针对真实感人脸模型匹配的细节控制和稳健创建问题，提出了线性人脸对象类模型的匹配提升技术。基于非统一抽样(NUS)的动态高斯金字塔分析(DGPA)方法，结合不等概率抽样和整群抽样策略，自适应地动态调整每级高斯金字塔图像的抽样分布，利用最优化算法由粗到精的计算全局近似最优解，获得精确的模型匹配。动态调节整群区域边界并利用再抽样率调节抽样密度，可以有效控制人像模型的细节表达效果，提高模型创建的稳健性。随机梯度下降的线性相关性扰动(CD-SGD)和学习率自适应(ALR)技术，提高了模型匹配的准确性和收敛速度。以MPI和AI&R人像库为测试样本，主观与客观评价的实验结果验证了该模型匹配提升技术的有效性。

An advanced matching technique for linear face class model is proposed, which can solve the problem of detailed controlling and robust iteration for the realistic facial modeling. A new method——Dynamic Gaussian Pyramid Analysis (DGPA), which combines Non-Uniform Sampling (NUS) method and Multi-Resolution Analysis, is presented. Integrating the PS Sampling and the Cluster Random Sampling, the distribution of the sampled points in each level images of the Gaussian pyramid is adjusted dynamically. In coarse-to-fine scheme, the minimization algorithm is used to compute the near global optimal solution that may fit to yield accurate model matching. Dynamic adjusting the boundary of the sampling cluster area and the resampling ratio, the detailed representations are effectively controlled, and the model creation is quite robust. An improved Stochastic Gradient Descent (SGD) algorithm based on the Correlative Disturbance (CD) and Adaptive Learning Rate (ALR) is exploited to accelerate iteration convergence and compute valid model parameters. With the examples of MPI Caucasian Face and AI&R Asian Face databases, experimental results in subjective evaluation and objective evaluation demonstrate the advanced model matching technique.