Exploration of illumination normalization based on divisive normalization

The changing environment, in reality, causes chief occlusions in face recognition. In other words, uncontrolled situations in face recognition is a choke point within practical applications of face recognition. Lighting normalization in recognition preprocessing contributes significantly to enhance the accuracy of recognition systems. Tremendous varying illumination conditions need to be cut down so as to achieve more compelling recognition results. Illumination normalization is a prominent concern in the cutting-edge merchant face recognition algorithms for a long time. The divisive normalization is an established method in canonical neural computation, which was developed to deal with responses in the primary visual cortex. And the divisive normalization turned to be effective by operating in the optical system and other sensory procedures. The thesis demonstrates a method by using divisive normalization as a tool to tackle with illumination variation problem in face recognition prior to recognition. The thesis will discuss a method based on the divisive normalization model, which combines the illumination estimation of adaptive smoothing and Retinex theory framework. The algorithm based on adaptive smoothing integrating discontinuity measurements and continuous convolution to generate image illumination normalization. The results gave evidence of performance improvements with the proposed procedure. It is evaluated based on the Extended Yale Face B database.Master of Science (Computer Control and Automation