Research And System Implementation Of 3D Face Reconstruction Algorithms Based On Local Feature Constraint

With the rapid development of computer vision,and intelligent terminal products related to human face are slowly coming into people’s view,and research related to human face is becoming more and more valuable.In the study of human faces,the most important thing is the edge contour information of the face and the partial information of the facial features.How to make the neural network learn the contour information and facial features of the face with high precision and efficiency is the three-dimensional face reconstruction and dense An important prerequisite for the alignment of key points on the face.At present,most face reconstruction and dense face key point alignment methods only focus on the global information of the face and lack the learning of the local information of the face,which makes it difficult for the neural network to predict when the facial deflection pose is large and self-occlusion.The key points of the occluded part and the reconstruction results lack realism.Today,driven by intelligent products,face reconstruction and dense face key point alignment have become two indispensable key tasks.However,in a complex and changeable natural environment,the facial area caused by the large deflection angle of the head is occluded or the self-occlusion problem caused by eyes,hair,etc.,is caused by face reconstruction and dense face key point alignment.challenge.In response to this challenging problem,this paper proposes two novel deep learning models to solve the problem of face reconstruction and dense face key point alignment under large poses and occlusions.The main work of the thesis is as follows:1.This article first proposes a multi-path global discriminative feature and local detailed feature complementary information aggregation method,fully mining the advantages of each feature of global information and local information.Use global feature information to predict the rough face geometry,and then use the attention transformation fusion mechanism to initially aggregate the local detailed features;finally,use the detailed advantages of the aggregated local features to make up for the shortcomings of the global features,making the network deflect the face The facial shape reconstructed in the large posture with larger angle and self-occlusion is smoother and more realistic.2.In order to further optimize the above method,this paper proposes a Shuffle-self method,which integrates global information into more complex and detailed local information.In addition,it also makes full use of the advantages of the coordinate attention mechanism that embeds the position information into the channel attention,and uses the method of feature aggregation in two different spatial directions to capture the global and local blocks while retaining the precise position.Dependency.enhance the network’s learning of interesting areas,so as to improve the accuracy of network regression.and achieve the purpose of more efficient and accurate 3D facial reconstruction.3.In order to closely fit the algorithm ideas of the above two methods,this paper uses the feature pyramid convolution(Pyramidal Convolution)and the more efficient RepVGG network as the basic network.Pyramidal Convolution expands its receptive field by using unequal-sized convolution kernels and the number of grouped channels during each convolution,so that multi-scale feature extraction of global information can also capture the correlation with local information;RepVGG is a simple and powerful multi-branch architecture network that uses reparameterization to decouple the training process and inference time of the network to achieve a perfect balance of accuracy and speed.4.Designed and implemented a three-dimensional face reconstruction and facial key point alignment system.The algorithm described in this article was systematically displayed.Ordinary users can log in to this system to perform corresponding system operations,so that the algorithm proposed in this article is obtained.Practical application.