The Research Of Face Illumination Normalization Algorithm And Implementation

As the development of the social, intelligent systems have become a necessity oflife. Especially, base on the practical feature the Automatic Face Recognition Systemare subject to a great deal concern. Because of face recognition system is moreconvenient and user-friendliness than the traditional fingerprint recognition or irisrecognition system, thus it has been widely used in lots of filed such as automaticauthentication, intelligent access control, intelligent monitoring and human computerinteraction. Due to a large number of inputs makes the rapid development of facerecognition technology, lots of companies and research institutes have launched thesuperior performance of face recognition system in recent years. However, these facerecognition systems are designed under controlled environment. Thus the performanceof these face recognition system would significantly decline under the uncontrolledenvironment such as illumination variations, expression variations and the change of theage and so on. Thus to design illumination invariant and robust face recognition systemhave become a central concern of the many researchers. Then in this article we willdiscuss the following aspects:1. Research of illumination problem and illumination processing foundation theoryin the filed of face recognition. In this section the illumination variations problem isproposed. Then some illumination processing theories are introduced, such asLambertian reflection model, illumination cone theory.2. Research of different illumination processing algorithms. In this sectionsummarizes the current illumination processing algorithms. We simple introduce somesimple algorithms such as the algorithms base on statistics method, the algorithms baseon image processing. At the same we summarizes the current illumination processingalgorithms which include the method base on self-quotient image (single scale, multiscale) algorithm, Retinex (single scale, multi scale) algorithm, illuminationnormalization base on Discrete Cosine Transform in Logarithm Domain, illuminationnormalization base on the large-scale features and small-scale features. All thepreprocessing methods are compared and analyzed under different face datasets. 3. The main contribution of this dissertation is proposing a new illuminationpreprocessing algorithm base on Robust Principal Component Analysis. The traditionpreprocessing algorithms just can remove the affection of the attached shadow, but theycan not eliminate the cast shadow’s affection. In order to solve this problem, weproposed a new method base on Robust Principal Component Analysis to furthereliminate the affection of the cast shadow. In this section, we full introduce our methodand through two complete experiments to prove the validity and reliable of our method.