| After Yuille et. al. introduced the use of deformable templates for estimating facial features, deformable templates have attracted attention because of their simplicity and flexibility. Other applications like lip reading, database retrieval, moving object tracking, medical image detection, and automatic navigation systems have used the deformable template model. In this thesis, a multiscale and multilayer deformable template is developed to describe not only the mouth contour but also the teeth contour. This model can decrease the computational time required to make the template converge to the correct minimum cost point by using Weighted Gaussian Approximation (WGA) to estimate the initial template estimation. Using the multiscale representation, we can reduce the computation time and improve the results of the deformable template. The contour description of bird in a poultry grading system is another application of MDT. We also applied the MDT to a poultry deboning system. The interaction between speech and video is explored as a means to better estimate the deformable template. We also explore the estimation of a new template using a linear prediction method, where we estimate the template parameters of next video frame as a linear combination of the templates of the previous template parameters. |
