Explainable AI (XAI) is essential for fostering trust in the predictions of deep neural networks (DNNs), especially within the domain of image classification. SHAP (SHapley Additive exPlanations), a theoretically sound method rooted in game theory, stands as a prominent XAI technique for attributing feature importance. However, SHAP’s inherent computational complexity, which grows exponentially with the number of features, restricts its practical application. Although numerous approximation methods have been proposed, they often deviate from the original SHAP formulation, potentially sacrificing accuracy and theoretical fidelity. In this paper, we introduce CorrSHAP, a novel approach that leverages image superpixel correlations to significantly accelerate SHAP value estimation while preserving the rigor of the original formulation. CorrSHAP efficiently quantifies the interdependence of superpixels within an image, enabling the SHAP calculation to consider only interdependent superpixel combinations. This targeted approach significantly reduces the computational burden without compromising the faithfulness of the explanation. Evaluation results demonstrate that CorrSHAP outperforms the comparable Monte Carlo SHAP method in explanation faithfulness, achieving this in a fraction of a second, with a 55-fold speed improvement. The source code is available on the link (https://github.com/vhasic/CorrSHAP).
Scanning Electron Microscope (SEM) images play a crucial role in defect detection and analysis in the semiconductor manufacturing process. However, traditional deep learning methods for image classification act as black boxes, hindering expert trust and impeding model debugging and improvement. This paper tackles this problem by applying state-of-the-art explainability methods tailored to expert needs. In addition, we present Carinthia-S, the first expert-validated dataset providing ground truth segmentation masks for SEM images of semiconductor manufacturing defects, used as ground truth to evaluate the explainability methods. Our results show that our proposed automatic ground truth segmentation approach achieved impressive performance with 91.33% correctly segmented images, and that experts are satisfied with the provided explanations. The Carinthia-S dataset is made publicly available.