Human Identification Performed With Skull’s Sphenoid Sinus Based On Deep Learning

The mainstream methods of human identification in forensic work are DNA comparison,fingerprint identification,facial recognition and so on.However,the corpses often occurred extreme phenomena such as decomposition,ossification,and incineration when large-scale disasters occurred.The traditional methods are not available.The ability to identify the deceased in a timely and accurately is quite crucial to speeding up post-disaster reconstruction and stabilizing the mood of the victims’ family members.The skull’s sphenoid sinus is usually selected by forensic experts as reference data to identify the deceased because it’s biospecific and solid.In the past,forensic radiologists performed human identification via artificial visual recognition by using sphenoid sinus’ CT images.This method usually implemented on a smallscale data set in existing studies,which couldn’t verify the reliability in large-scale disaster situations.To solve this problem,this thesis established the first large-scale 3D sphenoid sinus database and initially proposes the screening criteria of sphenoid sinus CT images.The largescale 3D sphenoid sinus database included 732 subjects with a total of 1475 sphenoid sinuses that were segmented and reconstructed by distinguishing the sphenoid sinus cavity in the CT metadata,which exceeds the scale of all the sphenoid sinus dataset currently known in the literature.The traditional artificial visual recognition of sphenoid sinus has low efficiency,high difficulty,strong subjectivity.Because there is no standardized process and unified indicators of sphenoid sinus recognition,it is difficult to be widely popularized and applied.To solve these problems,this thesis firstly attempts to distinguish the sphenoid sinus based on deep learning and proposes a human identification algorithm based on multi-view of the 3D sphenoid sinus.This algorithm collected multi-view joint images by scrip from the sphenoid sinus,and the MVSS-Net is proposed by this algorithm based on the residual learning and complementary relationship of the angel joint.in the end,this algorithm using MVSS-Net to extract the 3D shape description operator from the multi-view joint images to achieve human identification.According to the experimental results on the test set,the Top 1 accuracy of the algorithm on the test set of 132 objects is 93.94%,which is 28.79% higher than the artificial visual identification.The entire human identification process only takes 55 seconds,which is extremely fast.By compared with human beings,this algorithm is higher accuracy,more objective and much faster.It has the value of application in forensic human identification.The human identification algorithm based on multi-view of the 3D sphenoid sinus is faced with some problems.The 3D shape description operators between different sphenoid sinus objects are relatively similar.The process of multi-view joint image collection is redundant.The algorithm workflow is complicated.To fix these problems,this thesis proposes a human identification algorithm based on the point cloud of 3D sphenoid sinus.Firstly,this thesis proposed a point cloud pre-treatment algorithm based on filtering sampling which was used to process numbers of the sphenoid sinus point clouds to the same,because different sphenoid sinus point clouds have different numbers of points.Secondly,For the characteristics of point cloud data,the algorithm proposes the Point SS-Net based on the efficient channel attention mechanism and the large margin Softmax loss.This algorithm achieved human identification by 3D shape description operator which was extracted from sphenoid sinus point cloud through Point SS-Net.According to the test set of 132 subjects,the Top 1 accuracy rate of 99.24% was achieved by this algorithm,and only consuming 20.5 seconds in total.Compared with traditional algorithms,the Top 1 accuracy of this algorithm is improved by 18.18%,with the speed increasing by about 500 times.The Point SS-Net model complexity is reduced to 1/7compared with MVSS-Net.The experiments show that the algorithm has high accuracy,high level of automation,and fast identification speed,and can be used for rapid identification when large-scale disasters occur.