Face Alignment Algorithm Based On Feature Pyramid Hourglass Network

Face landmark location has achieved great development in recent years.Due to the small amount of data for face landmark location,it is difficult to capture complex facial images in real scenes,making the diversity of face information scarce.Therefore,based on the characteristic pyramid hourglass network,this paper proposes an algorithm that can be used for large-pose face landmark location.The main contents are as follows:First of all,using FU-Net Network to improve 2D large-pose face landmark accuracy,the input of the network is the four types of images:the original image together with its flip and rotation,and the mirror image after rotation.The training loss is a weighted sum of the loss functions of these four types of images.The FU-Net Hourglass network uses Dense Blocks to densely connect contextual features.In order to balance weights and time-consuming,two feature fusion networks are designed,which extracts stronger facial feature information,effectively improves the face landmark accuracy,and the model parameter quantity and network time are smallest.Secondly,the 3D face localization algorithm uses a feature pyramid network to enhance multi-scale feature fusion performance.In the middle layer feature extraction,deep separable convolutions and hole convolutions are used to replace ordinary convolutions to reduce the network calculation and time.The final prediction module combines deformable convolution and focuses more on learning the location area of facial landmark.The output features of the network are fused with the deformable convolution features,which can improve the accuracy of 3D face pose positioning and improve the overall performance of the algorithm.Finally,compared with the standard U-Net network,the FU-Net hourglass network has a smaller number of channels.The model parameter memory occupies 14MB,which is only 1/17 of the standard U-Net network.On the 300W dataset,the mean error of FU-Net is 3.5%,which is 1.12%lower than the standard U-Net;which proves that the improved loss function and the feature fusion method can effectively improve the face landmark accuracy.In the 3D face feature pyramid network,the forward network takes 9ms,the time is only 1/9 of the 3DDFA network,and 1/2 of the 3DSTN network.On the AFLW2000-3D dataset,the mean error of our algorithm is 4.01%.It is 1.31%lower than the 3DDFA network and 0.48%lower than 3DSTN.