Thursday, November 19, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News

SHAP Outperforms LIME in Explaining Industrial Image Quality Control Models

November 19, 2026
in Science News
Blake Davidson
By Blake Davidson Scienmag Editorial Profile - Data Science
Reading Time: 6 mins read
0
SHAP Outperforms LIME in Explaining Industrial Image Quality Control Models

SHAP Outperforms LIME in Explaining Industrial Image Quality Control Models

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

In an industrial manufacturing trial conducted at the Smart Production Lab of the FH JOANNEUM University of Applied Sciences in Austria, researchers Anja Stadlhofer and Vitaliy Mezhuyev demonstrated that Shapley Additive Explanations provide a more stable, computationally efficient, and informative interpretation of convolutional neural network predictions than Local Interpretable Model-agnostic Explanations during automated quality inspection. Published in Industrial Artificial Intelligence, the study investigated how post-hoc interpretability frameworks expose the internal decision-making mechanics of deep neural networks deployed for defect detection. Manufacturing enterprises transitioning into smart factory operations under Industry 4.0 paradigms generate vast volumes of operational data and increasingly rely on artificial intelligence algorithms to automate processes previously executed by human operators. However, because complex neural architectures function as opaque black-box models that conceal the intermediate criteria underlying their classifications, industrial adoption remains constrained by a lack of operational confidence. To resolve this barrier to deployment, the investigators evaluated the two leading local interpretability algorithms on an empirical image classification pipeline designed to distinguish acceptable manufacturing components from defective parts.

The experimental setup focused on quality verification for watch stands manufactured within the smart production facility, formulating the physical inspection task as a supervised binary classification problem. A deep convolutional neural network was tasked with categorizing component images into two operational classes: items designated as good, which possess the requisite structural quality to advance into final watch assembly, and items designated as defect, which contain flaws that preclude further fabrication. High-resolution imagery of the physical watch stands was gathered using a 64-megapixel Sony smartphone camera, with data collection accelerated by recording continuous video sequences of components positioned across varying perspectives, physical orientations, and shifting lighting conditions. Individual frames were subsequently extracted from the video recordings to construct a standardized dataset comprising exactly 12,014 images, balanced equally between 6,007 examples of flawless watch stands and 6,007 examples displaying visible defects. The extracted image files were then downsized to optimize processing throughput while retaining the critical geometrical features required for algorithmic discrimination.

Data engineering and model development proceeded on a high-performance computing cluster supported by an interconnected server network at FH JOANNEUM, utilizing an Anaconda virtual environment running Python and Jupyter-Notebook with FileZilla handling data transfer operations. For neural network construction and training, the authors employed the open-source Keras API operating on top of TensorFlow. The curated dataset of 12,014 pictures was partitioned according to an eighty-twenty ratio, allocating 80 percent of the acquired records—representing 4,806 good images and 4,806 defective images—into a dedicated training directory, while reserving the remaining 20 percent—comprising 1,201 good and 1,201 defective images—in a test directory to support independent verification. The directory hierarchy directly organized samples into subdirectories reflecting class identities, allowing the deep learning framework to parse ground-truth labels during iterative model execution.

Through an empirical development methodology exploring varied topological configurations, the researchers structured the convolutional neural network as a sequential stack of layers optimized for visual feature extraction. The final architecture integrated three major convolutional layers for hierarchical visual representation, interspersed with three max-pooling layers designed to progressively reduce intermediate three-dimensional spatial dimensions by aggregating pixel groups into maximum local activation values. These feature extraction stages fed into a flatten layer, which converted multidimensional feature representations into a single contiguous vector, concluding in a dense layer whose output neurons computed class probability distributions. Model compilation initiated optimization via mini-batch passes utilizing the fit method across ten complete training epochs, with each epoch processing the training data divided into batches of 50 images. Model training executed in approximately two minutes and twenty seconds, while the cumulative pipeline encompassing image loading, training passes, and verification finished within six minutes and thirty seconds.

Quantitative validation over the 2,402 independent test images confirmed that the tuned convolutional neural network attained an overall classification accuracy of approximately 99 percent. Detailed inspection of the resulting confusion matrix revealed robust diagnostic separation between manufacturing classes. Out of 1,201 test images exhibiting verified good components, the network correctly classified 1,197 instances as good while misclassifying only four acceptable items as defective. Concurrently, among the 1,201 test images displaying defective components, the model accurately recognized 1,199 parts as defective, erroneously categorizing just two defective items as good. Despite these exceptional statistical metrics, conventional accuracy calculations could not elucidate which physical attributes guided the model toward its decisions, underscoring the absolute necessity of post-hoc interpretability tools to verify whether classifications derived from genuine component characteristics or superficial background artifacts.

To expose the model’s inner decision logic, the researchers first implemented the LIME framework using library version 0.2.0.1, setting the explainer to probe the two top classification labels. The algorithm functions by creating a local, interpretable surrogate model to approximate the behavior of the complex neural network within the immediate vicinity of a specific test image. LIME segments the target image into contiguous clusters of similarly colored pixels known as superpixels using the quick shift algorithm. The explainer then generates artificial datasets through random perturbation, systematically toggling individual superpixels on or off by replacing inactive segments with a neutral gray tone. The trained convolutional neural network predicts class outputs for these perturbed variations, which are assigned weights based on an exponential distance metric measuring proximity to the original input. Fitting a weighted linear surrogate model yields numerical coefficients that reflect each superpixel’s marginal influence on the final classification decision.

Visual examination of LIME explanations demonstrated that the method effectively localized critical physical regions across the test samples. The framework presented explanations through two graphical formats: isolated top-five superpixel visualizations and colored overlay maps where green highlights indicated superpixels positively supporting the correct classification, while red highlights marked segments exerting negative or neutral influence. On flawless components, LIME confirmed that the model focused heavily on boundary edges and properly tightened screws, with the lower mounting screw exerting the strongest positive contribution. On defective components, the model identified misplaced screws, empty mounting holes, and edge profiles. However, the surrogate modeling technique exhibited notable operational vulnerabilities: repeating the sampling procedure on similar images produced subtle boundary shifts among borderline superpixels, reflecting sampling instability caused by sparse data generation and unconstrained local neighborhood definitions.

The team subsequently implemented the SHAP framework using package version 0.40.0, configuring the primary explainer with an evaluation batch size of 50 and utilizing a dedicated masking object to blur pixel regions during coalition sampling. Grounded mathematically in cooperative game theory, SHAP models each image feature as a player within a coalition and computes Shapley values that represent the average marginal contribution of that feature across all possible feature subsets. In the context of computer vision, SHAP groups contiguous pixels into superpixels and measures their exact additive impact relative to baseline expectations. The resulting visualizations overlay shades of red and blue across the test imagery: darker red hues identify superpixels that strongly increase prediction probability for a designated class, whereas darker blue hues denote superpixels that penalize or decrease that classification probability.

When applied across test images, SHAP generated distinct explanatory insights that contrasted sharply with LIME. Unlike LIME, which consistently placed significant weight on outer product edges, SHAP revealed that the neural network’s decisions were driven almost exclusively by component mounting holes and the precise placement of physical screws. On acceptable components, the darkest red highlights concentrated directly upon the two correctly positioned screws, while subtle blue highlights appeared over the deep black shadows cast beneath them. Because the network associated dark void-like circles with missing screws, the shadow was partially perceived as a hole indicative of a defective component; nevertheless, the overwhelming positive weight of the properly installed screws ensured an accurate overall classification. On defective components exhibiting surplus screws or vacant recesses, dark red regions highlighted the empty holes while misplaced screws received negative attribution, confirming that the network replicated logical human inspection principles.

Direct performance benchmarking under identical computational conditions on the high-performance cluster revealed notable efficiency contrasts between the two frameworks. In this specific industrial image classification setting, SHAP achieved faster execution than LIME, recording a core CPU execution time of 11.8 seconds and an overall wall running time of 2.29 seconds. In comparison, LIME recorded a CPU time of 12.2 seconds and a wall time of 4.8 seconds, with wall times remaining shorter than CPU times due to parallel server processing. Although existing academic literature frequently characterizes LIME as the faster alternative when analyzing low-dimensional text or tabular data, evaluating complex image superpixels inverted this expectation. Furthermore, because SHAP systematically distributes payouts across coalitions rather than relying on randomized local perturbation sampling, its calculated Shapley values provided strict mathematical consistency and complete replicability across repeated evaluations.

Comparative assessment across representation quality, applicability, and algorithmic reliability confirmed that SHAP offered superior analytical power for industrial image inspection. While LIME presented binary positive-or-negative masks that could be supplemented only through secondary heatmaps, SHAP natively rendered expressive, graduated color palettes that enabled quality control engineers to differentiate critical geometric drivers from minor secondary features. Additionally, the experimental findings yielded actionable physical improvements for the production environment: discovering that optical shadows cast by mounting screws were occasionally misinterpreted by the model as structural defects demonstrated that lighting configurations required immediate physical remediation. Stadlhofer and Mezhuyev concluded that while both frameworks successfully illuminate black-box networks, SHAP provides the most dependable, mathematically rigorous, and effective interpretability method for convolutional vision systems operating within industrial quality assurance pipelines.

Subject of Research: Interpretability methods for convolutional neural networks in industrial image classification

Article Title: Approach to provide interpretability in machine learning models for image classification

Article References: Stadlhofer, A., & Mezhuyev, V. (2023). Approach to provide interpretability in machine learning models for image classification. Industrial Artificial Intelligence, 1(1), Article 10. https://doi.org/10.1007/s44244-023-00009-z

Image Credits: AI Generated

DOI: 10.1007/s44244-023-00009-z

Keywords: Machine Learning Interpretability, SHAP, LIME, Convolutional Neural Networks, Industrial Quality Control, Computer Vision, Image Classification, Explainable Artificial Intelligence, Superpixels, Shapley Values, Deep Learning, Smart Factory, Industry 4.0, Surrogate Models, Model Transparency, Manufacturing Automation

Cite Scienmag News

Blake Davidson. (November 19, 2026). SHAP Outperforms LIME in Explaining Industrial Image Quality Control Models. Scienmag. https://scienmag.com/shap-outperforms-lime-in-explaining-industrial-image-quality-control-models/

Blake Davidson. "SHAP Outperforms LIME in Explaining Industrial Image Quality Control Models." Scienmag, 19 November 2026, https://scienmag.com/shap-outperforms-lime-in-explaining-industrial-image-quality-control-models/. Accessed 19 November 2026.

Blake Davidson. "SHAP Outperforms LIME in Explaining Industrial Image Quality Control Models." Scienmag. November 19, 2026. https://scienmag.com/shap-outperforms-lime-in-explaining-industrial-image-quality-control-models/

Tags: computer visionconvolutional neural networksdeep learningExplainable Artificial Intelligenceimage classificationindustrial quality controlIndustry 4.0LIMEmachine learning interpretabilitymanufacturing automationModel TransparencySHAPShapley valuessmart factorysuperpixelssurrogate models
Share26Tweet16
Previous Post

Survey Identifies Major Drivers Behind Older Adults Skipping COVID and Flu Vaccines

Next Post

Fracture Patients Show Lower NK Cells and Distinct Cytokine Profile

Related Posts

Transformer Framework Improves Precision in Perceptual Audio Loudness Estimation
Science News

Transformer Framework Improves Precision in Perceptual Audio Loudness Estimation

November 19, 2026
Statistical Analysis Reveals Climatological Patterns of Obonai-dashi Winds in Japan
Science News

Statistical Analysis Reveals Climatological Patterns of Obonai-dashi Winds in Japan

November 19, 2026
Indonesian Port Study Reveals Automation Requires Continuous Organizational And Operational Adaptation
Science News

Indonesian Port Study Reveals Automation Requires Continuous Organizational And Operational Adaptation

November 19, 2026
Study in Saudi Arabia Reveals Distinct Gender Differences in Mental Health Stigma
Medicine

Study in Saudi Arabia Reveals Distinct Gender Differences in Mental Health Stigma

November 19, 2026
Review Details Cellular Mechanisms and Bioactive Compounds Driving Cold-Adaptive Mammalian Thermogenesis
Science News

Review Details Cellular Mechanisms and Bioactive Compounds Driving Cold-Adaptive Mammalian Thermogenesis

November 19, 2026
Multivariate Tail Risk Measure Computes Copula Expectations Through One-Dimensional Formula
Science News

Multivariate Tail Risk Measure Computes Copula Expectations Through One-Dimensional Formula

November 19, 2026
Next Post
Fracture Patients Show Lower NK Cells and Distinct Cytokine Profile

Fracture Patients Show Lower NK Cells and Distinct Cytokine Profile

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Transformer Framework Improves Precision in Perceptual Audio Loudness Estimation
  • Statistical Analysis Reveals Climatological Patterns of Obonai-dashi Winds in Japan
  • Indonesian Port Study Reveals Automation Requires Continuous Organizational And Operational Adaptation
  • Study in Saudi Arabia Reveals Distinct Gender Differences in Mental Health Stigma

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading