DANNET: deep attention neural network for efficient ear identification in biometrics.

Area of Science:

• Computer Science
• Biometrics
• Artificial Intelligence

Background:

• Mask-wearing due to COVID-19 necessitates reliable biometric identification beyond facial recognition.
• Existing ear biometrics using convolutional neural networks (CNNs) face challenges in accuracy and efficiency.
• Ear biometrics are crucial for identification when facial features are obscured.

Purpose of the Study:

• To propose a novel deep learning method for accurate ear biometric identification.
• To enhance ear detection and segmentation precision using an encoder-decoder architecture with attention mechanisms.
• To address the limitations of current biometric systems in scenarios with partial facial occlusion.

Main Methods:

• Developed an encoder-decoder deep learning ensemble technique incorporating attention blocks.
• Employed an ensemble of two YSegNets for improved ear segmentation performance.
• Validated the method using combined datasets: EarVN1.0, AMI, and Human Face.

Main Results:

• Achieved a segmentation framework accuracy of 98.93%.
• The ensemble approach demonstrated superior performance compared to a single YSegNet.
• The method proved effective in robustly segmenting ear images, reducing false positives and negatives.

Conclusions:

• The proposed encoder-decoder deep learning ensemble technique offers a highly accurate solution for ear biometric identification.
• This method is particularly valuable for individual recognition in large gatherings and public spaces.
• The research provides a viable biometric identification solution for mask-wearing and other facial obstruction scenarios.