Segmentation and Normalization of Human Ears Using Cascaded Pose Regression

Being an emerging biometric characteristic, automated ear recognition is making its way into forensic image analysis for law enforcement in the last decades. One of the most important challenges for this application is to deal with loosely constrained acquisition scenarios and large databases of reference samples. The research community has come up with a variety of feature extraction methods that are capable of handling occlusions and blur. However, these methods require the images to be geometrically normalized, which is mostly done manually at the moment.

In this work, we propose a segmentation and normalization method for ear images that is using cascaded pose regression (CPR). We show that CPR returns accurate rotation and scale estimates, even for full profile images, where the ear has not been segmented yet. We show that the segmentation accuracy of CPR outperforms state of the art detection methods and that CPR improves the recognition rate of an ear recognition system that uses state of the art appearance features.