Ensemble-Based Hybrid Transfer Approach for an Effective 2D Ear Recognition System

Abstract

Person identification using ear images has gained significant attention recently. Transfer learning provides an effective platform for image classification, utilizing CNNs like AlexNet, ResNet, VGG16, and VGG19, which are fine-tuned for specific applications. Combining transfer learning with support vector machines (SVM) enhances people recognition via ear images. This paper integrates a hybrid transfer learning model with an ensemble technique to improve recognition accuracy. We use pre-trained CNN models, VGG16 and VGG19, for feature extraction and replace the fully connected layer with an SVM classifier. Using the SoftMax activation function, each model generates a probabilistic output, which is averaged for classification. The proposed ensemble model was validated on two datasets with variations in pose, illumination, and rotation. Simulation results show that the ensemble-based transfer learning approach outperforms its two anchor models and competes with state-of-the-art ear recognition techniques.

Keywords

• Ensemble learning
• transfer learning
• feature extraction
• classification
• accuracy