DXAI: Explaining Classification by Image Decomposition

Novel decomposition approach: Rather than traditional heatmaps that highlight important regions, DXAI separates images into two complementary parts—"class-agnostic" features (information irrelevant to classification) and "class-distinct" features (information the classifier uses)—that mathematically sum back to the original image.

Better for global features: The method is particularly effective for explaining classifications based on distributed characteristics like color or texture, rather than just localized objects, addressing a limitation of existing explanation techniques that focus on specific image regions.

Interpretability through reconstruction: By decomposing images into constituent parts that can be recombined, users can directly see what visual information the classifier is actually using versus ignoring, providing a more intuitive understanding of model decision-making than traditional explanation heatmaps.

Grounded in signal processing principles: The approach follows established analysis-and-synthesis methods from signal processing, meaning the decomposition is mathematically principled rather than ad-hoc, potentially making explanations more reliable and reproducible.

Practical availability: The authors have released open-source code, enabling practitioners and researchers to apply this explanation method to their own image classification systems and compare it against existing approaches.