Year: 2025
Type: article
Abstract: Achieving robust multimodal biometric identification requires advanced feature extraction strategies and effective integration of diverse data modalities. Conventional methods often encounter limitations such as computational complexity and degradation of critical information during feature transformation. Although deep learning models address feature extraction challenges, their heavy architectures hinder real-world deployment. Moreover, traditional fusion strategies, based mainly on simple concatenation, overlook critical intermodal correlations, leading to suboptimal recognition accuracy. In this study, we propose a lightweight Gabor Attention Network framework designed for efficient multimodal biometric recognition. Our approach utilizes learnable Gabor filters to capture detailed local and directional features with enhanced precision and reduced computational burden compared to standard convolutions. We further introduce a convolutional attention mechanism that adaptively refines intermediate feature representations, and a novel attention-driven fusion architecture that dynamically models and exploits intermodal dependencies. Extensive experiments on two multimodal datasets demonstrate that our model achieves superior performance compared to several state-of-the-art methods, attaining up to 99.49% accuracy and 0.35% Equal Error Rate, while maintaining high efficiency with only 10.6M parameters, 0.85 GFLOPs, and 60 FPS inference speed. These results highlight the effectiveness of our biologically inspired and attention-enhanced design for achieving high-accuracy, low-complexity multimodal biometric identification.
Authors: Phan Minh Than, Hoanh Nguyen
Institution: Industrial University of Ho Chi Minh City
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7160962067
