抄録
In this paper, we propose a method of 3D face recognition using parallel pyramid neural networks (NNs). We first compensate for the poses of 3D original facial images using feature points and geometrical measurement. Then, the shape and texture images are extracted from compensated 3D images, respectively. The dimensionalities of the shape and texture images are reduced using PCA, processed by LDA for enhanced generalization. The corresponding reduced shape and texture features are combined to form the integrated features for face recognition. In the second step, a method is proposed for face recognition based on parallel pyramid NNs. Compared with conventional NN, the performance of the proposed parallel pyramid NNs is improved by utilizing lower layer effectively. Experimental results for 60 persons with different poses, facial expression and illumination conditions demonstrate the efficiency of our algorithm. In particular, the proposed method achieves 99.17% recognition accuracy using only 120 features.
