Image Quality Assessment for Effective Ear Recognition

Abstract

Due to the recent challenges in access control, surveillance and security, there is an increased need for efficient human authentication solutions. Ear recognition is appealing, where the data acquisition procedure is contactless, non-intrusive, and covert. This article proposes a deep learning-based solution for effective ear recognition. We explore multiple strategies to enhance learning by using alternative ear datasets, with a wide range of ear poses. We investigate the performance of the proposed deep ear models in the presence of various image artifacts, which commonly occur in real-life recognition applications. Thus, this study focuses on identifying the robustness of the proposed ear recognition models in controlled and uncontrolled conditions (dataset dependent). In that direction, we propose using our efficient automatic ear image quality assessment model, which is designed to guide the proposed ear recognition system solution. By performing a set of ear-based recognition experiments on extended degraded ear datasets, we determined that the employment of the proposed automatic ear image quality assessment model improves ear identification performance from 58.72% to 97.25% for the USTB degraded dataset and from 45.80% to 75.11% for the degraded FERET dataset.