摘要
为了提高三维稀疏重建的精细度和鲁棒性,该文在改进模型参数先验分布描述的基础上,推导出一种最大后验概率参数估计方法即L1约束最小二乘重建方法。在BFM三维人脸库上测试重建算法的误差,L1约束最小二乘重建方法的重建结果精度优于无约束和L2约束最小二乘重建结果,也优于文献中的动态成分选择算法的结果,并且重建性能更鲁棒。实验结果表明:该方法与传统的无约束最小二乘或L2约束最小二乘等方法相比,能更可靠地克服实际中三维稀疏形变模型的表示基存在严重的多重共线性的问题,具有较好的重建性能。
The finesses and robustness of 3-D sparse reconstruction is improved by a modified prior-distribution description of the model parameters in an L1-constrained least squares(L1-LS) reconstruction method,which is a maximum a posteriori(MAP) parameter estimator.The algorithm reconstruction error was evaluated using a BFM 3-D face database.This method gave smaller reconstruction errors than the ordinary least squares(OLS) and L2-constrained least squares(L2-LS) methods and the dynamic component deformable model(DCDM) algorithm.In addition,the reconstruction performance of L1-LS is more robust.Experimental results demonstrate this method more reliably overcomes the severe multicollinearity problem existing in representation bases for the 3-D sparse morphable model than traditional sparse reconstruction methods like OLS or L2-LS to give better reconstruction results.
出处
《清华大学学报(自然科学版)》
EI
CAS
CSCD
北大核心
2012年第5期581-585,共5页
Journal of Tsinghua University(Science and Technology)
基金
国家"九七三"重点基础研究项目(2007CB311004)
国家自然科学基金资助项目(60972094)
关键词
信息处理
线性回归
三维人脸
稀疏重建
information processing
linear regression
3-D face
sparse reconstruction