| Facial landmark detection is a key part of tasks such as face recognition and threedimensional reconstruction of faces.It involves multiple research fields such as applied mathematics and computer vision.This paper focuses on the problem of low landmark detection accuracy of large-scale face images,based on the research of two kinds of facial landmark detection models called coordinate point regression and Heatmap regression,an optimization algorithm for face landmark detection task for large-scale face images is proposed.First of all,for the facial landmark detection model based on coordinate point regression,a 5-point affine transformation Coarse-to-Fine network is designed to detect landmark in largescale face images.The affine transformation matrix is calculated between the landmark prediction results of the Coarse part and the standard landmark,and the original face image pose is corrected using the transformation matrix,then sent the transformed image to the Fine network to fine-tune the landmark positions.And during the training process,this paper has carried out different processing methods on the input data of the Coarse and Fine structure networks,in order to further optimize the accuracy of face landmark detection.Secondly,for the facial landmark detection model based on Heatmap regression,the paper designs a feature fusion Stacked Hourglass Networks.Through the fusion of the feature information of the original Stacked Hourglass Networks down sampling part,the prediction layer of the face landmark detection model can receive the feature information from the global image.Realize the control of the overall content of the face image,and it is conducive to the detection of key points of the large-scale face images.Finally,through experimental verification,the Coarse-to-Fine structure has an average positioning error of 0.38%lower than the test of a single Coarse structure on a 300W dataset,indicating that the Coarse-to-Fine structure is useful for large-scale face landmark detection.The test results of the improved Stacked Hourglass Networks model on the 300W dataset show that the feature fusion Stacked Hourglass Networks face landmark detection algorithm compared to the original Stacked Hourglass Networks has reduced the average error on the 300W-full test set from 3.35%to 3.31%.And through further analysis of the results,it is concluded that the reduction of errors mainly comes from the 300W-challenge data set,which shows that the algorithm has a certain improvement effect on the landmark detection process of large-scale face images,and the average test error of the feature fusion Stacked Hourglass Networks on the CelebA dataset can reach 2.8%,which shows that the algorithm performs well on different face landmark detection datasets. |
