Research On Craniofacial Retrieval Based On Deep Neural Networks

Craniofacial reconstruction is a hot research topic in forensic anthropology.Craniofacial retrieval includes similar skull retrieval and reconstructed face retrieval.Similarity skull retrieval is to retrieve the most morphologically similar skull from a craniofacial database for a given skull.Reconstructed face retrieval retrieves the most similar face photos from the face photo database.In the craniofacial reconstruction method based on template deformation,the reference skull that is most similar to the skull to be reconstructed is first retrieved from the skull database and then the deformation between the two skulls should be calculated.The same deformation should be carried out on the face skin of the reference skull and finally the face skin after deformation would be taken as the reconstructed face.Therefore,the results of similar skull retrieval methods will directly affect the effect of craniofacial reconstruction.The reconstruction of the unknown skull face does not mean the end of craniofacial reconstruction.It is also necessary to retrieve the most similar face photos from the missing person photo database to help determine the identity of the unknown skull.Therefore,Reconstructed face retrieval has very important social value and research significance.The traditional craniofacial retrieval method relies on the expert experience,the operation is tedious and it has very high requirements for measurement accuracy.The computer-aided craniofacial retrieval method can effectively shorten the retrieval time,improve the retrieval accuracy and avoid the secondary damage caused by the measurement tool to the skull.The research contents of this thesis mainly include:(1)Aiming at the problems of complex operation and low accuracy of existing similar skull retrieval methods,this thesis improves and implements a skull retrieval method with improved mixed domain attention mechanism,which improves the accuracy of similar skull retrieval to 87.5%.First,data preprocessing is used to obtain skull multi-angle images to construct a skull image dataset;then,a network model with improved mixed-domain attention mechanism is used to extract features from the skull image.Finally,similar skull retrieval is performed using the extracted skull features.Experimental results show that this method can effectively extract skull features and improve the accuracy of similar skull retrieval.(2)Aiming at the problems of complex feature extraction operation and low measurement accuracy in existing reconstructed face similarity measurement methods,this thesis improves and implements a reconstructed face similarity measurement method based on deep neural network.The accuracy of the measurement is increased to 96.67%.Firstly,a pair of positive reconstructed face images and corresponding face photos were collected to construct a dataset;Then,the improved Inception model is used for feature extraction and the extracted features are dimensionality reduced by combining with Principal Components Analysis(PCA).The dimensionality reduced features are input into the neural network;Finally,the similarity between the reconstructed features and face photos is output.Experimental results show that this method can effectively improve the accuracy of face similarity measurement.(3)In view of the problems of low retrieval efficiency and poor retrieval accuracy of existing reconstructed face retrieval methods,a reconstructed face retrieval method based on deep neural network is designed and implemented,The TOP1 accuracy of retrieval of reconstructed faces is increased to 99.57%.Firstly,the datasets of reconstructed face and face photos were constructed;Then,a weighted convolutional neural network is used to extract the features of reconstructed face and face photos,the extracted face feature vectors are input into the Pseudo Siamese network for further feature extraction;Finally,KNN method is used to retrieve the most similar face photos.Experimental results show that this method can effectively improve the accuracy of reconstructed face retrieval.This thesis proposes several improved algorithms for the retrieval of skull and reconstructed face.These algorithms can effectively extract the depth features of craniofacial data.They improved the accuracy of similar skull retrieval and reconstructed face retrieval.Combining the practical application needs of forensic anthropology and criminal investigation,this research can not only provide reliable data support for craniofacial reconstruction technology,but also provide positive help for finding missing persons and identifying the skull origin information.