Automatic segmentation of human ear in the wild

Abstract

Ear biometrics has been a challenging and distinctive research area in recent times. The human ear possesses unique promising attributes that are being used by the researchers to carry out significant improvements in the field of human recognition using ear as a biometric. In order to achieve efficiency on any ear biometric system, the detection and segmentation of the human ear need to be performed precisely. Feeding accurately segmented images to the recognition system will result in higher recognition accuracy. In this paper, we present our work of segmentation of human ears from the images captured in unconstrained environment by employing the U-Net architecture on our own dataset and presented the results of ear segmentation. The U-Net model is also tested on the annotated web ears (AWE) segmentation dataset. We obtained 92.38% accuracy and 79.33% intersection over union (IoU) on the test data on our own dataset and 76.2% IoU on AWE segmentation dataset.