A subscription to JoVE is required to view this content. Sign in or start your free trial.
Review Article
This review synthesizes explainable AI (XAI) methods for visual inspection in industrial and medical domains. A PRISMA-guided search identified 75 studies, informing a domain-specific taxonomy and comparative analysis using standardized metrics. We contribute an evidence-backed taxonomy and a practitioner-oriented workflow for tool selection.
Explainable Artificial Intelligence (XAI) techniques enable transparency and trust in automated visual inspection systems by making black-box machine learning models understandable. While XAI has been widely applied, prior reviews have not addressed the specific demands of industrial and medical inspection tasks. This paper reviews studies applying XAI techniques to visual inspection across industrial and medical domains. A systematic search was conducted in IEEE Xplore, Scopus, PubMed, arXiv, and Web of Science for studies published between 2014 and 2025, with inclusion criteria requiring the application of XAI in inspection tasks using public or domain-specific datasets. From an initial pool of studies, 75 were included and categorized into post-hoc and intrinsic, which were then evaluated with respect to fidelity, robustness, complexity, and localization accuracy. Results show that gradient- and propagation-based methods offer efficient visual explanations suitable for near real-time inspection, though with coarse localization, while perturbation-based and surrogate-model methods provide more detailed attributions at higher computational cost but with reduced robustness. In addition, prototype-based networks and self-attention architectures illustrate trade-offs between interpretability and predictive performance. Selecting the most effective XAI method is not one-size-fits-all; it depends on the dataset, latency, and interpretability needs. This review introduces a unified taxonomy of XAI methods for visual inspection, compares different approaches in both industrial and medical domains using standardized metrics, and proposes a task-based selection workflow to guide practitioners in choosing the optimal method. Compared to prior reviews, this work offers a cross-domain comparative analysis grounded in quantitative benchmarks and outlines directions for standardized evaluation and user-centered validation.
It is essential for industries such as manufacturing, automotive, food packaging, pharmaceuticals, and healthcare to verify that products and systems are functioning adequately and safely. Automation and machine learning tools, which go beyond traditional visual inspection performed by human operators or rudimentary camera-based systems, improve accuracy, production speed, and consistency1. AI inspection allows real-time monitoring2 and defect detection in high-throughput environments3, reducing human error and operational costs4. The use of artificial intelligence, particu....
Access restricted. Please log in or start a trial to view this content.
2. XAI framework
Most state-of-the-art AI models-particularly deep neural networks-are often treated as black boxes; there is little visibility into how they make decisions12. That lack of transparency is a barrier to trust and explainability in diverse applications. As a result, XAI has become an indispensable part of trustworthy AI. No single approach fits all scenarios: some methods assume full access to model internals, whereas others treat the model as an immutable black box; some prioritize local explanations for individual predictions, while others provide global insight into a model's behavior.....
Access restricted. Please log in or start a trial to view this content.
This review examined 75 studies of XAI in visual inspection, resulting in a taxonomy of post-hoc and intrinsic methods and a benchmark-driven comparative analysis. The findings confirm that no single method universally outperforms others: gradient-based approaches such as GradCAM and LRP are efficient for real-time defect localization, but their explanations can be coarse in complex regions. Perturbation-based methods such as SHAP, LIME, and RISE provide finer detail and strong fidelity but often fail in practice due to .......
Access restricted. Please log in or start a trial to view this content.
The authors declare that there are no conflicts of interest regarding the publication of this work.
The author(s) declare that this work received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.
....Access restricted. Please log in or start a trial to view this content.
Access restricted. Please log in or start a trial to view this content.