基于主成分分析的唇部轮廓建模

摘要

研究了基于主成分分析(PCA)的唇部轮廓建模方法.首先,对5 000个样本唇部轮廓进行标定并对标定的坐标数据进行Procrustes分析,使数据归一化.然后,通过PCA算法寻找形变模式,在保持形变范围内最大限度地降低数据维数并利用所得到的均值和特征向量构建唇部轮廓模型.最后,利用PCA得到的前16种模式所建立的模型对5 000个样本原始的唇部轮廓进行重构.实验结果显示:PCA得到的前4种模式分别描述了唇部角度、下唇、尺度以及唇角等的形变信息,其余模式描述了唇部更细致的形变,模型重构的唇部轮廓与相应样本原始唇部轮廓的每个特征点之间平均差异均不大于0.6个像素宽.结果表明所建唇部模型能满足特征定位精度要求.%A modeling method of lip contours was proposed based on Principle Component Analysis (PCA). Firstly, the lip contours of 5 000 training samples were labeled, and the Procrustes analysis was performed on the coordinates gotten by labeling to normalize the data. Then the PCA was used to identify modes in data and compress the data dimension without losing the lip contour information. Furthermore, the lip contour model was constructed by using a mean value and eigenvectors gotten by PCA. Finally, the model constructed by the first 16 modes gotten by PCA was taken to reconstruct the lip contours of the original 5 000 samples. Experimental results indicate that the first 4 modes respectively describe the rotation, lower lip, scale and corners of the lip,and other modes describe more detailed lip variation. The mean difference of every feature point between the lip contour gotten by model reconstruction and the original contour is less than a width of 0. 6 pixels. The model can satisfy the precision requirements of the feature location.