Research On Face Detection Based On Skin Color Segmentation And AdaBoost Algorithm

Face detection is one of the most important tasks in the fields of computer vision and pattern recognition. It has been widely applied to such fields as content-based image retrieval, virtual reality, video surveillance, face recognition and personal identification, etc. Face detection is the precondition for face recognition. Face detection result is usually affected by the illumination, background, and head posture and so on, which make the detection process complicated and challenging. To solve these problems, we mainly study the methods to express facial features and face detection algorithms, and then make some improvements. The researches in this thesis mainly include:(1) We make some improvements to Multi-Gaussian skin color segmentation method on the basis of studying several skin color segmentation methods. The implementation of this improved method can be described as follows: (a) Obtain the probability density function of skin color respectively by fitting the distribution of skin color chromaticity Cb and Cr using the Gaussian surface in three regions of brightness Y; (b) Select appropriate thresholds to make the cumulative probability distribution of skin color close to 100%; (c) Segment the images using corresponding thresholds in each brightness region.(2) Haar-like intensity features are proposed after studying the theory of edge detection, which are representative of the information of image edge. The values of the intensity features are equal to the absolute values of corresponding Haar-like features.(3) This thesis constructs Haar-like multi-threshold features by combining the thresholds of Haar-like features and intensity features, which have lower error rate. Haar-like multi-threshold features divide the eigenvalue space into four regions with the thresholds of basic Haar-like features and intensity features. By calculating the weights of positive and negative samples of each region, we can judge whether the region belongs to human face or not, then mark it with a flag.(4) Cascade face detector is improved by using the preprocessing of skin color segmentation. The candidate face region of one color image is quickly achieved by Multi-Gaussian threshold segmentation. For the window being detected, if the amount of skin color exceeds a certain percentage, it will be detected by the AdaBoost classifier; otherwise, it will be discarded directly.