Performance Evaluation Of Face Recognition With Multispectral Image Fusion

As a typical biometric technology, face recognition has a wide range of applications in many areas, such as business, surveillance, video retrieval, and so on. Compared with other biometrics, face recognition has direct,friendly and convenient advantages in data collection. The data collected are generally single-spectrum image, visible image or infrared image. Visible imaging is vulnerable to the impact of illumination changes. Although better recognition results can be obtained under uncontrolled illumination conditions by IR image, it is vulnerable to the impact of temperature. Fusing IR with visible image is an effective solution for improvements of image quality and recognition performance.In this paper, three pixel-level fusion methods are used for enhancing robustness of face recognition, and objective indicators "entropy", "cross entropy" and "mutual information" are used to evaluate fusion image. In addition,for three face recognition methods using principal component analysis(PCA), linear discriminate analysis(LDA), and random projection and l 1 ?Minimization sparse representation respectively, a performance analysis is given by the CSU face recognition system. Experiment results show that the third method can handle the occlusion of eyeglasses, and get better in most conditions comparing with PCA and LDA.For the stability evaluation of algorithms, the subset bootstrap algorithm, an improved bootstrap, is suitable for face samples which don't meet the condition of independent identically distributed. With the confidence interval of equal error rate being the stability index, we compare the subset bootstrap technique with the conventional methods, and conduct an experiment of 50 times random sampling and another of 200 times. Experiment results show that the subset bootstrap method gives accurate evaluation of stability performance,and the result is more precise with the number of random experiments increasing. The face recognition based on random projection and l 1 ?Minimization sparse representation shows the best robustness with the smallest EER, but its stability performance is not as good as that based on LDA.