| 摘要 |
<p>The present invention is a quick human face alignment method. The method proposes using a series regression structure and a quick feature selection method to perform quick alignment on a human face shape in a human face picture. The system is divided into two phases. In the learning training phase, firstly human face shape initialization is performed on a real human face shape sample which is well labelled, then a pixel difference feature resolved from a current human face shape of each sample and a regression target are used to respectively train a plurality of regression devices, and these regression devices constitute a series structure. Finally, the current human face shape of each sample and each regression device training result are added as a final output. In the test phase, firstly a test sample arbitrarily initializes a human face shape, then the pixel difference feature resolved from the current human face shape and the regression target are put into a well-trained regression device, and are added to a regression device result, which is a final human face estimation shape. The method may quickly and accurately acquire human face shapes of human face pictures in various conditions (illumination, expressions and postures).</p> |
