DEEP TRANSFER LEARNING FOR EAR RECOGNITION

Abstract

Today, there is increasing talk of cross-sectoral insecurity, rising crime, and piracy. Moreover, the mobility of people, financial services transactions, and access to services require an urgent need to ensure the identity of individuals. Traditional security systems rely on previously acquired knowledge (PIN codes, passwords) or token-based access (keys, identifiers, badges). However, these systems are less reliable in many environments, as they are often unable to distinguish between truly authorized people and fraudsters. In this case, we selected one of these systems to study, which is a deep learning ear recognition system, or more precisely, a system that uses the human ear as a biometric. This system, it’s hard to copy. There are many advantages, such as ease of use and low cost. Our work can be seen as a two-stage process. Firstly, the data augmentation using different geometrical techniques is incorporated to overcome the lack of training samples required for training the deep learning model. Secondly, the feature extraction and classification task is performed through the four CNN algorithms to verify the person’s identity. AMI dataset is utilized to test and evaluate the proposed model’s performance. Our proposed method for the AMI database achieved an accuracy of 90% with Vgg16 and 92.22 % with Vgg19 and 91.11% with the exception model and 94 % with MobilenetV2. Experimental results conclude that the proposed work obtained good performance compared to existing methods.