Fast learning ear detection for real-time surveillance

Abstract

Fully automated image segmentation is an essential step for designing automated identification systems. This paper investigates the problem of real-time image segmentation in the context of ear biometrics. The proposed approach is based on Haar features arranged in a cascaded Adaboost classifier. This method, widely known as Viola-Jones in the context of face detection, has a limitation of an extremely long training time, approximately a month. We efficiently implement a modified training / learning method, which significantly reduces training time. This approach is trained about 80 times faster than the original method, and achieves ~ 95% accuracy based on four different test sets (> 2000 profile images for app. 450 persons). The developed ear detection system is very fast and can be used in a real-time surveillance scenario. Experimental results show that the proposed ear detection is robust in the presence of partial occlusion, noise and multiple ears with various resolutions.