Biometric identification using panoramic dental radiographic images withfew-shot learning

Determining identity is a crucial task especially in the cases of mass disasters such as tsunamis, earthquakes, fires, epidemics, and in forensics. Although there are various studies in the literature on biometric identification from radiographic dental images, more research is still required. In this study, a panoramic dental radiographic (PDR) imagebased human identification system was developed using a customized deep convolutional neural network model in a few-shot learning scheme. The proposed model (PDR-net) was trained on 600 PDR images obtained from a total of 300 patients. As the PDR images of the patients were very different in terms of pose and intensity, they were first cropped by the domain experts according to the region of interest and adjusted to standard view with histogram equalization. A customized data augmentation approach was applied in order for the model to generalize better while it was being trained. The proposed model achieved a prediction accuracy of 84.72% and 97.91% in Rank-1 and Rank-10, respectively, by testing 144 PDR images of 72 patients that had not been previously used in training. It was concluded that well known similarity metrics such as Euclidean, Manhattan, Cosine, Pearson, Kendall?s Tau and sum of absolute difference can be utilized in few-shot learning. Moreover, Cosine and Pearson similarity achieved the highest Rank 1 score of 84.72%. It was observed that as the number of rank increased, the Spearman and Kendall?s Tau metrics had the same success as Cosine and Pearson. Based on the superimposed heatmap image analysis, it was determined that the maxillary, mandibular, nasal fossa, sinus and other bone forms in the mouth contributed biometric identification. It was also found that customized data augmentation parameters contributed positively to biometric identification.