An augmentation-enhanced transformer model for forensic identification using panoramic dental x-rays

Forensic investigations depend on accurate human identification which requires use of panoramic dental radiographs because traditional biometric methods do not work. Deep learning methods fail to provide accurate results because they lack sufficient training data which makes it difficult to identify specific features and they cannot handle changes that happen after death. The study presents DentID-Net which functions as an anatomical identification system that employs dental identification methods from both living and deceased individuals. The system combines three interdependent elements which include (i) DentAug-Min which implements a morphology-based enhancement method for handling limited data situations through its capability to decrease overfitting risk at small sample sizes, (ii) The system uses dual-branch encoder which enables modeling both detailed dental structure and complete jaw shape through convolutional and transformer-based systems and (iii) IDMatch-Caps which operates as a capsule-based identity matcher that uses morphological consistency loss to maintain anatomical connections throughout identity assessment. DentID-Net achieved cross-validation results through five folds which showed 98.9% mean accuracy, 99.2% specificity and 0.999 AUC results which exceeded the performance of recent forensic dental identification systems. The results show that the system can generalize to new situations while maintaining low false match rates and operating efficiently in forensic laboratory environments. The developed system creates a strong yet understandable system which enables automatic identification of people through dental methods under actual forensic conditions.