Research On Face Reconstruction And Identity Recognition Method Of Digital 3D Skull

Craniofacial reconstruction and skull identification are important research contents in the field of craniofacial morphological informatics,and they are mainly used in forensic science,anthropology,archaeology and medical plastic surgery.Based on craniofacial digital model,computer-aided craniofacial reconstruction and skull identification is a new application of the intersection of information science and craniofacial morphology.It has the advantages of sufficient theoretical basis,strong objectivity and high efficiency.In recent years,it has become a research hotspot at home and abroad.In view of the problems existing in the current research of Craniofacial reconstruction and skull identification,such as inaccurate correspondence of dense points in the skull and face,highly subjective and cumbersome measurement process in the research of skull biometrics,insufficient expression of complex nonlinear mapping relationship between the local morphology of the skull and the face,and insufficient ability to express the intrinsic features of the skull and the face,this paper takes the knowledge of physical anthropology as a priori knowledge,combined with the theory of statistical analysis Image processing technology and in-depth learning methods are used to carry out in-depth research,further improve the research content of craniofacial morphoinformatics and support the practical application in related fields.The main research work and contributions are concluded as follows:1.A skull dense point correspondence method based on hierarchical optimization strategy and a face dense point correspondence method based on curvature graph are proposed.The dense point correspondence of the skull is divided into two processes: rough alignment and precise alignment.For rough alignment,k-means algorithm is used to eliminate mismatched point pairs.For precise alignment,k-d tree is added to improve the search efficiency of the algorithm,and geometric feature constraints are added to eliminate mismatched points twice.In the face dense point correspondence,the regional curvature graph descriptor of feature points is constructed to match the points with similar local shape,and the point correspondence search and matching strategy of curvature graph sub region is adopted to reduce the influence of missing region and improve the search efficiency.Then,on the premise of meeting the geometric consistency,the singular value decomposition method is used to calculate the rigid body transformation relationship between the surface point clouds to realize the rough alignment of the surface.Finally,the dynamic iteration coefficient is introduced to improve the iterative nearest point algorithm to realize the accurate alignment of face.The experimental results show that the above method can achieve accurate dense point correspondence between skull and facial skin,and provide a good data basis for subsequent skull face restoration and skull identification research.2.A skull gender recognition method based on wavelet transform and Fourier transform and a skull race recognition method based on improved convolution neural network and support vector machine are proposed.In the gender recognition method,wavelet transform and Fourier transform are used to quantitatively extract the features of the sagittal arc of the skull superior orbital margin and frontal bone respectively;Then,the features of superior orbital margin and frontal bone are fused,and the support vector mechanism is used to build a gender classifier to realize gender recognition.The experimental results show that this method explains the importance of classification of local regions of skull,avoids the cumbersome measurement of traditional methods,and the results are more objective and accurate.In the race recognition method,firstly,the structure of convolutional neural network lenet5 model is improved to extract the features of skull multi view image;Then a parallel support vector machine model is designed to construct a race classifier to realize race recognition.The experimental results show that the recognition accuracy of this method is better than the traditional method,and effectively solves the problems of strong subjectivity and large measurement error of the traditional method.3.A craniofacial reconstruction method based on region fusion strategy and and an improved generative adversarial network based craniofacial reconstruction method is proposed.The craniofacial reconstruction based on regional fusion divides the skull and face into five corresponding local feature regions;then uses the Gaussian process latent variable model to map the five regions of the skull and face to a low-dimensional latent space,respectively,and trains the least two-dimensional space in the latent space.Multiplying the support vector regression model established five mappings of the calvarial regions to the corresponding face regions.Finally,regional integration is carried out to achieve overall reconstruction.Design a hierarchical structure of generator and discriminator,each structure includes a global network and five local networks(left eye,right eye,nose,mouth and frame outline),the generator also has a fusion network for Synthesize face images from the outputs of the global and local networks.Additionally,a self-attention mechanism is introduced into Generative Adversarial Networks for modeling widely separated spatial region relationships.The adversarial loss,pixel loss,feature matching loss and local transfer loss are combined as the loss function of the network.Finally,the above improved generative adversarial network model is used to craniofacial reconstruction.The experimental results show that the proposed method can support the expression of complex nonlinear mapping relationship between craniofacial local morphology,and can better improve the accuracy of face reconstruction.4.A method of skull face recognition based on least squares canonical correlation analysis and skull face recognition based on fusion of view features and shape features is proposed.The skull face recognition based on least squares canonical correlation analysis builds a statistical shape model of the skull and face,and the skull and face are projected into the shape parameter space.Then,the main correlation information of skull and face is analyzed by least squares canonical correlation,and a global correlation analysis model and local correlation model are constructed to measure the global correlation and local correlation of skull and face.Finally,through the established correlation analysis model,the matching degree of each face in the unknown skull and face database is calculated,and the recognition result is obtained.Skull face recognition based on fusion of view features and shape features uses a multi-view neural network to learn the multi-view features of the skull and face,constructs scale-invariant wave core features according to the idea of eigenvalue normalization,and extracts the shape features of the skull and face;then The view feature and shape feature are fused by kernel canonical correlation analysis,and the intrinsic feature vector representation of skull and face is obtained.Finally,skull face recognition is realized by calculating the correlation coefficient of skull feature vector and facial feature vector.The experimental results show that the proposed method improves the face recognition rate and reduces the complexity of the skull face recognition problem.This method does not need to accurately extract and represent the intrinsic morphological relationship between the skull and the face,and can confirm the identity of the skull only by analyzing the correlation between the skull and the face.