Toward Personal Identification Using Multi-Angle-Captured Ear Images: A Feasibility Study

The ear is an effective biometric feature for personal identification. Although numerous studies have attempted personal identification using frontal-view images of the ear, only a few have attempted personal identification using multi-angle-captured ear images. To expand the extant literature and facilitate future biometric authentication technologies, we explore the feasibility of personal identification using multidirectionally captured ear images and attempted to identify the direction-independent feature points that contribute to the identification process. First, we construct a convolutional neural network model for personal identification based on multi-angle-captured ear images, after which we conduct identification experiments. We obtained high identification accuracies, exceeding 0.980 for all the evaluation metrics, confirming the feasibility of personal identification using multi-angle-captured ear images. Further, we performed Gradient-weighted Class Activation Mapping to visualize the feature points that contribute to the identification process, identifying the helix region of the ear as a key feature point. Notably, the contribution ratios for ear images in which the inner ear was visible and not visible are 97.5% and 56.0%, respectively. These findings indicate the feasibility of implementing personal identification using multi-angle-captured ear images for applications, such as surveillance and access control systems. These findings will promote the development of future biometric authentication technologies.