Ear biometrics in forensic identification: from ear similarity quantification to kinship verification driven by deep learning approaches.

Abstract

Kinship verification using biometric traits is crucial for finding missing children, rapid forensic identification, and social media analysis. Ear biometrics is gaining attention due to its uniqueness, permanence, and non-intrusiveness. However, current research on ear-based kinship verification is limited, and the factors associated with ear similarities across kinship-related individuals require to be evaluated. To fill this gap, our study developed deep learning models to quantify the similarity between ear images and conduct the kinship verification task. Two ear image datasets, namely SCED and CNKE, were collected from Chinese subjects. The SimiNet model, which utilized a pre-trained ResNet50 as its backbone, was constructed to evaluate the cosine similarity between ear image pairs. For kinship verification, the VTrans model was established by combining a pre-trained VGG16 with Transformer modules. Based on the cosine similarity, the SimiNet model obtained 93.53% accuracy and an AUC (Area Under the Curve) of 0.98 for personal verification. The similarity analysis further revealed that kinship-related individuals of the same sex displayed higher ear similarity scores. The VTrans model attained 71.17% accuracy and an AUC of 0.76 on the CNKE dataset. Heatmaps revealed that the VTrans model focused on the helix and the upper half of the ear during kinship verification. The model code has been provided in Github ( https://anonymous.4open.science/r/SimNet_VTrans-EB41 ) to facilitate refinement in future research. Our study has successfully established deep learning models for ear similarity quantification and kinship verification, providing effective tools for biometric forensic identification.