Face Detection In Color Images With Complex Backgrounds

The study of human face detection originated from face recognition. It is to determine whether faces exist in the input image, and get the relevant information about the faces such as positions of the faces and facial features, etc. At present, the application of face detection has gone far beyond the domain of face recognition with its more and more academic and practical values. This thesis proposes an algorithm for face detections in color images in presence of varying complex backgrounds, multi-pose and complex facial expressions. It is composed of three parts:First, optimize the skin-tone model. We improved the elliptical-cluster skin-tone model proposed by Rein-Lien Hsu which has defects of missing detection and false detection. Together with two steps of pre-processes: light compensation and high pass filter filtration, our improved skin-tone model can extract skin areas from the color images with complex backgrounds much more precisely.Second, propose an algorithm for face detection with multi-pose. We designed a morphological filter to eliminate the scattered dots and smooth the binary images. And then, we proposed an adjacent-recursive algorithm to number each skin area, and check each of them with three acceptance criterions we designed. The human face areas from the images with complex backgrounds and multi-pose could be picked out.Third, extract facial features from faces with complex expression. We designed eye-model with luminance as parameter. With five steps: face edge confining, model computation, face area reconstruction, image enhancement, center computation, we can extract eyes from face area and get the positions precisely. With the position of eyes, we can confine the bound of mouse with physiology rules. By designing mouse-models with both luminance and chrominance as parameters, we can extract lip and teeth from the confined area, and compute the position of mouse's center precisely.Experimental results show a face detection rate of 94.3%, eyes' detection rate of 93.4%, and mouse's position detection rate of 92.6%. Our algorithm demonstrates a strong adaptability over a wide range of facial variation in 3D poses and expressions in color images with complex backgrounds.