Skull Ethnicity And Gender Discrimination Based On Deep Learning

The skull gender and ethnic identification is to identify the gender and ethnic of a person ethnic based on the skull.The identification of skull ethnic and gender is a popular research topic in the field of forensic anthropology and archaeology,which also plays an important role in detecting criminal cases,advancing historical and cultural excavation,and medical cosmetic surgery.However,the three-dimensional skull model is not easy to obtain and the data is complex.This article focusing on skull ethnic and gender identification aims to reduce the amount of calculation and improve the accuracy of identification.The main research content and method innovations are as follows:1.Ethnic identification based on the morphological characteristics of the skull and Back Propagation Neural Network(BPNN)According to the ethnic differences in the morphological characteristics of the skull,a skull ring boundary extraction algorithm is proposed,which extracts the overall outline of the skull,the boundary information of the eyes and the nose as feature vectors,and inputs them into BPNN for ethnic identification.Firstly,two network structures [36:12:2] and[36:6:2] with different parameters are used in our method,and the network is optimized by using the Adam algorithm and adding L2 regularization terms to avoid appearing in the training process Over-fitting.Afterward,the optimal parameter combination and network structure are selected to compare with a variety of classification methods.Finally,the foreigner skull data from the ethnic Caucasians together with the skull data from Han and Uyghur nationalities are identified by this three-class discrimination,in order to verify the application of our method.2.Combining skull feature image and convolutional neural network for ethnic identificationThe depth,curvature,and elevation values of the three-dimensional skull model are calculated and then the three values are used as the three channel values of the RGB image to generate a two-dimensional feature image of the skull.After the network training,the skull features can be automatically extracted from the feature image.In this paper,five classical convolutional neural networks are used to compare,and the results reveal that the performance of VGGNet16 is more superior than other methods.So the network is optimized by adjusting the parameters based on this network,upgrading the ethnic discrimination method with the higher degree of automation and discrimination accuracy.3.Combining the multi-view skull information with the improved Lenet5 for gender identificationCompared with the differences in skull ethnic characteristics,the differences in skull gender characteristics are relatively small.In response to this problem,the combination with the multi-view features of the skull to perform gender identification is proposed.The three-dimensional skull model is rotated and then projected onto the two-dimensional plane.Each skull generates 5 two-dimensional auxiliary images with different views and inputs them into the convolutional neural network,which is built as a multi-branch network structure.The network contains 5 branches with equal structures.The skull features at different views can be obtained by convolution and then multiple views features are combined for discrimination.The complete and incomplete skulls were identified respectively.In the experiment,the network performance was tested at different learning rates,and the skull with missing mandible was identified for incomplete skulls,which fully proved the versatility of our method.Consequently,this paper uses different skull representation methods and inputs them into the neural network,contributing to reduce the amount of calculation and improve the accuracy and degree of automation.Therefore,this research has more widely practical application compared with the classic skull gender and ethnic identification methods.