|
Humans can recognize a face with binocular vision, while computers typically use a single face image. It is known that the performance of face recognition (by a computer) can be improved using the score fusion of multimodal images and multiple algorithms. A question is: Can we apply stereo vision to a face recognition system? We know that human binocular vision has many advantages such as stereopsis (3D vision), binocular summation, and singleness of vision including fusion of binocular images (cyclopean image). For face recognition, a 3D face or 3D facial features are typically computed from a pair of stereo images. In human visual processes, the binocular summation and singleness of vision are similar as image fusion processes. In this paper, we propose an advanced face recognition system with stereo imaging capability, which is comprised of two 2-in-1 multispectral (visible and thermal) cameras and three recognition algorithms (circular Gaussian filter, face pattern byte, and linear discriminant analysis [LDA]). Specifically, we present and compare stereo fusion at three levels (images, features, and scores) by using stereo images (from left camera and right camera). Image fusion is achieved with three methods (Laplacian pyramid, wavelet transform, average); feature fusion is done with three logical operations (AND, OR, XOR); and score fusion is implemented with four classifiers (LDA, k-nearest neighbor, support vector machine, binomial logical regression). The system performance is measured by probability of correct classification (PCC) rate (reported as accuracy rate in this paper) and false accept rate (FAR). The proposed approaches were validated with a multispectral stereo face dataset from 105 subjects. Experimental results show that any type of stereo fusion can improve the PCC, meanwhile reduce the FAR. It seems that stereo image/feature fusion is superior to stereo score fusion in terms of recognition performance. Further score fusion after image/feature fusion does not benefit much (or at all) for performance improvement which may imply that the power of image/feature/score fusion shall not be doubly applied to a face recognition process pipeline.
|
