The State of Post-Hoc Local XAI Techniques for Image Processing: Challenges and Motivations

TL;DR

The paper surveys post-hoc local explainability techniques for image processing, clarifying key XAI terminology and mapping motivations to challenges. It reviews local methods such as ICE, counterfactuals, LIME, and SHAP, and discusses regulatory and industry drivers for explainability. It highlights major open problems—lack of formalism, explanation interpretability, causal explanations, and evaluation metrics—and proposes future directions including multi-modal explanations and richer ground-truth alignment. The work emphasizes the practical importance of trustworthy explanations for safety-critical AI systems and offers a roadmap for building more reliable, stakeholder-aware XAI tools.

Abstract

As complex AI systems further prove to be an integral part of our lives, a persistent and critical problem is the underlying black-box nature of such products and systems. In pursuit of productivity enhancements, one must not forget the need for various technology to boost the overall trustworthiness of such AI systems. One example, which is studied extensively in this work, is the domain of Explainable Artificial Intelligence (XAI). Research works in this scope are centred around the objective of making AI systems more transparent and interpretable, to further boost reliability and trust in using them. In this work, we discuss the various motivation for XAI and its approaches, the underlying challenges that XAI faces, and some open problems that we believe deserve further efforts to look into. We also provide a brief discussion of various XAI approaches for image processing, and finally discuss some future directions, to hopefully express and motivate the positive development of the XAI research space.

Figures (8)

• Figure 1: ICE: a) Understanding why the purple point is assigned to the negative class by considering what would happen if either feature $x_{1}$ (cyan line) or feature $x_{2}$ (purple line) was changed. b) The effect of changing feature $x_{1}$: This results in the point being classified as positive, if feature $x_{1}$ had a higher value. c) The effect of changing feature $x_{2}$: This results in the point being classified as positive if feature $x_{2}$ had a lower value. ice
• Figure 2: Counterfactual Explanations: a) Find out the explanation for which the data point (purple point 1) was classified negatively to determine what needs to be changed to yield the correct classification (cyan point 2). b) Seeking sparse counterfactual examples where only a few features (i.e just feature $x_{1}$ are changed. c) There may be multiple potential ways to modify the input (brown point 1) to change the classification points. ice
• Figure 3: Left: Image of a bowl of bread. Middle and right: LIME explanations for the top 2 classes (bagel, strawberry) for image classification made by Google's Inception V3 neural network. limepage
• Figure 4: Illustration of Shapley Values
• Figure 5: Illustration of SHAP Explanation shappart