Research On Facial Landmarks Localization Algorithm For Occluded Face Image

Facial landmark location,known as facial landmark detection or face alignment,refers to the technique of finding the coordinates of feature points such as eyes,nose,and mouth on the image of face.It is an important part of computer vision tasks such as face recognition,face reconstruction,and head posture estimation.In recent years,deep learning has greatly improved the accuracy of face alignment,but posture,expression,lighting,occlusion,and blur still affect the improvement of the accuracy of face alignment.Wearing a mask to travel in the context of the epidemic has made it more and more urgent to study the occlusion factors of face-related algorithms.This paper optimizes the occlusion influence in the face alignment.Facial landmark location based on deep learning is generally divided into two types,one is the Coordinate-based regression model(CBR),and the other is the Heatmap-based Regression model(HBR).The coordinate regression model directly outputs the predicted coordinates of N ?2 dimensions;the heat map regression model outputs the heat map of N channels,and then maps the coordinates of the feature points from the heat map.,and then maps the coordinates of the feature points from the heat map.The heatmap regression model improves the generalization ability of the model.Generally speaking,the result of the heatmap regression model is better than that of the coordinate regression model.However,the prediction results of the heat map regression model may show scattered feature points when facing face images affected by factors such as posture and occlusion,because the heat map regression model cannot effectively express the correlation between feature points.In this paper,in view of the failure of face positioning caused by blocking interference feature extraction and the lack of occlusion labels for face feature points,an improved algorithm for face alignment is proposed.The main work of this article is as follows:1.A custom generation algorithm for occlusion is proposed to realize the amplification and labeling of the occlusion face data set,and to solve the problem of the complexity of occlusion and the lack of corresponding occlusion labels for facial feature points.By superimposing random size,shape,color,texture,and transparency occlusion on the original image,the facial feature points are accurately marked according to the area and transparency of the superimposed occlusion.Use the GAN network to generate occlusion(such as makeup,beard,glasses)that closely fits the face,and mark the feature points in the corresponding area according to the type of occlusion.2.A facial feature point positioning algorithm with adaptive weight of occlusion,named occ HRNet,is proposed to solve the problem of occlusion destroying the geometry of the face and the problem of interference with feature extraction in the occlusion area.Based on HRNet,a loss function of the adaptive weight of occlusion is designed to obtain a smaller weight for the occlusion point so that the impact of occlusion on feature extraction is reduced.In the network output stage,an auxiliary module is added to predict the degree of occlusion of feature points,and the occlusion is linearly transformed as the adaptive weight of the heatmap regression task.Generate point feature maps,edge feature maps,area feature maps,and cutting maps of the face based on the predicted coordinates,and enhance the geometric constraints between the feature points through the fusion of the multidimensional feature maps of the points,lines,and surfaces with the original image,so as to obtain more accurate and rich facial features.3.It simulates the application scenarios of mask blocking and face positioning that have become more common in the epidemic environment in recent years,and tests the practicality of the improved algorithm in this paper.The occ HRNet was used to perform mask-blocking face positioning tests on the Masked 300 W and COFW dataset respectively to obtain its accuracy on the standard data set.At the same time,the actual situation of face positioning without feature point labels is simulated on the MAFA dataset,and the positioning effect of masks covering faces is artificially evaluated by the positioning results of a single picture.A variety of ablation experiments and comparison experiments with other advanced algorithms are designed on related data sets such as COFW and 300 W.After a number of experimental data,it is proved that the improved algorithm has higher accuracy and robustness in the positioning of feature points of obscured face images.