Ear Symmetry Evaluation on Selected Feature Extraction Algorithms in Ear Biometrics

Abstract

The human ear has an intriguing shape and like most parts of the human body, bilateral symmetry is observed between left and right.  Occlusions of the ear is a major problem in ear recognition, however, if ear symmetry is established, then reconstructing partially occluded ear images will be possible from the other ear, also the left ear of an individual’s test image can be matched against the right ear in the gallery database (or vice-versa). This paper presented an evaluation of the relationship between left and right ear using four selected feature extraction algorithms: Principal Component Analysis (PCA), Speeded Up Robust Features (SURF), Geometric feature extraction and Gabor wavelet based feature extraction techniques in terms of performance issues given by of False Acceptance Rate (FAR), False Rejection Rate (FRR), and Genuine Acceptance Rate (GAR).The approach was evaluated on non-public ear dataset and simulated in MATLAB Environment. For these selected feature extraction algorithms, the right ears of the subjects are used as the gallery, and the left ear as the probe. The experimental results suggest the existence of some degree of symmetry in the human ears but the ear are not exactly identical as the recognition accuracy of the system declined for three (PCA, SURF, and Gabor wavelet) of the feature extraction algorithms, FRR rising to over 84% for SURF. However, Geometric feature extraction reported relatively high recognition accuracy with FRR of 12.50% and GAR of 87.50%.

Keywords: Ear symmetry, Gabor wavelet, Occlusion, Principal Component Analysis (PCA), Speeded Up Robust Features (SURF).