SHARPEN-CAM: efficient hierarchical SHAP-based visual explanation for deep convolutional neural networks

Deep learning models, particularly convolutional neural networks (CNNs), have achieved remarkable success in solving complex computer vision problems. However, their lack of interpretability poses challenges for adoption in high-stakes domains. To address this, numerous saliency methods have been proposed to generate visual explanations of CNNs’ decision processes. Existing methods often suffer from low explanation quality, high computational cost, or a lack of theoretical foundation. In this work, we introduce SHARPEN-CAM, a novel visual explanation method that adopts a hierarchical “reunion-redistribution” strategy to assign importance coefficients to activation maps. SHARPEN-CAM first reunites similar activation maps into groups and calculates SHAP values as group coefficients. It then redistributes these group coefficients to individual activation maps using Aumann-Shapley values derived from the integrated gradient method. Extensive experiments on ImageNet and VOC datasets demonstrate that SHARPEN-CAM outperforms state-of-the-art methods in terms of explanation quality, localization capability, and computational efficiency. By providing a clearer theoretical basis and practical implementation, SHARPEN-CAM offers a promising solution for interpreting deep neural networks, enhancing their transparency and trustworthiness.