Enhancing Occluded Face Recognition Through Device-Edge Collaboration and Cross-Domain Feature Fusion with Heatmaps

Face recognition under occlusion presents a persistent challenge in computer vision, primarily due to difficulties in capturing and effectively integrating visible and obscured facial features. This paper introduces a novel device-edge collaborative framework that advances occluded face recognition through an innovative cross-domain feature fusion technique. Our method employs a software-generated, occlusion-aware heatmap created from standard visible-light images, which emphasizes key facial landmarks even when regions of the face are partially covered. This heatmap provides a complementary data layer that enhances traditional visible facial features, allowing the cross-domain feature fusion process to synthesize a more comprehensive set of features for reliable recognition under occlusion. The combined feature set leverages both visible and simulated heatmap information to improve model robustness, even in complex occlusion scenarios. Additionally, a delay-optimized task scheduling method is applied on the edge, ensuring efficient resource allocation, faster processing, and minimal latency. Although this project is ongoing, preliminary evaluations indicate the potential effectiveness of this approach, with expected outcomes to include high recognition accuracy and reduced latency. This framework provides a scalable solution for occluded face recognition by integrating complementary data sources without requiring specialized hardware.