MorphoSHAP: Rethinking the Unit of Attribution in Explanation for Deep Visual Models

Redefining cooperative Shapley game attribution from raw pixels to hierarchical morphological structures for real-time model auditing.
Explainable AI
Computer Vision
Mathematical Morphology
Shapley Values
Author

Anirudh Prabhakaran

Published

October 1, 2026

Read the full preprint on arXiv.

Overview

Traditional explainable AI (XAI) methods for computer vision often struggle with noise and computational latency, particularly when attributing importance at the raw pixel or arbitrary superpixel level.

Working under the guidance of Prof. Gianni Franchi at ENSTA, IP Paris, I architected MorphoSHAP, a novel explainable AI framework. MorphoSHAP redefines cooperative Shapley game attribution by moving away from pixels and instead leveraging hierarchical morphological structures via the Tree of Shapes.

By formulating a structural attribution vocabulary that characterizes “shape players” by their scale and signed values, this method unifies spatial saliency with concept-based textual explanations.

Key Results

  • Achieved State-of-the-Art (SOTA) Insertion/Deletion AUC across 5 standard benchmarks.
  • Accelerated explanation latency by 6x over standard superpixel KernelSHAP (achieving 0.13s/image).
  • Made real-time production model introspection and dataset-level performance monitoring highly viable.