Research And Improvements On Active Shape Models

Active shape model is a facial feature point locating method, which uses local texture model to search locally around a single feature point and global statistical model to limit the shape composed of all the feature points; the two steps repeat iteratively and converge to the optimal shape finally. Active shape model is a fast algorithm, generally applied to the occasion requiring high real-time.However, active shape model has the following three significant shortcomings: (1) Active shape model of convergence speed and the probability of falling into local maxima are dependent on the initial shape. When the initial shape starts far away from the real shape, the number of iterations will increase by a large margin, the probability of falling into local maxima will rise. (2) 1D texture model of active shape model only samples the pixels along the normal line; it can be used to describe the feature points on the edge of face and facial features, but not for non-edge or inner point. (3) Local search strategy of active shape model is equivalent to the exhaustive search on the normal line, traversing the points on the normal line to find the point having minimum template matching function value. This search strategy is extremely simple and not uses any optimizing search strategy, therefore less efficient.For these three shortcomings, this paper contributes to improve active shape model in the following three aspects: (1) Improving the methods of locating the initial shape of active shape model to get the initial shape more accurately and quickly. (2) Combining active shape model local search and mean-shift search process to improve the ability to describe the local texture model and the efficiency of the local search strategy. (3) Extending the 1D texture model of active shape model, studying the difficulties that exist in it and finding some solutions.