Face verification with feature fusion of Gabor based and curvelet based representations

  • Authors:

  • Fan Ou;Zhaocui Han;Chong Liu;Zongying Ou

  • Affiliations:

  • School of Mechanical Engineering, Dalian University of Technology, Dalian, China 116024;School of Mechanical Engineering, Dalian University of Technology, Dalian, China 116024;School of Mechanical Engineering, Dalian University of Technology, Dalian, China 116024;School of Mechanical Engineering, Dalian University of Technology, Dalian, China 116024

  • Venue:
  • Multimedia Tools and Applications
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

Face verification has broad potential in varieties of multimedia applications, such as security access control, surveillance monitor, image retrieval and intelligent human machine interface. However, the existence of variable lighting, pose, facial expression, aging and other random factors often causes the occurrence of recognition errors. Further upgrading the performance of face verification systems is not only a challenging but also an urgent task. Information fusion had proved to be one of the promising approaches in upgrading verification performance since more cues and evidences were provided. Gabor feature face representation and Curvelet feature face representation were chosen for fusion processing, since both representations are good at depicting image intrinsic pattern but with different emphases. After calculating the Gabor features and Curvelet features of a face image, the mutually correlated projection pairs of these two individual mode features were first yielded by canonical correlation analysis (CCA) method. Then, an integrate projection set can be built by simply grouping these two mutual correlated projection sequences term by term correspondingly. The integrate projection set possesses most of the information contained in Gabor features and Curvelet features and is optimally reorganized in a correlation sense. To further enhance the discriminant capabilities, a linear discriminant analysis (LDA) post-processing is applied on the integrate projection set to yield the final fusion feature set. The experiment results testing on MBGC data set show that the proposed fusion approach does reduce the error rates significantly as compared with using individual mode feature alone. FRR100 and FAR1000 were reduced about 30% and more.