Multispectral Image Fusion Methods And Applications In Face Recognition

Human face recognition is a very difficult and important task in the field of pattern recognition, which has widely potential applications. However, face recognition based only on the visual light spectrum has shown difficulties in performing consistently under uncontrolled operating conditions. Face recognition accuracy degrades quickly when the lighting is dim or not uniform. The use of thermal infrared (IR) images can improve the performance of face recognition under uncontrolled illumination conditions. But eyeglasses may result in loss of useful information around the eyes in thermal face images since glass material blocks a large portion of thermal energy, which are important features for recognition. This thesis focuses on the methodology of multi-spectral image fusion for improving the performance of face recognition system.Firstly, a new multi-modal image fusion algorithm based on singular value decomposition (SVD) image slicing is proposed. The original image is decomposed using SVD and then recombined into three layers of different energy resolution. The three layers account for different proportion of the total original image energy and are called low, high and ultrahigh resolution layers. The low resolution layers are fused using a weighed averaging and sharpening scheme. Three fusion schemes based on intensity selection, local energy and wavelet transform are proposed to fuse the high resolution layers. The ultrahigh layers are discarded as noise. Two objective evaluation methods are used to evaluate the proposed image fusion algorithm. Experiment results show its validity and the advantages over some traditional image fusion algorithms.Secondly, a multi-spectral image fusion scheme based on double-match is presented in this thesis. It is performed by introducing the local gradient into the local energy and then selecting the low frequency coefficients according to the type and the value of match measures. Results from two objective evaluation methods without requiring the reference image and visual effects show that the image fusion scheme based on double match measure is effective. In which, local features are fully applied, and contour and edge details of the fused image are efficiently remained, so the block effect can be effectively avoided, and scheme presented outperforms some traditional image fusion schemes.Finally, a face recognition algorithm based on visible light and far infrared image fusion is presented by an integration of advantages from SVD image slicing and double match fusion scheme. Experimental results based on a comparison of several...