate as well. This is the innovation in this paper.Secondly, Simulation results using improved PJFM's on the different kind of images indicate that the values of invariant moment are very different for different images and are nearly the same for different versions of the same image.Thirdly, it is proved that the improved PJFM's has the higher description ability and the lower NIRE than original PJFM's through reconstruction of face image experiments. At the same time, we found that the improved PJFM's is so perfect that the main information of original images could be recovered by lower order of PJFM's than original PJFM's.Finally, face images from ORL database were digitalized and recognized by improved algorithm first time in this paper. The training sample sets select 380 face images from 400 images in ORL database and the database of invariant moments was established by improved PJFM's.Then two groups of testing face images recognized by the minimum Euclidean distance classifier, the experimental results are ideal and the average recognition rate is 89.35%.