Precision of Global Motion Parameter Detection in Interactive Semantic/Model-Based Image Coding

Video network is beginning to have various kinds of applications such as using a vertical blanking interval. Model-based coding is suitable for the narrow band network. Interactive model-based image coding, which had been developed for non-realtime environment, needs to be specified by the number and the precision of interactive operations. This paper first analyzes the noise in input feature points during the global motion parameter detection. Then a new structure filter for reducing the noise in the feature points on a facial wireframe model is developed. The structure filter keeps original facial shape composed of feature points and reduces variation in the mutual distance between the feature points. In evaluating the precision for applying the least-squares method for detecting depth and global motion parameters, the number of conditions of input data is considered from the point of variance of error of the estimated values. Based on the structure filter with the precision evaluation, the allowance error in the interactive operation for the encoding process is specified by the pixel accuracy for feature point position.