Facial feature localization without considering the appearance of image context

Abstract

Facial feature detection plays an important role in various applications such as human computer interaction, video surveillance, face tracking, and face recognition. Efficient face and facial feature detection algorithms are required for applying to those tasks. This dissertation presents the algorithms for all types of face images in the presence of several image conditions. There are two main steps corresponding to face and facial feature detection algorithms. First, the faces are detected from an original image. Canny edge detection is applied to find the edge of the image. A candidate face region can be found from the region having the number of pixels corresponding to average face template. Then, the matching value is calculated and applied to find the actual face. Second, facial feature detection is applied to the actual face obtained from the previous step. A proposed neural visual model (NVM) is used to recognize all possibilities of facial feature locations. The input parameters are obtained from the face characteristics and the locations of facial features which are independent of the intensity information. For the better result, an image processing technique called dilation is applied to remove some irrelevant feature regions. In addition, the algorithms can be extended to cover rotational invariance problem by using Radon transform to extract the main angle of the face. With more than 1,000 experimental images, the algorithms are successfully tested on various types of faces with color intensity, gray intensity, binary intensity, object occlusion such as sunglasses, scarf, and hand, facial expression, lighting effect, noise and blurry images, as well as color and sketchy images from animated cartoon. In particular, the method achieves more than 94% detection rate on the average.