Biometric Recognition using Deep Learning

Abstract

Inforensicscience,security,identity,andverification,biometricsarees- sential. Over the past three decades, research has concentrated on develop- ing dependable systems with human behavioural and biological characteris- tics, each with its own advantages and difficulties. Thehumanearisemergingasapromisingbiometricmodalitydueto its unique characteristics and advantages over other methods like face, fin- gerprint, or iris scanning.Ear biometrics offers a seamless process without requiringphysicalmovementsfromusers. Recentyearshaveseenincreased attention and progress in ear biometrics, covering detection, preparation, feature extraction, verification, and identification. Taking advantage of this great development in deep learning due to improvements in computer pro- cessing power, data availability, and innovative algorithms.More specifi- cally, convolutional neural networks (CNNs), which have achieved remark- able success in areas such as computer vision tasks. We conducted a comprehensive experimental study introducing novel methodologies to improve ear identification.Two main contributions are devoted in this work; the first one introduces a practical approach called Mean-Class Activation Maps with CNNs (Mean-CAM-CNN) to tackle is- sues related to image classification by focusing on discriminative regionsof ear images.The Mean-CAM framework addresses intra-class variability by using a guided mask to crop relevant areas based on mean heat maps. This croppedregion is then utilised for traininga CNN, thereforeenhancing discriminative classification performance.In the second contribution, we aimed to combine deep convolutional generative adversarial network DC- GANmodelwithMean-CAMtechnique,whereDCGANmodelusedaspre- processingsteptocolouriseandenhancetheearimagesoftheuseddatasets. The proposed approach was evaluated on two ear recognition datasets, AMI andAWE,showingsignificantimprovementswithRank-1recognitionrates of 100 % for AMI and 76.25 % for AWE