基于正交投影的主动外观模型匹配算法

Improving Convergence of AAM(Active Appearance Model) Fitting Algorithm Based on Orthogonal Projection

为了提高传统主动外观模型匹配算法的收敛速度,提出了一种基于正交投影的主动外观模型匹配算法,该算法通过引入正交空间投影算法,将形状参数与纹理参数分别映射到不同的线性子空间,并利用梯度下降算法依次求解形状参数与纹理参数。实验结果表明,该方法在保证精确度的条件下,提高了收敛速度。

Aim. Cootes et al proposed AAM and AAM fitting algorithm and othersEa made improvements. But, to our knowledge, there does not as yet exist any papers that improved the slow convergence of AAM fitting algorithm to make it relatively fast while retaining the same high accuracy as Cootes et al. We now propose doing so. In the full paper, we explain our improvements in some detail. In this abstract, we just add some pertinent remarks to listing the two topics of explanation. The first topic is： the AAM fitting algorithm based on gradient descent. In this topic, we point out that for each iteration, if the transformation coefficient is only related to the values of current shape parameters, then, the transformation coefficient of AAM fitting algorithm may reduce the number of dimensions and computation load will be greatly reduced. The second topic is. the AAM fitting algorithm based on orthogonal projections. In this topic, through making use of the line of thinking on orthogonal projections in Ref. 5 by Baker et al, we project shape parameters and appearance parameters respectively into different linear subspaces and seek the solutions of the parameters, thus improving the AAM fitting algorithm. Finally, to verify our fitting algorithm, we perform three computer simulations. The first simulation uses respectively our algorithm and the fitting algorithm based on gradient descent to position human eyes respectively so as to verify the accuracy of our algorithm. The second simulation uses the convergence rate relative to initial displacement to verify the convergence of our algorithm. The third simulation verifies the convergence speed and accuracy of the algorithm by computing root-mean-square. The simulation results, shown in Figs. 2 through 4 in the full paper, indicate preliminarily that our algorithm has not only high accuracy but also relatively fast convergence.