Multiscale Dynamic Graph Representation for Biometric Recognition With Occlusions.

Area of Science:

• Computer Science
• Biometrics
• Artificial Intelligence

Background:

• Occlusion significantly degrades the performance of Convolutional Neural Networks (CNNs) in biometric recognition systems.
• Existing methods struggle to generalize effectively when faced with occluded biometric data.

Purpose of the Study:

• To develop a unified framework that integrates CNNs and graph models to address occlusion challenges in biometric recognition.
• To enhance the robustness and accuracy of biometric systems in real-world scenarios with occlusions.

Main Methods:

• Proposed a novel multiscale dynamic graph representation (MS-DGR) framework.
• Recrafted deep features into a feature graph (FG) where nodes represent local regions and edges indicate non-occluded co-occurrences.
• Utilized dynamic graph matching to discard occluded regions based on node similarity and topological structure.
• Incorporated a multiscale strategy for diverse region representation.

Main Results:

• The MS-DGR framework effectively identifies and discards occluded regions in biometric samples.
• Demonstrated significant improvements in recognition accuracy in both natural and simulated occlusion scenarios.
• Achieved superior performance compared to baseline methods in extensive experiments.

Conclusions:

• The proposed MS-DGR framework offers a robust solution for occlusion problems in biometric recognition.
• The integration of graph models and CNNs with a multiscale approach enhances system accuracy and generalization.
• The framework provides illustrative inference by visualizing paired nodes, aiding in understanding occlusion handling.