| Title |
Deep Learning based Pose-invariant Ear Recognition |
| Authors |
박현정(Hyeonjung Park) ; 남기표(GiPyo Nam) ; 김익재(Ig-Jae Kim) |
| DOI |
https://doi.org/10.5573/ieie.2019.56.8.47 |
| Keywords |
Ear recognition; Convolutional neural networks; K-ear database; Ensemble |
| Abstract |
Recently, ear recognition has been getting attention as a biometrics to identify an individual. According to significant enhancement of the recognition performance based on deep learning in the various fields of biometrics, a lot of deep learning-based approaches has been studied in ear recognition. However, there is a limit to improve the performance because of some issues such as the lack of large-scale ear databases and non-consideration of the characteristics for ears. To overcome these problems, this paper introduces a new database, called the K-Ear database, that can be used for research of ear recognition in unconstrained environment. The K-Ear database includes ear images of various environments such as pose changes from the front of a ear to 60°, illumination changes, and the partial occlusion by accessories. This paper also proposes a deep learning-based ear recognition model robust to pose variation by using the K-Ear database. To consider the characteristics of an ear, this paper performs the zero padding as a preprocessing on the input image. VGG-16 and ResNet50 model was utilized, to extract the feature vector. Finally, score level ensemble is performed to enhance the performance. Experimental results show that the proposed method is superior to the single model by more than 30% based on rank-1 accuracy in extreme pose variation. |