End-to-End Pixel-Wise Ear Segmentation with U-Net and ResNet-50 Encoder

Ear biometrics has emerged as a promising alternative in biometric recognition systems, offering robustness in unconstrained environments where traditional modalities such as face recognition may fail on its own, but can be enhanced by ear. Ear segmentation, in particular, plays a crucial role in downstream recognition by isolating discriminative ear regions and reducing background interference. However, existing approaches to ear detection and segmentation are commonly susceptible to severe occlusions, ear accessories, and variable illumination, and their performance deteriorates on images captured in the wild. To address these limitations, we introduce a tailored ear-segmentation architecture based on a U-Net with a ResNet-50 encoder. Trained and validated on the Annotated Web Ears (AWE) dataset, our method achieves a mean Intersection over Union (IoU) of 77.1% and a pixel-wise accuracy of 99.7%, outperforming the Convolutional Encoder--Decoder (CED) baseline. We further evaluate on the EarSegDB-25 dataset, where our approach attains a test-set IoU of 94.76%, significantly surpassing previous ear segmentation methods based on the original U-Net architecture. High pixel-wise accuracy across methods is largely attributable to background dominance; in contrast, the improved IoU achieved by our approach more accurately reflects gains in ear region segmentation performance. Leveraging a ResNet-50 encoder, our model demonstrates robust performance under occlusion and illumination challenges, achieving state-of-the-art results on AWE and EarSegDB-25 and showing strong potential for biometric applications in unconstrained environments.