Finding Initial Estimates of Human Face Location

This paper describes a method to nd initial estimates of face location in an image where the orientation and viewpoint of the faces are not known. Features such as eyes, nose and mouth are detected from the image using quadrature phase lters and grouped into potential face candidates. AAne invariants are used in grouping to overcome the problem of variation in viewpoint. An eecient searching algorithm is proposed to group these features based on the constraints in the geometry. Each face candidate is then evaluated using a belief network which assigns probabilities to each face candidate and rejects improbable ones. A result of 93% accuracy in detecting viewpoint variations is obtained. Human face recognition has been an interesting problem attempted by numerous researchers over the years. There are many important applications such as criminal identiication, visual surveillance and human computer interfacing which continuously provide the drive and motivation for research in this area. Most human face recognition algorithms have assumed that the location of human face in the image is known, or that the face can be easily extracted from the background. However, this is not true of most applications. Hence, face detection and localization still remains as an important problem to be solved. Recent work on face detection are attempted using various techniques: neural networks (Rowley et al. 9]), shape statistics (Leung et al. 5]), bandpass ltering (Graf et al. 3]), ellipse tting (Jacquin and Elefthe-riadis 4]), and colour (Wu et al. 11]). The neural networks and shape statistics approach works only for fronto-parallel faces with little variation in viewpoint. The methods based on bandpass ltering and ellipse tting works only for head and shoulder images with very little background clutter. Wu et al. 's method of using colour and fuzzy logic works for more general scenes but it cannot cope with diierent hair colour, or when the skin coloured regions in the image doesn't form an elliptical shape of a face. Yow and Cipolla 12]'s work on face detection under diierent viewpoints makes use of Gaussian derivative lters to detect features. This technique has been shown to work well but it has the problem of having too many false candidates and being unable to reject false candidates. In this paper, we describe the use of quadrature phase lters in feature detection which can signiicantly improve the robustness of the system and reduce the number of false detection. A …