Multimodal ear recognition using classical and deep learning approaches

Abstract

Authentication is crucial in computing. Traditional methods using newlinepasswords and tokens are insecure due to risks like loss and compromise. newlineBiometric security systems address these issues. The human ear is a reliable newlinebiometric for identification, unaffected by age or cosmetics. It offers newlineadvantages over fingerprints and faces, being easy to collect and process. newlineResearchers initially used 2D ear images but shifted to 3D for improved newlineaccuracy. Despite the ear being one of the fastest advancing biometric features, newlineit has some restrictions, such as being affected by posture, lighting, and scaling newlinein 2D images. To solve these problems, the researchers attention has shifted to newlinethe three-dimensional space where the inherent features of the 3D ear greatly newlineimproved identification accuracy. However, some studies reveal that using a newlinecombination of both 2D and 3D ear features provides more accurate results than newlineusing only 3D features. This thesis introduces three distinct approaches for ear newlinerecognition. It achieves superior accuracy on large datasets, outperforming newlineexisting methods, and enhances recognition by integrating multiple biometric newlinetraits. newlineThe first method proposed involves the utilization of diverse deep newlinelearning (DL) models, including LeNET, AlexNet, and Visual Geometry Group newline(VGG)-16, for the evolution of the ear recognition approach. newline