An Improved Supervised Descent Method based Face Alignment Algorithm

Abstract

Aiming at existing problems about Supervised Descent Method (SDM), such as inaccurate facial feature extraction and the poor final alignment effect resulted by local optimum, an improved SDM (ISDM) based face align- ment algorithm is proposed. Firstly, an improved multi-scale Histograms of Gradient (IMHOG) feature extraction method based on multi-layers is raised, which expresses more refined facial features and makes faces be recognized more easily. Meanwhile, the social spider optimization (SSO) is applied to op- timize the estimated shape in iteration globally, which can avoid the local opti- mal. And it makes the estimated shape closer to the real shape so that the final alignment effect is more precise. Experiments have shown that the proposed al- gorithm can get better results than previous algorithms in LFPW, AFLW and 300-W datasets.