A novel multi-view 3D face registration method based on principal axis analysis and labeled regions orientation called local orientation registration is proposed.The pre-registration is achieved by transforming the mu...A novel multi-view 3D face registration method based on principal axis analysis and labeled regions orientation called local orientation registration is proposed.The pre-registration is achieved by transforming the multi-pose models to the standard frontal model＇s reference frame using the principal axis analysis algorithm.Some significant feature regions, such as inner and outer canthus, nose tip vertices, are then located by using geometrical distribution characteristics.These regions are subsequently employed to compute the conversion parameters using the improved iterative closest point algorithm, and the optimal parameters are applied to complete the final registration.Experimental results implemented on the proper database demonstrate that the proposed method significantly outperforms others by achieving 1.249 and 1.910 mean root-mean-square measure with slight and large view variation models, respectively.展开更多
基金supported by the New Century Excellent Talents of China (NCET-05-0866)
文摘A novel multi-view 3D face registration method based on principal axis analysis and labeled regions orientation called local orientation registration is proposed.The pre-registration is achieved by transforming the multi-pose models to the standard frontal model＇s reference frame using the principal axis analysis algorithm.Some significant feature regions, such as inner and outer canthus, nose tip vertices, are then located by using geometrical distribution characteristics.These regions are subsequently employed to compute the conversion parameters using the improved iterative closest point algorithm, and the optimal parameters are applied to complete the final registration.Experimental results implemented on the proper database demonstrate that the proposed method significantly outperforms others by achieving 1.249 and 1.910 mean root-mean-square measure with slight and large view variation models, respectively.
