Grad-CAM++ 기반 보조 히트맵 입력을 활용한 경량 얼굴 인식 모델의 강건성 향상

초록

This study proposes a method to enhance lightweight face recognition models by integrating Grad-CAM++–based Heatmaps as auxiliary input channels. Without modifying the backbone architecture, the proposed structure explicitly guides visual attention through automatically generated Grad-CAM++ Heatmaps. This guidance accelerates early convergence and improves macro F1 under occlusion and brightness variations. A Zero/Random/Shuffle ablation experiment verified that the Heatmap channel is a causal factor in the performance improvement, confirming the robustness of the proposed approach. Overall, the Heatmap input enhances model interpretability and visual focus, demonstrating its practicality for deployment in lightweight face recognition systems.

키워드