A fine-grained network for human identification using panoramic dental images.

Area of Science:

• Forensic Science
• Computer Vision
• Artificial Intelligence

Background:

• Panoramic dental images are crucial for identifying unknown individuals in forensic investigations.
• Deep learning models have shown promise but often overlook the significance of tooth contours.
• Existing methods lack specific architectures to integrate tooth contour information effectively.

Purpose of the Study:

• To develop a fine-grained human identification model using deep learning that specifically incorporates tooth contour information.
• To enhance the accuracy of identifying individuals from panoramic dental X-rays by focusing on subtle dental features.
• To address the limitations of current deep learning approaches in utilizing tooth-specific details for identification.

Main Methods:

• A bilateral branched architecture was designed, with separate branches for image and mask feature extraction.
• An elementwise reweighting mechanism integrated image and mask features.
• An improved attention mechanism focused on informative regions, and ArcFace loss was enhanced for hard samples.

Main Results:

• The model achieved an average rank-1 accuracy of 88.62% and rank-10 accuracy of 96.16%.
• Performance was validated on a large dataset of 23,715 panoramic dental X-ray images from 10,113 patients.
• The approach demonstrated superior performance in distinguishing individuals based on fine-grained dental features.

Conclusions:

• The proposed deep learning model effectively utilizes tooth masks and contours for accurate human identification from dental radiographs.
• The novel architecture and enhanced loss function significantly improve fine-grained identification capabilities.
• This method offers a promising advancement for forensic odontology and identification processes.