Multi-Pose Facial Landmark Detection Technology Based On CNN

Facial landmark detection aims to estimate the location of landmarks for a given face,such as eyebrows,the corner of the eyes,the nose,the cheek,etc.This problem has received much attention as evidence by the recent advancement in both the megodology and performance.It is the premise of many face analysis tasks,such as face recognition,expression recognition and 3D face reconstruction.However,there is still much room for improvement in the large-pose facial landmark detection.The difficulties of multipose facial landmark detection lie in the following three aspects: 1)Some landmarks may disappear due to the large-pose.2)The field selection of facial landmarks may be different because of the large-pose;3)The performance of facial landmark detection for multi-rotation faces in plane is poor.This paper aims to study and realize multi-pose facial landmark detection technology.We define multi-pose as yaw and rotation-in-plane(RIP).Considering the difficulty of multi-pose facial landmark detection,we propose an improved HG(Stacked Hourglass Network)network.The main improvements are as follows: 1)We introduce the attention mechanism which enables the network to adaptively enhance the extraction of local features of visible landmarks;2)According to the symmetry of faces,the mirror loss function is introduced to enhance the modeling of shape features;3)A 360-degree facial landmark detection algorithm based on rotation calibration is proposed.The method proposed in this paper improves the performance of multi-pose facial landmark detection.The experiment part carries out detailed analysis and ablation experiments for each module we proposed.At the same time,it compares and analyzes with state-of-the-art algorithms on the public data set.The effectiveness of the method we proposed is verified.