2D and 3D information fusion based ear recognition

In order to solve pose and illumination variation problems in ear recognition, an information fusion method was proposed to fuse 2D and 3D ear information at the decision level. For a 2D ear, the ear images will become nonlinear manifold structure due to pose variation, so the manifold learning method, isometric mapping (Isomap), was used to extract features. For a 3D ear, the 3D local binary pattern (3DLBP) method was adopted for feature extraction. Then 2D ear recognition and 3D ear recognition were implemented separately. Finally, results from the 2D and 3D were fused at the decision level. Experiments were done on a database of 79 persons, one of which has eight 2D ears with pose variation and six 3D ears with illumination variation. It is found that both the recognition rate and verification rate are significantly improved compared with 2D ear recognition and 3D ear recognition alone.