Automatic Human Identification from Panoramic Dental Radiographs Based on Deep Learning

The ability to accurately and efficiently identify individuals using panoramic dental radiographs (PDRs) is crucial in human identification, particularly in scenarios such as large-scale disasters and criminal investigations. However, this task poses significant challenges due to the complex anatomical structures of teeth, variations in image quality, and the need for scalable data processing methods. To address these issues, we conducted a study using a dataset of 16,383 PDR images from 2,653 individuals, applying a standardized preprocessing pipeline that included medial filter, contrast enhancement to improve input data consistency. We evaluated the performance of multiple deep learning architectures, including ResNet50, VGG16, and InceptionResNetV2. Our findings revealed that InceptionResNetV2, combined with the triplet loss function, outperformed other models. With input images resized to 299×299 for feature extraction, this model achieved remarkable results, including 99.81 % training accuracy, 99.71 % validation accuracy, and 99.69 % test accuracy. The best validation loss was 0.0121, and the test loss was 0.0134. Additionally, we developed an automated recognition software system based on this model, capable of processing each image in just 33 milliseconds on average. This system offers a fast and reliable tool for human identification, with promising applications in forensic science, disaster victim identification, and personalized medical record management.