Dental biometric systems: a comparative study of conventional descriptors and deep learning-based features

Abstract

Dental biometrics utilizes the evidence divulged by radiographic dental images for human identification. Dental biometrics is commonly used to recognize dead individuals by comparing their before death (ante-mortem) and after death (post-mortem) dental images with the attributes such as tooth contours, restorations, number and shape of the teeth, and relative positions. In recent years, conventional local image descriptors and deep learning based features have shown excellent performances in different applications due to their excellent flexibility and capacity. Regardless of dental biometrics' potential, the efficacy of human identification using dental radiographic images with advanced machine learning methods has not been adequately analyzed so far. In this paper, we investigate various facets of conventional hand-crafted microtextural (12 different descriptors) and deep learning-based features (8 different architectures) for dental biometrics. The dental features of single tooth images (segmented with Mask RCNN) are extracted and the features are matched with various distance functions and fusion techniques. Also, pretraining and fine-tuning transfer learning methods are employed while evaluating deep learning based methods. The empirical analysis, performed on a dataset of 100 dental images and fully reproducible, demonstrates the potential of local microtextural and deep learning tools for dental biometrics. The experiments showed that deep learning based methods with majority voting outperform other methods where Inception architecture has higher identification accuracy. All of the deep learning based methods have at least than 96% Rank-1 accuracy with majority voting.