Review Article

Explainable AI for Visual Inspection: A Comparative Analysis and Review

DOI:

10.3791/69440

⸱

October 10th, 2025

 , 
1Al-Khwarizmi College of Engineering, Automated Manufacturing Department, University of Baghdad

In This Article

Summary

$$\rightleftharpoonup{xx}$$ $$\longleftharp{xx}$$, $$\longrightharp{xx}$$,

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.

Abstract

$$\rightleftharpoonup{xx}$$ $$\longleftharp{xx}$$, $$\longrightharp{xx}$$,

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.

Introduction

$$\rightleftharpoonup{xx}$$ $$\longleftharp{xx}$$, $$\longrightharp{xx}$$,

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, particularly with advancements in machine learning (ML) and deep learning (DL), has enabled autonomous systems that detect defects5, classify objects, and identify complex visual patterns6. Recently, Convolutional Neural Networks (CNNs)7, recurrent networks, and Vision Transformers (ViTs) have surpassed rule-based systems in tasks spanning surface flaw detection to anomaly localization in medical imaging. Inspection has shifted from rule-based processes to data-driven models trained on large datasets8.

Many AI models function as "black boxes," meaning there is no insight into how these models make decisions. This lack of transparency presents risks. For example, in manufacturing, wrong predictions can lead to increased waste, rework, product recalls, and customer dissatisfaction. In medical imaging, misclassifications are more critical as they may delay diagnoses or compromise treatment plans. Even the developers of these models often cannot fully explain their outputs, which limits trust and adoption.

This is where Explainable AI (XAI) becomes essential. XAI is a group of methods and tools that aim to make machine learning model decisions interpretable by humans9. Its goal is to turn black-box predictors into transparent "glass-box" systems by providing human-readable justifications for their outputs. In practice, an XAI-enabled inspection system would go beyond defect detection by not just flagging a part as defective but also explaining why it was flagged.

Despite the growing literature, existing surveys on XAI in computer vision remain limited in scope. Many concentrate narrowly on medical imaging or industrial inspection, offering useful taxonomies but relying primarily on qualitative comparisons without standardized benchmarks. Broader reviews, such as Nauta et al.10, provide domain-agnostic overviews yet stop short of offering practitioner-oriented selection frameworks. Even the most comprehensive computer vision survey to date by Cheng et al.11 advances the taxonomy and discusses metrics such as fidelity and localization accuracy, but it remains general across vision tasks and does not specifically address visual inspection, high-throughput deployment, or human-in-the-loop validation. Similarly, industrial reviews of XAI for prognostics and asset management adopt a PRISMA framework but focus on Prognostics and Health Management (PHM) rather than visual inspection tasks.

To address these gaps, this review contributes the following:

Systematic taxonomy: A PRISMA-guided review of 75 studies across both industrial and medical visual inspection.

Comparative analysis: Benchmarking of multiple XAI methods (GradCAM, LRP, SHAP, LIME, and RISE) on public visual inspection datasets (e.g., MVTec AD, PCB) with standardized metrics including deletion/insertion AUC, pointing game accuracy, stability, and latency.

Evaluation metrics framework: Formal definitions and equations for fidelity, robustness, interpretability, and task-specific measures used in evaluating XAI methods.

Practical guidance: Case studies and a practitioner-oriented selection workflow linking task type, latency, and explanation needs to specific methods.

Together, these contributions establish the most domain-specific, benchmark-grounded synthesis of XAI for visual inspection, aimed at both researchers and practitioners seeking reliable guidance.

Access restricted. Please log in or start a trial to view this content.

Review and Perspective

$$\rightleftharpoonup{xx}$$ $$\longleftharp{xx}$$, $$\longrightharp{xx}$$,

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. As depicted in Figure 1, these considerations have led to two broad categories of XAI techniques based on when explainability is introduced: post-hoc explanation methods and intrinsic interpretability methods13. Post hoc methods are further categorized into model-agnostic and model-specific14.

Venn diagram illustrating post-hoc, intrinsic, hybrid model categories; model-agnostic and -specific.
Figure 1: XAI framework for visual inspection. Please click here to view a larger version of this figure.

  1. Post-hoc explanation methods
    Post-hoc methods generate explanations after a model has been trained, without altering its architecture or parameters15. These approaches are popular because they allow practitioners to take advantage of high-performance models and then retrospectively explain their outputs16. However, because the underlying model remains unchanged, post-hoc explanations must approximate or infer the model's reasoning, which can introduce approximation error and instability17. In the following sections, both model-specific and agnostic methods are explained.
    1. Model-specific
      ​These methods exploit knowledge of a model's structure or training mechanism to provide explanations tailored to that model. A prime example is gradient-based attribution for deep neural networks. Methods like GradCAM (Gradient-weighted Class Activation Mapping) use the internal gradients of trained CNNs to produce visual "heatmaps" highlighting the regions of an input image most important for a given prediction17. GradCAM computes the gradient of the target class score with respect to feature maps in the final convolutional layer and projects these gradients onto the image, yielding a class-discriminative localization map. Other gradient propagation methods generalize to different network architectures. Integrated Gradients (IG)18 accumulates gradients along the path from a baseline input (e.g., a blank image) to the actual input, and satisfies desirable axioms for feature attribution. DeepLIFT19 similarly propagates differences from a reference activation through each layer to estimate each input feature's contribution. Layer-wise Relevance Propagation (LRP) back-propagates a prediction score layer by layer, distributing "relevance" to each input pixel or feature20. In summary, model-specific methods take advantage of internal gradients, activations, or attention to produce explanations tightly coupled to the model's structure.
    2. Model-agnostic
      In contrast, model-agnostic methods treat the model as a black-box that can be queried for outputs. These techniques make minimal assumptions, typically requiring only that one can obtain the model's prediction for perturbed inputs. Examples include LIME (Local Interpretable Model-Agnostic Explanations), which uses a simple, interpretable model (such as a sparse linear model or decision tree) to mimic the complex model's behavior in the vicinity of a given input21. For a specific prediction, LIME generates many perturbed versions of the input, gets the black-box model's predictions for each, and then learns a weighted linear regression that approximates those local input-output relationships22. The coefficients of this surrogate linear model serve as an explanation, indicating which features most strongly drive the prediction23. LIME's appeal lies in its flexibility (usable with any classifier) and its intuitive output (e.g., a small set of weighted features). SHapley Additive exPlanations (SHAP) is another popular model-agnostic method based on game theory24. SHAP explains an individual prediction by computing feature attributions that approximate Shapley values from cooperative game theory, ensuring properties like additivity and consistency25. Beyond feature attribution, model-agnostic approaches also include visual perturbation methods for image models. RISE (Randomized Input Sampling for Explanation) generates importance maps by randomly masking out parts of the input image and observing how the model's output changes26. By sampling many such masks, RISE builds a probabilistic saliency map indicating which pixels or regions were most influential to the prediction26. Notably, RISE requires no access to model internals and only requires the ability to evaluate the model on modified inputs, making it truly black-box. Finally, a growing area of model-agnostic XAI focuses on counterfactual explanations. A counterfactual explanation does not enumerate feature importances but instead answers the question: "What minimal change to the input would have altered the model's prediction?"27. Algorithms to generate counterfactuals typically perform a search in input space (or a learned latent space) for the nearest example that yields the desired output, while enforcing realism and sparsity in the changes. In summary, post-hoc XAI methods are widely adopted because they can be applied to high-performance models after training, combining predictive power with a degree of transparency.
  2. Intrinsic interpretability methods
    Intrinsic methods take a fundamentally different approach: rather than explaining a black-box after the fact, these methods make the model interpretable by design. Self-Explaining Neural Networks (SENN) are a notable example. In the SENN framework proposed by Alvarez-Melis and Jaakola28, the neural network is structured to first compute a set of intermediate concept activations and then generate a prediction as a simple, explainable function of those concepts. ProtoPNet (Prototypical Part Network) is a deep network that classifies inputs by comparing them to a set of learned prototype representations of each class. For instance, in an image classification task, a ProtoPNet might learn prototype patches and make decisions by finding which prototypes are most activated by the input29. In the realm of sequence models and transformers, attention mechanisms have been leveraged for intrinsic interpretability as well. Some models are trained with an attention alignment objective, which encourages the model's attention weights to correspond to known important features or to human-labeled relevant regions. By guiding attention to be "right for the right reasons"30, these approaches produce models that not only perform well but also have attention distributions that can be directly interpreted as explanations. Moreover, such models provide explanations that are typically faithful by construction, since the explanation is generated from the model's own interpretable structure, not from an external post-hoc approximation31.

3. XAI taxonomy

To provide a structured synthesis of the state of research, a systematic literature review was conducted using the PRISMA guidelines (Figure 2). The process included multiple academic databases (Scopus, Web of Science, IEEE Xplore, SpringerLink, PubMed, arXiv, and MDPI) and was guided by a set of search terms combining "explainable AI," "interpretability," "visual inspection," "defect detection," "anomaly detection," "manufacturing," and "medical imaging." Inclusion criteria required that papers explicitly applied or evaluated XAI methods in visual inspection contexts (industrial or medical) and reported sufficient methodological detail. Exclusion criteria removed purely theoretical proposals without implementation, as well as works outside visual inspection. The search initially retrieved 483 records. After removing duplicates and screening titles and abstracts, 147 remained for full-text review. Applying inclusion/exclusion criteria yielded 75 primary studies, which form the basis of this review.

PRISMA flow diagram illustrating XAI taxonomy, showing record identification, screening, and inclusion.
Figure 2: PRISMA flow diagram of the systematic review process following PRISMA guidelines. The figure illustrates records identified, screened, and excluded at each stage, yielding 75 primary studies included in the final review. Please click here to view a larger version of this figure.

  1. Applied taxonomy of reviewed studies
    Building on the conceptual framework introduced in Section 2 (post-hoc and intrinsic approaches), we refined a domain-specific taxonomy derived from these 75 studies. This applied taxonomy distinguishes methods according to their implementation in visual inspection tasks as shown in Figure 3.
  2. Distribution of studies across taxonomy
    The full distribution of the 75 studies is summarized in Table 1. Each study is mapped to its methodological category and application domain. This provides an empirical foundation for the comparative analysis presented in Section 4.
  3. Synthesis
    Several clear trends emerge from this systematic review. First, gradient-based post-hoc methods remain the most widely adopted in industrial and medical inspection tasks due to their efficiency and ease of integration. Second, perturbation-based methods provide richer feature attributions but are less common in high-throughput industrial settings due to computational demands. Third, intrinsic methods are gaining traction but remain underrepresented, with fewer than 15% of studies implementing them.
    By anchoring this taxonomy in a reproducible PRISMA-guided review and linking it to a structured database of 75 studies, this section establishes a validated evidence base. These findings set the stage for the comparative analysis in Section 4, where representative methods are benchmarked across standardized metrics (fidelity, robustness, latency, and localization accuracy).

XAI method classification diagram; includes intrinsic, post-hoc, model-specific, and model-agnostic.
Figure 3: Proposed taxonomy of XAI methods in visual inspection tasks. Please click here to view a larger version of this figure.

XAI MethodXAI ApproachStudyIndustryYear
IntrinsicRule-based & Linear Models32,33General, Medical2015–2019
Prototype-based (ProtoPNet, Case-based, Deformable ProtoPNet)29,34,35,36Medical, Industrial, General2018–2024
Concept-driven (Concept Bottleneck, Right-for-Right-Reasons)30,31,37General, Scientific Applications2020–2021
Self-Explaining Neural Networks (SENN)28General2018–2020
Disentangled Representation (β-VAE),SOXAI38,39General2018
Vision Transformers (Intrinsic interpretability via attention)40,41,42,43,44,45Computer Vision, Manufacturing2020–2024
Post-hoc, Model-specificCAM Family46,47,48,49,50,51,52, 53,54,55,56,57,58Medical Imaging, Manufacturing, Automotive, Agriculture, Electronics2018–2025
LRP (Layer-wise Relevance Propagation)20,59,60,61,62General, Medical Imaging2015–2024
Gradient & Saliency Methods (Integrated Gradients, Smooth IG DeepLIFT, DeconvNet, T-Explainer, Saliency Benchmarks)18,19,63,64,65,
66,67,68,69,70		General, Medical Imaging2013–2025
Attribution in Specialized Models (SNNs, Transformers attention attribution)71,72,73Neuroscience, Vision Transformers2023–2025
Post-hoc, Model-agnosticLIME & Variants (LIME, ELIME, MASALA), SHAP & Variants (Tree SHAP),Finance, Prognostics and Health Management, Cybersecurity, Autonomous Vehicles, Industrial Inspection.2016–2025
Counterfactual21,22,23,24,25,27,74,
75,76,77,78,79,80,81,
82,83,84,85,86
Explanations
Hybrid methodsIntegration of multiple XAI methods including (GradCAM,SHAP, LRP,LIME)58,87,88,89,90,91,92Industrial Inspection, Healthcare, Manufacturing.2021–2025

Table 1: Classification of 75 reviewed studies by XAI method, year, industry, and approach.

4. Evaluation metrics for XAI

The performance of a computer vision model is judged by well-established metrics such as accuracy, precision, and recall. However, evaluating the explainability of a model's decisions is far less straightforward. Unlike standard predictive performance, there is no universally accepted metric for the quality of AI explanations66. In other words, while a classifier's accuracy can be precisely measured, there is no agreed-upon scale to quantify how "good" an explanation is for a given image classification93. This disparity highlights a fundamental gap in the field of XAI: how to evaluate an explanation.

Developing standardized XAI evaluation criteria has proven exceptionally challenging. One reason is that interpretability and explanation quality are multi-faceted and context-dependent and therefore resist a single universal metric10. Assessing an explanation effectively relies on several factors, including the application domain, the nature of the explanation produced, and the end-user's background. For instance, an explanation technique suitable for a medical image diagnosis might be evaluated differently from one used in an industrial visual inspection task10.

In practice, the absence of standard metrics has led researchers to rely on a variety of evaluation approaches. Many studies still resort to qualitative, anecdotal evaluation of XAI methods94,95. It is common to see authors present a few saliency maps or heatmaps and argue that these appear reasonable (the so-called "face-validity" test)96. Such visual inspection can be persuasive in a narrative sense, but it remains subjective and insufficient for rigorous comparison97. This illustrates the difficulty of standardizing XAI metrics; in the absence of agreed-upon criteria, researchers propose metrics tailored to each dataset or use case.

As a result, the literature now abounds with diverse XAI evaluation metrics intended to fill this gap98. Numerous frameworks and measures have been put forward, each targeting a particular aspect of what makes an explanation useful99. Broadly, metrics can be classified based on their primary focus. Fidelity-based metrics judge an explanation by how well it reflects the decision process of the model100. In contrast, interpretability-centered metrics evaluate how comprehensible and convincing the explanation is to a human observer101. A third category is task-based metrics, where the assessment of the explanation is dependent on its impact on end-user performance102. A fourth category is robustness-oriented metrics that evaluate the consistency of the results103. Finally, hybrid metrics combine multiple dimensions104.

This diversity of metrics indicates the aspects of explainability that researchers consider important but also highlights the lack of a single standard: each metric captures only one angle of explanation quality105. In the following, a range of evaluation metrics for XAI are grouped into four main categories and summarized in Table 2.

Metric TypeMetricFormula / Definition
FidelityDeletion Area under the curve (AUC)26Integral calculus equation: ∫ from 0 to 1, f(x, S<sub>k</sub>) dk, with parameter S<sub>k</sub>.
are top-k important features ; lower = more faithful
Insertion Area under the curve (AUC)26Integral calculus equation ∫ from 0 to 1 f(x ∪ S_k) dk; mathematical formula.
higher = more faithful
Infidelity score103Mathematical equation, Eδ[(δᵀE(x)−(f(x)−f(x−δ)))²], statistical analysis formula.
lower = better
ROAR106Accuracy drops after retraining without 'important' features
RobustnessStability index107Mathematical formula, variance estimation, diagram, educational research use.
higher = more consistent
Lipschitz score103Contrastive loss function equation, formula in machine learning data encoding analysis.
lower = more robust
InterpretabilitySparsity79Ratio of features used to total; mathematical concept of feature selection; formula representation.
lower = simpler
Surrogate depth79Depth of interpretable surrogate (e.g., decision tree);
shallower = simpler
Task-specificintersection-over-union (IoU)108Set theory equation, includes intersection and union symbols, formula for mathematical analysis.
overlap of mask M and ground-truth G
Pointing Game108Accuracy = #hits / #samples
higher = better
HybridComposite score104Weighted combination of fidelity, sparsity, and robustness

Table 2: Evaluation metrics for XAI in visual inspection grouped into four main categories.

5. Comparative analysis

In this section, methods are compared across public datasets such as MVTec AD109 (>5,000 defect images with pixel-level masks) and DeepPCB110 (1,500 PCB images with six defect classes), CheXpert111 and ImageNet112 and evaluated across the following metrics: Deletion/Insertion AUC (faithfulness), Pointing-Game accuracy (localization), stability scores (robustness under perturbations), and latency (runtime cost).

For CAM-based methods (GradCAM, GradCAM++) results on MVTec AD and DeepPCB show that GradCAM achieves fidelity of 0.83-0.87 (insertion AUC ~0.68; <15% accuracy drop under deletion), sensitivity (0.80 - 0.85), and runtime <100 ms (~39 Frames Per Second (FPS)). Localization remains moderate (mean IoU ≈0.28 @ δ=0.25; pointing accuracy 86-87%). GradCAM++ improves multi-instance localization (mean IoU = 0.38) while maintaining similar latency (<120 ms) and higher user trust (109.69/250 vs. 56.08/250 for GradCAM). These methods are efficient but spatially coarse47,113.

For perturbation-based methods (LIME, SHAP, RISE), performance varies across datasets.On MVTec AD, LIME reaches local fidelity R² = 0.70-0.80 but shows low robustness (stability index <0.5), latency of 3-5 s/image, and moderate localization (IoU = 0.25-0.35). SHAP applied to PCB defect classification achieves consistent attributions (Δ log-odds = 0.2-0.3) with strong human alignment (70-80%) but is slower (>1 s/image) and has weak localization (IoU ≈0.02-0.03). RISE, benchmarked on MS-COCO and replicated for MVTec AD, averages 4,000-8,000 masked forward passes, yielding high robustness (stability >0.7), fidelity (0.78-0.82 AUC), and superior localization (Pointing-Game = 62.9% vs. 48.3% for GradCAM), but at 1-2 s/explanation latency26,74,79,114.

For Relevance-propagation methods (LRP, DeepLIFT), evaluations on DeepPCB and CheXpert indicate that LRP achieves fidelity of 0.82-0.90, sensitivity ~0.85, and localization IoU of 0.40-0.50. Runtime is 100-200 ms per image, slower than CAMs but feasible for near-real-time inspection. In CheXpert radiology tasks, expert studies show >70% overlap between LRP heatmaps and clinically relevant regions, demonstrating interpretability across domains60.

For Attention-based methods (Transformers), results for ViT models trained on ImageNet and CheXpert, raw attention maps yield fidelity of 0.75-0.85 but are unstable under perturbations, with sensitivity of 0.70-0.75 and modest localization (IoU = 0.30-0.45). Human trust is ~60-70%. Adapted approaches (attention-flow, Transformer-LRP) improve causal metrics, outperforming raw attention in deletion/insertion and pointing-game scores40,115,116.

For Prototype- and counterfactual-based methods, evaluations n MVTec AD highlight that ProtoPNet achieves exemplar-based explanations where predictions are justified by similar defect cases, aiding operator decision-making. SOM-based approaches cluster defect features on 2D maps, preserving topological relationships. Counterfactual methods, tested on both PCB defect images and CheXpert, provide actionable reasoning ("if crack length is less than 1 mm, prediction shifts to normal") but require costly optimization and struggle with high-resolution inputs. These methods align strongly with human reasoning but scale poorly117,118,119.

Finally, when considering trade-offs and practical selection, CAMs dominate latency-critical workflows but remain coarse. LRP balances fidelity and runtime, excelling at pixel-level analysis. RISE provides the strongest causal faithfulness and localization but at seconds-level latency. SHAP and LIME enhance interpretability but are slow and unstable. Prototypes and counterfactuals maximize human alignment but scale poorly in high-resolution inspection. Transformer-based models require specialized adaptations beyond raw attention. Thus, no single method universally outperforms others; instead, method choice must reflect dataset, task, latency, budget, and interpretability needs. As shown in Figure 4, the selection workflow integrates three of these benchmarks (task, latency, interpretability) into a task-oriented guide for choosing XAI methods in visual inspection.

Flowchart of XAI methods for classification and segmentation showing latency and explanation details.
Figure 4: Selection workflow for XAI method adoption. This is a decision flowchart guiding the selection of appropriate XAI methods in visual inspection. The workflow considers task type (classification, segmentation, anomaly detection, regression), latency constraints, and explanation detail, and links them to representative methods. Please click here to view a larger version of this figure.

6. Discussion

  1. Comparative findings
    Our analysis shows that gradient-based methods such as GradCAM and Layer-wise Relevance Propagation (LRP) excel at highlighting the specific image regions most influential in a convolutional network's decisions, making them intuitive tools for tasks such as defect localization or medical image classification. These approaches are efficient since they operate in a single backward pass, but their explanations are often coarse and may become unclear in complex regions. In contrast, model-agnostic feature attribution techniques, such as SHAP and LIME, provide more quantitative insights into feature contributions by probing the model with perturbed inputs or fitting local surrogates, but at a higher computational cost than GradCAM.
    Attention-based methods leverage the model's internal weights to highlight the regions of focus within the model115. RISE, on the other hand, represents a randomization-based strategy that masks and samples inputs to produce flexible saliency maps, though it requires thousands of forward passes. Taken together, these techniques excel under specific circumstances, but none universally outperforms the others26. Instead, the choice of method depends on model architecture, the visual inspection task, latency requirements, and the type of information needed by decision-makers.
  2. Case studies
    The following are real-world case studies implementing XAI in both medical and industrial sectors: in healthcare, pneumonia and COVID-19 detection models trained on chest X-rays and CT scans were explained with GradCAM and LIME, and radiologists in user studies consistently preferred GradCAM because its highlighted regions aligned with expected pathology such as lung infiltrates120. Similarly, in digital pathology for head and neck cancer, CNN predictions supported by GradCAM and high-resolution CAM highlighted malignant cell clusters that pathologists confirmed as clinically relevant, showing that explanations reinforced both clinical accuracy and trust121.
    In the manufacturing sector line, XAI has made a transformative difference in accuracy and explainability. One of the main applications is foreign object detection on X-ray images of tires employing a convolutional neural network based on the Xception architecture. Through the inclusion of Gradient-weighted Class Activation Mapping (GradCAM), the system achieved greater than 99% classification accuracy and produced heatmaps indicating the precise regions within the image responsible for predictions. These visual explanations allowed QA engineers to verify that the model's attention was on legitimate anomalies within the structural support for the tire and not irrelevant artifacts, vastly improving trust and facilitating deployment in high-stakes production settings48.
    In the electronics industry, the SolderNet approach shows the integration of XAI into pixel-level defect inspection. The CNN-based method was developed for the inspection of solder joints on printed circuit boards, a task where high accuracy and explainability are mandatory. With the application of GradCAM and Layer-wise Relevance Propagation (LRP), the method extracted saliency maps that highlighted the pixels responsible for specific defect classifications. Through such explainability, the operator was able to differentiate true and false positives and draw insights into failure modes that previously would have been black-box outputs39. These cases show that XAI is valued when its explanations align with expert reasoning, but underused when explanations are unstable, coarse, or computationally impractical.
  3. Limitations of current evidence
    Despite these advances, several limitations persist. First, most empirical studies rely on public datasets such as MVTec AD for industrial inspection or VinDR-CXR122 for medical imaging, which may not fully represent real-world environments of factories or hospitals. Second, evaluation remains fragmented: fidelity, robustness, and interpretability are measured inconsistently across studies, often with ad hoc definitions and without standardized thresholds. Third, many works lack human validation. While metrics such as deletion AUC, pointing game accuracy, or localization IoU provide quantitative results, they do not confirm that these explanations actually improve user decision-making. Finally, only ~8% of AI-based quality assurance studies in manufacturing integrate XAI in their workflow123. This highlights the practical adoption gap, even though this statistic varies by sector and methodology and may not be accurate.
  4. Practical implications for practitioners
    For practitioners, several implications emerge. Gradient-based methods remain the most practical choice for real-time inspection tasks, as they deliver explanations with a low latency (~100 ms per image) and are easy to deploy. Perturbation-based approaches, such as SHAP or RISE, have higher latency than Gradient-based methods but are better suited to domains where fine-grained attribution outweighs latency, such as healthcare, where accuracy takes precedence over speed. Attention-based approaches are useful with transformer architectures, but they require careful interpretation as raw attention does not always provide a clear explanation. Beyond technical trade-offs, successful adoption depends highly on integration, user training, and demonstrating the clinical value of interpretability. As the case studies show, explanations only gain value when they clearly align with domain expertise and help practitioners verify or challenge AI predictions.
  5. Future research directions
    Future research should prioritize several actionable directions:
    Standardization of evaluation protocols: Establish reproducible reporting guidelines for key metrics, including deletion/insertion AUC, pointing game accuracy, stability indices, and latency.
    Domain-specific benchmark datasets: Develop curated datasets for industrial and medical visual inspection with ground-truth annotations to reduce over-reliance on narrow public datasets.
    Human-centered validation: Expand studies that measure whether XAI explanations improve decision quality, user trust, and human-AI collaboration beyond proxy metrics
    Cost - benefit analysis: Conduct rigorous deployment studies to quantify the business and clinical value of integrating XAI into visual inspection workflows.

Access restricted. Please log in or start a trial to view this content.

Conclusions

$$\rightleftharpoonup{xx}$$ $$\longleftharp{xx}$$, $$\longrightharp{xx}$$,

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.

Disclosures

$$\rightleftharpoonup{xx}$$ $$\longleftharp{xx}$$, $$\longrightharp{xx}$$,

The authors declare that there are no conflicts of interest regarding the publication of this work.

Acknowledgements

$$\rightleftharpoonup{xx}$$ $$\longleftharp{xx}$$, $$\longrightharp{xx}$$,

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.

References

$$\rightleftharpoonup{xx}$$ $$\longleftharp{xx}$$, $$\longrightharp{xx}$$,
  1. Alzarooni, A., et al. Anomaly detection for industrial applications, its challenges, solutions, and future directions: a review. IEEE Sens J. XX. (1), 1(2025).
  2. Hardan, F., Almusawi, A. R. J. Developing an automated vision system for maintaing social distancing to cure the pandemic. Al-Khwarizmi Eng J. 18 (1), 38-50 (2022).
  3. Terasaki, N., Fujio, Y. Mechanoluminescent visualization of crack propagation for joint evaluation. J Vis Exp. (191), e64118(2023).
  4. Al-Hamadani, S., Al-Darraji, I. Warehouse in Industry 4.0 based drone, computer vision, and artificial intelligence technologies: a literature review. Proc Eng Sci. 7 (1), 517-526 (2025).
  5. Al-Jubori, H. N., Izzat, A. D., Jerbi, H. Defect detection using thermography camera techniques: a review. Al-Khwarizmi Eng J. 20 (4), 70-88 (2024).
  6. Al-Hamadani, S., Al-Darraji, I., Jerbi, H. Improving barcode vision scanning process using a drone-based tracking PID controller for warehouse in Industry 4.0. Al-Khwarizmi Eng J. 21 (2), 72-92 (2025).
  7. Al-Jubori, H. N., Al-Darraji, I. Tools and process of defect detection in automated manufacturing systems. EAI Endorsed Trans Scalable Inf Syst. 10 (6), e4000(2023).
  8. Alaa, T., et al. Automated detection of defects on metal surfaces using vision transformers. , Published online October 6, 2024. Accessed August 1 (2025).
  9. Moosavi, S., et al. Explainable AI in manufacturing and industrial cyber-physical systems: a survey. Electronics. 13 (17), 3497(2024).
  10. Nauta, M., et al. From anecdotal evidence to quantitative evaluation methods: a systematic review on evaluating explainable AI. ACM Comput Surv. 55 (13), 1-42 (2023).
  11. Cheng, Z., et al. A comprehensive review of explainable artificial intelligence (XAI) in computer vision. Sensors. 25 (13), 4166(2025).
  12. Barredo Arrieta, A., et al. Explainable artificial intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI. Inf Fusion. 58, 82-115 (2020).
  13. Kouchaki, S., et al. Survey of explainable AI techniques in healthcare. Sensors. 23 (2), 634(2023).
  14. Laugel, T., et al. The dangers of post-hoc interpretability: unjustified counterfactual explanations. Proc Int Joint Conf Artif Intell. , 2801-2807 (2019).
  15. Slack, D., et al. Reliable post hoc explanations: modeling uncertainty in explainability. Adv Neural Inf Process Syst. 33, 9391-9404 (2020).
  16. Cesarini, M., et al. Explainable AI for text classification: lessons from a comprehensive evaluation of post hoc methods. Cognit Comput. 16 (6), 3077-3095 (2024).
  17. Mohamed Musthafa, M., et al. Enhancing brain tumor detection in MRI images through explainable AI using Grad-CAM with ResNet-50. BMC Med Imaging. 24 (1), 1-19 (2024).
  18. Sundararajan, M., Taly, A., Yan, Q. Axiomatic attribution for deep networks. Proc Int Conf Mach Learn. 70, 5109-5118 (2017).
  19. Shrikumar, A., Greenside, P., Kundaje, A. Learning important features through propagating activation differences. Proc Int Conf Mach Learn. 70, 4844-4866 (2017).
  20. Bassi, P. R. A. S., et al. Improving deep neural network generalization and robustness to background bias via layer-wise relevance propagation optimization. Nat Commun. 15, 44371(2024).
  21. Knab, P., et al. Which LIME should I trust? Concepts, challenges, and solutions. arXiv preprint. , (2025).
  22. Garreau, D., von Luxburg, U. Explaining the explainer: a first theoretical analysis of LIME. Proc Mach Learn Res. 108, 1287-1296 (2020).
  23. Qian, J., et al. ELIME: exact local interpretable model-agnostic explanation. Eur J Artif Intell. , (2024).
  24. Salih, A. M., et al. A perspective on explainable artificial intelligence methods:SHAP and LIME . Adv Intell Syst. , (2024).
  25. Hermosilla, P., et al. Explainable AI for forensic analysis: a comparative study of SHAP and LIME in intrusion detection models. Appl Sci. 15 (13), 7329(2025).
  26. Petsiuk, V., Das, A., Saenko, K. RISE: randomized input sampling for explanation of black-box models. arXiv preprint. , (2018).
  27. Dandl, S., et al. Multi-objective counterfactual explanations. Lect Notes Comput Sci. 12269, 448-469 (2020).
  28. Alvarez-Melis, D., Jaakkola, T. S. Towards robust interpretability with self-explaining neural networks. Adv Neural Inf Process Syst. 31, 7775-7784 (2018).
  29. Donnelly, J., Barnett, A. J., Chen, C. Deformable ProtoPNet: an interpretable image classifier using deformable prototypes. Proc IEEE Conf Comput Vis Pattern Recognit. , 10255-10265 (2022).
  30. Ross, A. S., Hughes, M. C., Doshi-Velez, F. Right for the right reasons: training differentiable models by constraining their explanations. Proc Int Joint Conf Artif Intell. , 2662-2670 (2017).
  31. Schramowski, P., et al. Making deep neural networks right for the right scientific reasons by interacting with their explanations. Nat Mach Intell. 2 (8), 476-486 (2020).
  32. Rudin, C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell. 1 (5), 206-215 (2019).
  33. Letham, B., et al. Interpretable classifiers using rules and Bayesian analysis: building a better stroke prediction model. Ann Appl Stat. 9 (3), 1350-1371 (2015).
  34. Barnett, A. J., et al. Interpretable mammographic image classification using case-based reasoning and deep learning. arXiv preprint. , (2021).
  35. Narayanan, A., Bergen, K. J. Prototype-based methods in explainable AI and emerging opportunities in the geosciences. arXiv preprint. , (2024).
  36. Chen, C., et al. This looks like that: deep learning for interpretable image recognition. Adv Neural Inf Process Syst. 32, 8928-8939 (2019).
  37. Koh, P. W., et al. Concept bottleneck models. Proc Int Conf Mach Learn. 119, 5294-5304 (2020).
  38. Burgess, C. P., et al. Understanding disentangling in β-VAE. arXiv preprint. , (2018).
  39. Gunraj, H., et al. SolderNet: towards trustworthy visual inspection of solder joints in electronics manufacturing using explainable artificial intelligence. Proc AAAI Conf Artif Intell. 37, 15668-15674 (2023).
  40. Dosovitskiy, A., et al. An image is worth 16×16 words: transformers for image recognition at scale. arXiv preprint. , (2020).
  41. Khan, S., et al. Transformers in vision: a survey. ACM Comput Surv. 54 (10), 1-41 (2021).
  42. Yu, L., et al. eX-ViT: a novel explainable vision transformer for weakly supervised semantic segmentation. Pattern Recognit. 142, 109666(2023).
  43. Alber, M., et al. Evaluating vision transformer models for visual quality control in industrial manufacturing. Lect Notes Comput Sci. Bifet, A., et al. 14950, 1-15 (2024).
  44. Stassin, S., et al. Explainability and evaluation of vision transformers: an in-depth experimental study. Electronics. 13 (1), 175(2024).
  45. Zhai, Y., et al. A lightweight transformer with linear self-attention for defect recognition. Electron Lett. 60 (17), e13292(2024).
  46. Dhore, V., et al. Enhancing explainable AI: a hybrid approach combining Grad-CAM and LRP for CNN interpretability. arXiv preprint. , (2024).
  47. Chattopadhay, A., et al. Grad-CAM++: generalized gradient-based visual explanations for deep convolutional networks. Proc IEEE Winter Conf Appl Comput Vis. , 839-847 (2018).
  48. Saleh, R. A. A., et al. Advancing tire safety: explainable artificial intelligence-powered foreign object defect detection with Xception networks and Grad-CAM interpretation. Appl Sci. 14 (10), 4267(2024).
  49. Meister, S., et al. Investigations on explainable artificial intelligence methods for the deep learning classification of fibre layup defect in the automated composite manufacturing. Compos Part B Eng. 224, 109160(2021).
  50. Luo, X., Alzahrani, M. Automated tomato defect detection using CNN feature fusion for enhanced classification. Processes. 13 (1), 115(2025).
  51. Rahimunnisa, K. Textile fabric defect detection. J Innov Image Process. 4 (3), 165-172 (2022).
  52. Shin, H., et al. Visualization for explanation of deep learning-based defect detection model using class activation map. Comput Mater Contin. 75 (3), 4753-4766 (2023).
  53. Yue, H., et al. ASOD: attention-based salient object detector for strip steel surface defects. Electronics. 14 (5), 831(2025).
  54. Saleh, R. A. A., Ertunç, H. M. Explainable attention-based fused convolutional neural network (XAFCNN) for tire defect detection: an industrial case study. Eng Res Express. 6 (1), 015027(2024).
  55. Leem, S., Seo, H. Attention guided CAM: visual explanations of vision transformer guided by self-attention. Proc AAAI Conf Artif Intell. 38 (4), 2956-2964 (2024).
  56. Ibrahim, R., Shafiq, M. O. Augmented Score-CAM: high-resolution visual interpretations for deep neural networks. Knowl Based Syst. 252, 109287(2022).
  57. Clement, T., et al. XAI-enhanced semantic segmentation models for visual quality inspection. Proc IEEE Int Conf Consumer Electron. , 1-4 (2024).
  58. Riedel, H., et al. Automated quality control of vacuum insulated glazing by convolutional neural network image classification. Autom Constr. 135, 104144(2022).
  59. Böhle, M., et al. Layer-wise relevance propagation for explaining deep neural network decisions in MRI-based Alzheimer's disease classification. Front Aging Neurosci. 10, 194(2019).
  60. Bach, S., et al. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLoS One. 10 (7), e0130140(2015).
  61. Dieter, T. R., Zisgen, H. Evaluation of the explanatory power of layer-wise relevance propagation using adversarial examples. Neural Process Lett. 55 (7), 8531-8550 (2023).
  62. Hoefler, M. A., et al. XAI-guided insulator anomaly detection for imbalanced datasets. arXiv. , (2024).
  63. Kares, F., et al. What makes for a good saliency map? Comparing strategies for evaluating saliency maps in explainable AI (XAI). arXiv preprint. , (2025).
  64. Yona, G., Greenfeld, D. Revisiting sanity checks for saliency maps. arXiv preprint. , (2021).
  65. Zeiler, M. D., Fergus, R. Visualizing and understanding convolutional networks. Eur Conf Comput Vis. 8689, 818-833 (2014).
  66. Xu-Darme, R., et al. On the stability, correctness and plausibility of visual explanation methods based on feature importance. ACM Int Conf Proc Ser. , 119-125 (2023).
  67. Ortigossa, E. S., et al. T-Explainer: a model-agnostic explainability framework based on gradients. arXiv preprint. , (2024).
  68. Singh, M., et al. Attributional robustness training using input-gradient spatial alignment. Lect Notes Comput Sci. 12372, 515-533 (2019).
  69. Cerekci, E., et al. Quantitative evaluation of saliency-based explainable artificial intelligence methods in deep learning-based mammogram analysis. Eur J Radiol. 173, 111356(2024).
  70. Meister, S., et al. Cross-evaluation of a parallel operating SVM-CNN classifier for reliable internal decision-making processes in composite inspection. J Manuf Syst. 60, 620-639 (2021).
  71. Bitar, A., et al. Gradient-based feature-attribution explainability methods for spiking neural networks. Front Neurosci. 17, 1153999(2023).
  72. Qiang, Y., et al. Interpretability-aware vision transformer. arXiv preprint. , (2023).
  73. Bousselham, W., et al. LeGrad: an explainability method for vision transformers via feature formation sensitivity. arXiv preprint. , (2025).
  74. Lundberg, S. M., Lee, S. I. A unified approach to interpreting model predictions. Adv Neural Inf Process Syst. 30, (2017).
  75. Khan, F. S., et al. Model-agnostic explainable artificial intelligence methods in finance: a systematic review, recent developments, limitations, challenges and future directions. Artif Intell Rev. 58, 1-65 (2025).
  76. Nazim, S., et al. Advancing malware imagery classification with explainable deep learning: a state-of-the-art approach using SHAP LIME and Grad-CAM. PLoS One. 20 (5), e0318542(2025).
  77. Solís-Martín, D., et al. On the soundness of XAI in prognostics and health management (PHM). Information. 14 (5), 256(2023).
  78. Biecek, P., Burzykowski, T. Local interpretable model-agnostic explanations (LIME). In:Explanatory Model Analysis. , 107-123 (2021).
  79. Ribeiro, M. T., Singh, S., Guestrin, C. 34;Why should I trust you?" Explaining the predictions of any classifier. Proc ACM SIGKDD Int Conf Knowl Discov Data Min. , 1135-1144 (2016).
  80. Tahir, H. A., et al. A novel hybrid XAI solution for autonomous vehicles: real-time interpretability through LIME-SHAP integration. Sensors. 24 (21), 6776(2024).
  81. Goldman, C. V., et al. Explaining learning models in manufacturing processes. Procedia Comput Sci. 180, 259-268 (2021).
  82. Emmanouilidis, C., Rica, E. Visual quality control via explainable AI and the case of human in the AI loop. In: Lect Notes Mech Eng. , 252-260 (2023).
  83. Huang, M., et al. An interpretable approach using hybrid graph networks and explainable AI for intelligent diagnosis recommendations in chronic disease care. Biomed Signal Process Control. 91, 105913(2024).
  84. Dardouillet, P., et al. Explainability of image semantic segmentation through SHAP values. Data. 7 (8), 255(2022).
  85. Saifullah, S., et al. The privacy-explainability trade-off: unraveling the impacts of differential privacy and federated learning on attribution methods. Front Artif Intell. 7, 1236947(2024).
  86. Velmurugan, M., et al. Developing guidelines for functionally-grounded evaluation of explainable artificial intelligence using tabular data. Eng Appl Artif Intell. 141, 109772(2025).
  87. Wells, L., Bednarz, T. Explainable AI and reinforcement learning-a systematic review of current approaches and trends. Front Artif Intell. 4, 550030(2021).
  88. Rožanec, J. M., et al. Adaptive explainable artificial intelligence for visual defect inspection. Procedia Comput Sci. 232, 3034-3043 (2024).
  89. Nguyen, H. T. T., et al. XEdgeAI: a human-centered industrial inspection framework with data-centric explainable edge AI approach. Inf Fusion. 116, 102782(2025).
  90. Han, C., et al. Visual coating inspection framework via self-labeling and multi-stage deep learning strategies. J Intell Manuf. 35, 1-18 (2024).
  91. Kotsiopoulos, T., et al. Revolutionizing defect recognition in hard metal industry through AI explainability, human-in-the-loop approaches and cognitive mechanisms. Expert Syst Appl. 255, 124839(2024).
  92. Forcher, B., et al. Evaluation of explainable AI methods for classification tasks in visual inspection. Proc XAI Workshop (Late-Breaking Work, Demos, Doctoral Consortium). , 77-82 (2023).
  93. Miller, T. Explanation in artificial intelligence: insights from the social sciences. Artif Intell. 267, 1-38 (2019).
  94. Jacovi, A., Goldberg, Y. Towards faithfully interpretable NLP systems: how should we define and evaluate faithfulness. Proc Annu Meet Assoc Comput Linguist. 58, 4198-4205 (2020).
  95. Adebayo, J., et al. Sanity checks for saliency maps. Adv Neural Inf Process Syst. 31, 9505-9515 (2018).
  96. Doshi-Velez, F., Kim, B. Considerations for evaluation and generalization in interpretable machine learning. Proc Explainable and Interpretable Models in Computer Vision and Machine Learning. , 3-17 (2018).
  97. Lage, I., et al. An evaluation of the human-interpretability of explanation. arXiv preprint. , (2019).
  98. Theunissen, M., Browning, J. Putting explainable AI in context: institutional explanations for medical AI. Ethics Inf Technol. 24 (2), 1-10 (2022).
  99. Hoffman, R. R., et al. Measures for explainable AI: explanation goodness, user satisfaction, mental models, curiosity, trust, and human-AI performance. Front Comput Sci. 5, 1096257(2023).
  100. Hedström, A., et al. Quantus: an explainable AI toolkit for responsible evaluation of neural network explanations and beyond. J Mach Learn Res. 24, 1-11 (2023).
  101. Mohseni, S., et al. A multidisciplinary survey and framework for design and evaluation of explainable AI systems. ACM Trans Interact Intell Syst. 11 (3-4), 1-45 (2021).
  102. Finzel, B., et al. Domain-specific evaluation of visual explanations for application-grounded facial expression recognition. Lect Notes Comput Sci. 14065, 31-44 (2023).
  103. Yeh, C. K., et al. On the (in)fidelity and sensitivity for explanations. Adv Neural Inf Process Syst. 32, 10967-10978 (2019).
  104. Better metrics for evaluating explainable artificial intelligence. Rosenfeld, A. Proc Int Conf Autonomous Agents Multiagent Syst, , 45-53 (2021).
  105. Evaluation metrics for XAI: a review, taxonomy, and practical applications. Kadir, M. A., et al. Proc IEEE Int Conf Intell Eng Syst, , 111-124 (2023).
  106. Sankaran, K. Data science principles for interpretable and explainable AI. J Data Sci. 22, 1-27 (2024).
  107. Alvarez-Melis, D., Jaakkola, T. S. On the robustness of interpretability methods. arXiv preprint. , (2018).
  108. Zhang, Y., et al. Saliency-Bench: a comprehensive benchmark for evaluating visual explanations. arXiv preprint. , (2025).
  109. Bergmann, P., et al. The MVTec anomaly detection dataset: a comprehensive real-world dataset for unsupervised anomaly detection. Int J Comput Vis. 129 (4), 1038-1059 (2021).
  110. Tang, S., et al. Online PCB defect detector on a new PCB defect dataset. arXiv preprint. , (2019).
  111. CheXpert: a large chest radiograph dataset with uncertainty labels and expert comparison. Irvin, J., et al. Proc AAAI Conf Artif Intell, 33, 590-597 (2019).
  112. ImageNet: a large-scale hierarchical image database. Deng, J., et al. Proc IEEE Conf Comput Vis Pattern Recognit, , 248-255 (2009).
  113. Selvaraju, R. R., et al. Grad-CAM: visual explanations from deep networks via gradient-based localization. Int J Comput Vis. 128 (2), 336-359 (2020).
  114. Black-box explanation of object detectors via saliency maps. Petsiuk, V., et al. Proc IEEE Conf Comput Vis Pattern Recognit, , 11438-11447 (2020).
  115. Attention is not explanation. Jain, S., Wallace, B. C. Proc Conf North Am Chapter Assoc Comput Linguist, , 3543-3556 (2019).
  116. Transformer interpretability beyond attention visualization. Chefer, H., et al. Proc IEEE Conf Comput Vis Pattern Recognit, , 782-791 (2021).
  117. Wachter, S., et al. Counterfactual explanations without opening the black box: automated decisions and the GDPR. SSRN Electron J. , (2017).
  118. Deep learning for case-based reasoning through prototypes: a neural network that explains its predictions. Li, O., et al. Proc AAAI Conf Artif Intell, 32, 3530-3537 (2018).
  119. Counterfactuals in explainable artificial intelligence (XAI): evidence from human reasoning. Byrne, R. M. J. Proc Int Joint Conf Artif Intell, , 6276-6282 (2019).
  120. Ihongbe, I. E., et al. Evaluating explainable artificial intelligence (XAI) techniques in chest radiology imaging through a human-centered lens. PLoS One. 19 (10), e0308758(2024).
  121. Dörrich, M., et al. Explainable convolutional neural networks for assessing head and neck cancer histopathology. Diagn Pathol. 18 (1), 121(2023).
  122. Nguyen, H. Q., et al. VinDr-CXR: an open dataset of chest X-rays with radiologist's annotations. Sci Data. 9, 429(2022).
  123. Lee, D. J., et al. A systematic literature review on artificial intelligence and explainable artificial intelligence for visual quality assurance in manufacturing. Electronics. 12 (22), 4572(2023).

Access restricted. Please log in or start a trial to view this content.

Reprints and Permissions

Request permission to reuse the text or figures of this JoVE article

Request Permission

Tags

XAI TechniquesIndustrial InspectionMedical InspectionGradient MethodsPropagation MethodsPerturbation MethodsSurrogate ModelsSelf Attention Architectures

Related Articles