Performance evaluation of enhanced deep learning classifiers for person identification and gender classification

Area of Science:

• Computer Science
• Biometrics
• Artificial Intelligence

Background:

• Periocular biometrics is crucial for person authentication.
• Existing systems face challenges in accuracy, overfitting, and computational efficiency.
• Deep learning offers potential but requires specialized architectures for periocular recognition.

Purpose of the Study:

• To propose an enhanced deep learning classifier (EDLC) paradigm for person identification and gender classification using periocular images.
• To address limitations of current biometric systems, including accuracy and efficiency.
• To introduce novel methods for region of interest (ROI) extraction and feature representation.

Main Methods:

• A novel Hexagon-shaped ROI extraction technique for periocular regions.
• Feature extraction using the Laplacian transform.
• Implementation of three custom EDLCs: dilated axial attention CNN, self-spectral attention-based relational transformer net (SSA-RTNet), and parameterized hypercomplex convolutional Siamese network.
• Hyperparameter optimization using an adaptive coati optimization algorithm.

Main Results:

• SSA-RTNet achieved maximum accuracies of 99.8% (UBIPr) and 99.67% (UFPR) for person identification.
• SSA-RTNet obtained accuracies of 98.4% (UBIPr) and 99.68% (UFPR) for gender classification.
• The enhanced models demonstrated considerable improvements over existing competitive models.

Conclusions:

• The proposed EDLC paradigm, particularly SSA-RTNet, offers a significant advancement in periocular biometric systems.
• The novel ROI extraction and feature extraction methods contribute to improved performance.
• The study highlights the potential of advanced deep learning techniques for robust person identification and gender classification.