CONICET | Buscador de Institutos y Recursos Humanos

Automated, non invasive ear detection in images andvideo is becoming increasingly required in several con-texts, including nonivasive biometric identification,biomedical analysis, forensics, and many others. Inbiometric recognition systems, fast and robust ear de-tection is a crucial step within the recognition pipeline.Existing approaches to ear detection are susceptibleto fail in the presence of typical everyday situationsthat prevent a crisp imaging of the ears, like partialocclusions, ear accessories, or uncontrolled cameraand illumination conditions. Even more, most of theproposed solutions work efficiently only within a pre-viously detected rectangular region of interest, whichlimits their applicability and lowers the accuracy ofthe overall detection. In this paper we evaluate the useof Convolutional Neural Networks (CNNs) togetherwith Geometric Morphometrics (GM) for automaticear detection in the presence of partial occlusions, anda Convex Hull algorithm for the ear area segmenta-tion. A CNN was trained with a set of ear imageslandmarked by experts using GM to achieve high con-sistency. After training, the CNN is able to detect earsover profile faces, even in the presence of partial oc-clusions. We analyze the performance of the proposedear detection and segmentation method over partiallyoccluded ear images using the CVL Dataset.

enviar mensaje