Advancing Biometric Identity Recognition with Optimized Deep Convolutional Neural Networks

Abstract

Biometric identity recognition, capitalizing on unique physical attributes, represents an increasingly explored research field within the biometrics community, with implications spanning surveillance, crowd analytics, automated identity checks, and user device access.Ear images, in particular, offer a robust data source for devising effective personal identification systems.The biometric field has seen a surge in the application of machine learning algorithms, specifically deep neural network architectures such as Convolutional Neural Networks (CNNs) and transfer learning methods, to enhance ear recognition systems.This study evaluates leading deep CNN architectures -ResNet, DenseNet, MobileNet, and Inception -for their efficacy in creating ear recognition systems resilient to varying imaging conditions.The AMI and WPUT datasets, publicly accessible ear image datasets, were utilized to train and assess the proposed models.The models demonstrated substantial success, achieving rank-1 accuracies of 96% and 83% on the AMI and WPUT datasets, respectively.Additionally, the Gradient-weighted Class Activation Mapping (Grad-CAM) visualization technique was employed to elucidate the models' decision-making processes, revealing a reliance on auxiliary features like hair, cheek, or neck when available.The use of Grad-CAM not only enhances understanding of the decision-making processes within the CNNs but also highlights potential areas of improvement for the proposed ear recognition systems.