Palatal Rugoscopy for Biometric Identification based on ResNet50 Feature Extraction and Machine Learning Classifiers

The Palatal Rugoscopy have highly stable and unique morphological aspects and can be used in odontology as a ‘fingerprint'’ of the mouth for reliable human identification in Forensic Dentistry. Though AI is increasingly being used in biometrics, the need for local datasets, especially for the Iraqi population is stark from College of Dentistry Teaching Hospital at the University of Tikrit. In addition, deep learning on newly collected small medical dataset results in significant overfitting and computational burden. Based on this, this study aims to present a lightweight hybrid ML-based system customized for a new manually curated dataset of palatal rugoscopy from Iraq population. The pipeline only retrains the parameters of the ResNet50 backbone for deep feature extraction, without retraining the whole model, which can be beneficial due to the number of parameters. High dimensional features are extracted and then classified using robust classifiers such as SVM, K-NN, ANN and the model generalization was guaranteed through targeted geometric and radiometric data augmentation. It is observed that SVM has the highest overall performance which yields 98.00% accuracy, 97.00% precision, 98.00% recall, and 97.33% F1-Score among the tested classifiers, even with a small number of initial samples. This first-of-its-kind study in the region introduces a highly efficient, computationally lightweight proof-of-concept for forensic identification which can be used to develop scalable, automated biometric systems in forensic Dentistry.