LOW BIT-RATE IMAGE CODING FOR FACIAL MOVEMENT
Reinhardt, Randal Todd
MetadataShow full item record
It is well known that video information uses large bandwidths for real-time transmission as in television broadcasting. Solving video bandwidth problems has focused on the use of image compression based on redundancy removal. However, this type of image compression alone is not. capable of reducing the bit-rate sufficiently to allow real-time video through low bandwidth channels such as telephone lines. This method presented in this 'thesis takes a different approach. By limiting the type of images to facial images, this method encodes face expressions into motion parameters for the purpose of low bit-rate transmission. Through these parameters, movement of the person can be reconstructed by using a simulated face. Data requirements per frame can be reduced dramatically by the method allowing the frame rate to increase to the minimum video rate standard of 30 frames per second. The implementation of the coding method requires an encoder and decoder. Since video is given as a sequence of images, a reference image is transmitted from the encoder to the decoder at the start. The encoder's operation is to extract and measure the features of the face. This process uses edge detection, object extraction, and object identification to find features. The extracted features are encoded with respect to the grid and the warping method used. Two methods of grid manipulating (action units and grid movement) and two method of warping (bilinear and cubic spline) are discussed in this thesis. Choice of method is based on operation speed, and algorithm complexity. The decoder was realized by the algorithms of the Attractor/Repellor for eyes, and mouth movement, 2-D motion algorithms for the head, and nonlinear scaling for head rotation. The thesis demonstrated that the correct features can be found for 80 percent of the frames in a video clip database. It was also demonstrated that the decoder using the bilinear warping can successfully simulate eye and mouth movement as well as head motions including horizontal and vertical head turning and head rotation.