Fitting of semantic wire-frames using principal components analysis of a set of facial images

A new, fast and efficient method of automatic fitting of wire-frames for semantic model-based coding of head-and-shoulders video sequences is proposed. The method utilises the principal components analysis (PCA) of a codebook of facial images. The PCA of the facial codebook is performed only once when all the images from the facial codebook are manually pre-fitted with semantic wire-frames of the same structure. Neither PCA nor manual wire-frame fitting of the codebook are a part of the on-line processing and so do not influence the speed of analysis of an unknown image. The algorithm consists of two stages. In stage one the approximate position of the subject's head is estimated. In stage two, the accurate positions of the important facial features (the left eye, the right eye, the lips and the nose) are established. Both stages use the codebook of facial images. The information about the geometry of the human face is utilised in the second stage only. This increases the speed and reliability of the algorithm. The results obtained after analysis of a widely used set of images are presented.