Face Alignment Under Complex Scenes

In the field of computer vision and image perception,research of face images has been one of the attractive topics,and face alignment is a challenging problem in this research.Face alignment is a process of locating facial landmarks and it plays an important role in many visual tasks like face recognition,expression analysis.In this paper,we start with the method based on model regression and present a two-layer regression framework.We test and verify our two-layer shape regression framework on two widely used datasets(LFPW and Helen)for alignment.Specifically,we have conducted the following research:1.We implemented explicit shape regression(ESR)algorithm proposed by Cao et.and our implementation is comparable to which was reported by the original authors.By analyzing the implementation and experimental results of this algorithm,we discussed the potential defects and proposed an improved method;2.We alalyse the physiological differences of facial landmarks and categorize the holistic shape constraints into two parts: inter-component constraints and intra-component constraints.We compute the offset distance that landmarks shift from their normal location when one makes various expression and we proposed the notion of feature point stability;it provides a theoretical explanation for our algorithm;3.We proposed a two-layer shape regression framework to estimate coordinates of landmarks.In the first layer,we regress the entire facial landmarks in a holistic manner and obtain an initial estimation of coordinates of the feature points;based on the initial estimation done by the first layer,we can estimate the local region for sub-shapes and we regress the landmarks of each sub-shape independently in the second layer;then we synthesis the final prediction of landmarks from estimation of each sub-shape.Our algorithm is improved based on ESR and we respectively increased the accuracy rate by 8.8% and 7.9% in the two datasets(LFPW and Helen).