线性人脸对象类模型的匹配提升技术  

Improved Matching Algorithms for Linear Face Class Model

作　　者：付昀[1] 郑南宁[1] 

机构地区：[1]西安交通大学人工智能与机器人研究所,西安710049

出　　处：《中国工程科学》2005年第2期47-56,共10页Strategic Study of CAE

基　　金：国家自然科学创新研究群体基金资助项目 (60 0 2 43 0 1) ;国家自然科学基金资助项目 (60 2 0 5 0 0 1) ;河南省重大科技基金资助项目 (0 2 2 2 0 2 0 40 0 )

摘　　要：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 propos ed, which can solve the problem of detailed controlling and robust iteration for the realistic facial modeling. A new method——Dynamic Gaussian Pyramid Analysi s (DGPA), which combines Non-Uniform Sampling(NUS) method and Multi- Resolution An alysis, is presented. Integrating the PS Sampling and the Cluster Random Samplin g, the distribution of the sampled points in each level images of the Gaussian p yramid is adjusted dynamically. In coarse-to-fine scheme, the minimization alg orit hm is used to compute the near global optimal solution that may fit to yield acc urate model matching. Dynamic adjusting the boundary of the sampling cluster are a and the re-sampling ratio, the detailed representations are effectively contr o lled, and the model creation is quite robust. An improved Stochastic Gradient De scent (SGD) algorithm based on the Correlative Disturbance(CD) and Adaptive Lear ning Rate (ALR) is exploited to accelerate iteration convergence and compute val i d model parameters. With the examples of MPI Caucasian Face and AI&R Asian Face databases, experimental results in subjective evaluation and objective evaluatio n demonstrate the advanced model matching technique.

关 键 词：人脸建模 模型匹配 随机梯度下降 非统一抽样 多分辨率分析 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]