DHI-GAN

DHI-GAN implements a semisupervised generative adversarial network framework for dental-based human identification that enhances identity discrimination under limited sample sizes.

Key Features:

• Semisupervised GAN with integrated classifier: Employs a GAN that jointly generates samples and classifies identities in a "classifying while generating" paradigm.
• Identity embedding-guided architecture: Incorporates identity embeddings to preserve individual-specific informative features during generation.
• Parallel spatial and channel fusion attention block: Applies a parallel spatial and channel fusion attention mechanism to focus on regional details and abstract features and integrates this block into the classifier.
• Loss function optimization (ArcFace + focal loss): Combines ArcFace loss and focal loss to mitigate small-sample problems and improve discrimination.
• Sample control parameters: Introduces two parameters to regulate generation and selection of samples so that high-quality, informative samples are fed into the classifier.

Scientific Applications:

• Dental-based human identification: Improves identification accuracy on dental data, achieving a reported top-one accuracy of 92.5% on real-world datasets.
• Forensic and low-sample scenarios: Reduces required training samples and supports identification in cases where data are scarce or difficult to obtain.
• Model adaptation: Provides a GAN-based semisupervised training strategy that can be incorporated into other classification models in bioinformatics and forensic science.

Methodology:

Semisupervised GAN with an additional classifier trained under a classifying-while-generating paradigm, using identity embedding guidance, a parallel spatial and channel fusion attention block integrated into the classifier, optimization with ArcFace and focal loss, and two sample-control parameters to regulate generated-sample inclusion.