Human Identification with Dental Panoramic Images Based on Deep Learning

Human identification by means of dental panoramic X-ray images has been achieved due to end-to-end deep learning using a convolutional neural network. This paper proposes a novel attention-based multi-supervision network (AMNet) for human identification. AMNet includes an attention-based supervision mechanism to obtain an accurate attention mask, feature fusion branches to combine multilevel global features, and part feature branches to obtain discriminative local features. After extracting features of dental panoramic X-ray images, matching scores between the gallery and query features are calculated with cosine similarity to determine if the query image and gallery image are from the same identity. Our training dataset has 22,172 images from 9490 subjects. The proposed method achieved 88.72% rank-1 accuracy and 95.79% rank-5 accuracy on the query set with 665 dental panoramic X-ray images from 503 different subjects.