Research On Multiple Craniofacial Similarity Calculation Method Based On Convolutional Neural Network

Computer aided craniofacial reconstruction is a new technology combining modern anatomy with computer as a tool for digital craniofacial restoration,which has wide application prospects.Experts and scholars in the related fields have been studying for many years on the algorithm for craniofacial restoration.However,the study of similarity calculation in craniofacial restoration is still lacking.In this paper,the convolutional neural network in deep learning is used as the basic research method.According to the characteristics of experimental data,four different convolutional neural networks structures have been built,and a multiple craniofacial similarity calculation method based on convolutional neural network has been implemented.The specific research contents as follows:1.A method of mapping 3D information to 2D information based on depth map is realized.This method uses the dimension projection,block processing,interpolation repair and normalization of 3D craniofacial for 3D craniofacial point cloud data,building a corresponding 2D depth map with depth information.The data retaining the original 3D information,and the convenience of 2D information is embodied in data processing and training.2.A face recognition method based on a eight layer convolutional neural network is proposed.This method is on the basis of repeated experiments based in the input data,and a convolutional neural network is constructed,which is made of “convolution layer-down sampling layer-convolution layer-down sampling layer-local connected layer-local connected layer-fully connected layer-softmax output layer”,while setting up the network learning rate and the configuration of the parameters of each layer,realizing the automation of calculating craniofacial similarity.Experiments show that compared with commonly used craniofacial similarity calculating method,the method is automated,which does not need to calibrate feature points and feature calculating,and it has high accuracy and efficiency.3.A method of fusion of craniofacial data based on RGB-D is realized.This method combines the depth map of 3D craniofacial data and the 2D craniofacial image data with weighting,setting depth information as the fourth dimension.This method overcomes the limitations of a single data form,and improving the calculating and matching accuracy of craniofacial similarity.4.A Craniofacial similarity computing system based on convolutional neural network is designed.This system can realize data training,visualization of convolutional neural network and calculation of craniofacial similarity.