Dental Recognition Methods For Human Identification In Forensics

Forensic odontology aims to deal with human identification using the dental images(that is dental recognition),and it is of great practical values in Forensics.Additionally,the superiority of the computer-aided dental recognition techniques has been highlighted in many studies.Nevertheless,existing traditional methods for dental recognition lack the necessary robustness in practice when there are large shape variations for dental images and a larger dental image database.As a preliminary of this thesis,a novel deep learning method has proposed to assist in the dental image recognition,called LCANet(learnable connected attention network for human identification using dental images).While promising performance has been demonstrated,it fails to achieve a high cosine similarity score to perform human identification for one individual.Additionally,the current dental image dataset simply contains two dental images for one individual,and it is possible to establish considerable differences between two images of one individual.Firstly,the shortcomings of the learnable connected attention network for human identification using dental images are figured out as a preliminary of this thesis.Then,for exhaustively improving the recognition performance of the aforementioned method,this thesis proposes a new deep supervised neural network to accurately segment 2-D panoramic dental X-ray images.On the other hand,by assisting in the recognition supervised neural network with correct segmentation labels,experiments are conducted to justify if the segmentation results help to improve the recognition accuracy.In light of this,this thesis then focuses on tackling the challenging but natural few-sample recognition of the dental images,and proposes some carefully designed architectures.The contributions of this thesis can be summarized as follows:(1)This thesis proposes a new generative adversarial network to accurately segment 2-D panoramic dental X-ray images.Within the general framework,it designs an effective instantiation,called the improved attention module,which is simple and can effectively model the global context.The LCANet could extract tooth features by using a data-driven convolutional neural network,which learns the shape and data priors simultaneously.However,at the same time,it fails to achieve a high cosine similarity score to perform human identification for one individual.Accordingly,this thesis proposes a new generative adversarial network to accurately segment2-D panoramic dental X-ray images,and then it attempts to improve the recognition performance of the previous dental recognition method by assigning to it correct segmentation inputs.Extensive experiments demonstrate that the segmentation method proposed in this thesis can achieve 97.27% of accuracy,91.69% of dice similarity coefficient and 84.44% of jaccard index,significantly superior to the existing methods.Moreover,by employing the segmentation results on the LCANet,it further improves the Rank-1 accuracy by 0.72%,and the Rank-10 accuracy by 1.19%.(2)This thesis proposes a novel method for human identification by automatically and accurately matching 2-D panoramic dental X-ray images,which called Bilateral-Branch Network with mask guiding and feature reweighting for dental recognition.The core idea is to assist in the recognition supervised neural network by assigning to it correct segmentation labels.Additionally,this thesis develops a new loss function that acts as a more effective alternative to previous approaches for dealing with the loss imbalance.This work aims to tackle the challenge of few-sample dental recognition and further investigate the more effective way to employ the segmentation results.Concretely,this thesis proposes a novel supervised neural network,which called Bilateral-Branch Network with mask guiding and feature reweighting for dental recognition.In the experiments,empirical results demonstrate that the proposed method can achieve a new state-of-the-art performance.A total of 2140 dental images from 1000 different individuals are used to evaluate our method.The evaluating database is much larger than all of the previous study of automatic dental recognition,and our method reaches 87.81% rank-1 accuracy and 96.67% rank-10 accuracy.