Classical methods of video coding exploit the stochastic properties of images such as temporal and spatial correlation. By exploiting structural features of images, new methods can be introduced for video coding and image compression, which are useful at very low bit rate transmission systems such as telephony. Since these methods are mainly used in videophone, their main object is coding of face images. With this assumption, to code these images it is needed to locate important features of the face in one frame and determine their variation with respect to previous frames. This method consists of two levels, analysis level and synthesis level. In the analysis level, the location of face features is determined automatically and their variation with respect to previous frames is encoded using a feature code book. In the synthesis level the image can be reconstructed by the received codes and using the same feature code book.
The “cut and paste” technique is used for model-based coding in this paper. The given method uses pictures of eyes and mouths and a single full face image. The main problems tackled in realizing an automatic model based system were to locate the eyes and mouth in a face, to drive pictures of eyes and mouth areas from a lead moving sequence, and to select the eye and mouth pictures that are best match for the eye and mouth areas in the sequence. The method which is used in this paper can detect the eye locations without any constraint. Based on the location of eyes and the relationship among face features, the mouth location and dimensions of sub images needed in cut and paste method are estimated.
It is shown that the coding system based on realized automatic offline model reduces transmission rate to a great extent and the reconstructed images have such a good quality that the pronounced words in the reconstructed image sequence are “lip- readable”.
)