Feature-granularity-based spurious correlation and causal analysis in CNNs Exploring Feature Granularity to Uncover Spurious Correlations

This study presents a novel method for analyzing object correlations in convolutional neural network (CNN) explanation maps, addressing limitations in current approaches. Our method integrates foreground-background separation, superpixel segmentation, and linear regression modeling to quantify the contribution of objects within an image. By employing F-tests and p-value calculations, the method effectively identifies and spurious correlations. Experimental results demonstrate that this approach not only improves the efficiency and stability of the analysis but also enhances the interpretability of CNN models, providing a more robust foundation for the reliability of image classification models. Additionally, by constructing aCorrelation network based on the analysis results, the study facilitates the visualization of the decision-making processes, offering significant support for the practical application and deployment of CNNs.