Research And Implementation Of Facial Landmark Detection And 3D Face Reconstruction

With the rise of computer vision,face-related intelligent products have gradually entered people's lives.Therefore,the research significance and value of face analysis and recognition have become more important.To do a good job of face analysis and recognition,the most important information is the facial features and contour information.How to accurately and efficiently make the neural network learn the apparent and structural-semantic information of the face is an important prerequisite for face landmark detection and face reconstruction.At present,most face landmark alignment and face reconstruction methods only focus on the apparent information of the face but lack the learning of the face structure information.This makes it difficult for the neural network to reason about the face landmark of the occluded part with large poses and occlusions and it will reconstruct a poor result.Face landmark detection and 3D face reconstruction have become two indispensable tasks in the development of human life.This paper aims to propose a new deep learning model for solving face landmark detection and face reconstruction.Especially in an unconstrained natural environment,the large head pose and self-occlusion of the face have brought new challenges to face landmark detection and face reconstruction.Aiming at this problem,this paper proposes a new deep learning model to solve face landmark detection and face 3D reconstruction under large face pose and occlusion.The main contributions of the paper are as follows:1.For the problem of large pose and self-occlusion of the head in an unconstrained natural environment,this paper proposes a graph structure-constraint to constrain the geometric structure information of the face,with the purpose of fully mining the apparent and structural semantic information of the face.In the case of large posture and occlusion,the network can also infer from the structural information that the occluded part will not only stay on the learning of episodic semantic information.2.We propose a vector graph structure reasoning network loss function,which includes a distance graph structure inference loss function and a directed graph structure inference loss function.The distance graph structure loss function constrains the relative distance between the face components,and the direction graph structure loss function constrains the relative method between the face components.So that the neural network can more effectively extract the spatial structure information of the face.3.This paper introduces the DAMD-Net network as a basic network.DAMD-Net is a dual attention mechanism and an efficient end-to-end three-dimensional face alignment framework.Through deep separable convolution,tightly connected convolution,and lightweight channel attention mechanism,a stable network model is established.Through the channel and spatial attention mechanism,important apparent and structural-semantic information of the face will be extracted and the secondary semantic information will be ignored,reducing redundant information and improving the accuracy of network regression.4.For training,we propose a weight mask which assigns different weight to global and local distance graph structures and compute a weighted loss.5.A face landmark detection and face reconstruction platform is designed.The algorithm in this paper is integrated so that ordinary users can log in to this platform for face landmark detection and face reconstruction.