Research On Facial Landmark Localization Based On Deep Learning

The purpose of facial landmark localization is to automatically locate the vital part of the face,such as eyes,eyebrows,nose,mouth and so on.It is also an important prerequisite for face tracking,pose estimation,3D face reconstruction,facial expression analysis,facial beauty,face recognition and other face applications.Though facial landmark localization has a long research history,the performance of facial landmark localization is still restricted by the various factors such as occlusion,illumination,expression,detector and pose.With the emergence of deep learning,the convolutional neural networks are widely used in facial landmark localization tasks and have achieved excellent performance.The paper proposes a facial landmark localization method based on the convolution neural network(CNN)and another method based on hourglass network.The main works are as follows:(1)This paper proposes a facial landmark localization method based on CNN,this method based on the residual network,and optimizes its main module named residual module,and proposes a bottleneck residual module based on dilated convolution and a bottleneck residual module based on attention mechanism.The feature maps with different perceptive fields are fused by residual module based on dilated convolution to expand the perceptive field and enhance the representation of the convolutional neural network.The attention mechanism is used to get the important information and suppress the useless information such as background.Secondly,this paper uses the stacked residual modules based on dilated convolution and residual modules based on attention mechanism to further enhance the network performance and improve the precision and accuracy of facial landmark localization.The experimental results show that using the data on the 300W for facial landmark localization,this method improves the accuracy of the original residual network by 6.45%,and exceeds some current facial landmark localization methods.(2)This paper proposes a facial landmark localization method based on hourglass network,this method inspired by the stacked hourglass network which has achieved excellent performance in human body pose estimation,and use the stacked hourglass network for facial landmark localization.Since the diff-erent size of input and output of hourglass can bring quantization error,this paper modified the original hourglass network to avoid the error by reducing the input size and increasing the number of up-sampling.Furthermore,this paper improves the up-sampling method of the original hourglass network,and uses the transpose convolution to replace nearest neighbor interpolation of the original hourglass network.The experimental results show that using the data of 300W for facial landmark localization,this method improves the accuracy of the original hourglass network by 5.2%,and exceeds some current facial landmark localization methods.