Ensembles of Deep Learning Models and Transfer Learning for Ear Recognition

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Visual recognition systems rely on effective feature extraction.
• Convolutional Neural Networks (CNNs) excel at capturing deep visual features like appearance, color, and texture.
• Ear recognition is a challenging biometric task requiring robust feature representation.

Purpose of the Study:

• To develop a novel ear recognition system using ensembles of deep CNNs.
• To leverage Visual Geometry Group (VGG)-like architectures for discriminative feature extraction from ear images.
• To enhance recognition performance by combining multiple deep learning models.

Main Methods:

• Training various depth CNNs on ear images with random initialization.
• Evaluating pretrained CNN models as feature extractors and fine-tuning them.
• Constructing ensembles of the best-performing CNN models.
• Testing the system on controlled and uncontrolled ear image datasets (AMI, AMIC, WPUT).

Main Results:

• The proposed ensembles of deep CNN models achieved state-of-the-art performance in ear recognition.
• Significant improvements were observed compared to recently published results.
• The system demonstrated effectiveness across datasets with varying image conditions.
• Visual explanations highlighted regions used by models for decision-making.

Conclusions:

• Ensembles of deep CNNs, particularly VGG-like architectures, offer a powerful approach for robust ear recognition.
• The proposed method significantly advances the state-of-the-art in biometric identification.
• The system's ability to provide visual explanations enhances interpretability and trust.