A Study On Face Detection Technology With Varying Illumination

Face detection technique has been proven to be an important and popular research direction in the field of computer science nowadays. It means the process of number, position, size and poses assurance in input-picture. Its research combines a great deal of disciplines, such as computer vision, image processing, pattern recognition, biological information discerning, physiology, psychology, etc. Face detection is a challenging task since human faces have many variant factors: different scales, varying illumination, partial occlusion, multipose, etc. Moreover, human faces are nonrigid objects. They contribute substantially to the variation of facial appearance in an image. Varying illumination is one of the bottle-neck that restricts the development of face detection. How to extract the illumination invariant features is one of the important problems to be solved.In this paper, we do some researches on face detection with varying illumination. It is described as follows:Firstly, to detect skin from color images effectively, an optimal color space description for skin detection is presented in the paper. Skin color and non-skin color are regarded as two patterns; an optimal color space description for skin detection is extracted based on the Fisher criteria, Foley-Sammon transformation, mean separability and covariance separability; the final, relatively accurate skin regions are obtained by segmentation.Secondly, based on the skin region segmentation, the face candidate regions are further obtained according to the confirmation in a region of face shape information. Also, with regard to illumination, we improved a method which based on the quotient image (QI), and a correction-coefficient choice method based on real-valued genetic algorithm (RGA) is presented in this thesis.Thirdly, with regard to the problem of the feature extraction with varying illumination, an algorithm based on an improved local binary pattern (LBP) as face feature is presented. Then, support vector machine (SVM) is adopted following an improved LBP to train and classification, and thus the final face detection results are obtained.Finally, the above method is tested on a great deal samples, and an evaluation scheme for the results is provided. In the two experiments, the accuracy of 95.11% and 86.37% are obtained respectively. They show the validity of the above methods.