Research On Pose Robust Face Landmark Detection

Face landmark detection,is an important research issue in the field of computer vision,and is also called face alignment.The goal of face alignment is to locate facial key points for any given face picture.Facial key points refer to the points defined on the facial features and contours.Face landmark detection has a vital influence on face recognition and expression recognition.However,the face pictures obtained in reality are affected by the factors such as lighting,pose,expression,occlusion,these are difficult for face landmark detection.For example,there are various face poses in reality,and the faces in different poses have different appearances,which make the research on face landmark detection still challenge.This paper focuses on pose factor and proposes the pose robust face landmark detection algorithm.The main work and contributions are as follows:Firstly,we propose a face landmark detection algorithm based on an adaptive SDM model.The algorithm divides face images with different poses into different pose categories through clustering,thus divides a difficult problem into relatively simple subproblems for processing.The face poses in each category are similar and the variance is small.At the same time,in each category,a good initial shape is given that is closer to the true shape,which makes the landmark regression more efficient and accurate.Following the rule of coarse-to-fine,the adaptive feature block size is used for feature extraction.The feature block size will get smaller with the increase of the iteration,which makes it possible to extract discriminative features that are useful to face landmark detection.Experimental results on LFPW,HELEN and 300 W datasets show that our algorithm is robust to pose and improves the accuracy of landmark detection.Then we propose a deep face alignment network based on Wing loss.The algorithm is a cascaded regression method,which is divided into two stages in total,and each stage uses a convolutional neural network for face landmark detection.A transformation mechanism is added between the first stage and the second stage to alleviate the impact of pose factor on face landmark detection.At the same time,a landmark heatmap is added to enable the network to use the entire face information.We analyze the effect of the existing loss function on the mean error of face landmark detection,and find that they only perform well for large errors.Then we propose a Wing loss,which can take into account the impact of large and small errors on face landmark detection.Experimental results on 300 W,AFLW and MTFL datasets show that the algorithm is robust to pose,and it is also robust to lighting,expression,and occlusion.