Enhancing the Robustness of Adaptive Class Activation Mapping (AD-CAM) Against Noisy Facial Expression Data Using Preprocessing and Adaptive Normalization

In real-world computer vision applications, visual data is often corrupted by noise, reducing both the accuracy and interpretability of deep learning models. This study proposes an enhanced AD-CAM framework that integrates noise-aware preprocessing and adaptive normalization to improve robustness in both prediction and visual explanation. Experiments were conducted on the FER2013 facial expression dataset augmented with Gaussian, salt-and-pepper, and speckle noise. Using ResNet-50 as the backbone, the proposed method demonstrated significant gains across multiple evaluation metrics, including Robust Accuracy (RA), Drop Coherence (DC), Area Under Robustness Curve (AURC), and Signal-to-Noise Ratio (SNR). Compared to the baseline, the model achieved over 10% accuracy improvement and up to 0.16 DC reduction under noise. Qualitative visualizations showed that the improved model consistently highlighted semantically relevant facial regions, maintaining interpretability even under severe input degradation. These results support the adoption of noise-aware interpretability frameworks for more reliable and trustworthy deployment in real-world vision systems.