[Paper Review] Quantum artificial vision for defect detection in manufacturing

Daniel Guijo, Victor Onofre+8 more|arXiv (Cornell University)|Aug 9, 2022

Integrated Circuits and Semiconductor Failure Analysis4 citations

TL;DR

This paper proposes and benchmarks two quantum machine learning approaches—quantum Support Vector Machine (QSVM) on gate-based quantum hardware and QBoost on a quantum annealer—for defect detection in automotive manufacturing using the GDXray+ dataset. It demonstrates that both quantum methods outperform classical counterparts in precision, with QBoost enabling industrially relevant problem sizes on current NISQ devices without extensive preprocessing.

ABSTRACT

In this paper we consider several algorithms for quantum computer vision using Noisy Intermediate-Scale Quantum (NISQ) devices, and benchmark them for a real problem against their classical counterparts. Specifically, we consider two approaches: a quantum Support Vector Machine (QSVM) on a universal gate-based quantum computer, and QBoost on a quantum annealer. The quantum vision systems are benchmarked for an unbalanced dataset of images where the aim is to detect defects in manufactured car pieces. We see that the quantum algorithms outperform their classical counterparts in several ways, with QBoost allowing for larger problems to be analyzed with present-day quantum annealers. Data preprocessing, including dimensionality reduction and contrast enhancement, is also discussed, as well as hyperparameter tuning in QBoost. To the best of our knowledge, this is the first implementation of quantum computer vision systems for a problem of industrial relevance in a manufacturing production line.

Motivation & Objective

• To evaluate the feasibility of quantum machine learning for industrial defect detection in manufacturing.
• To compare quantum algorithms (QSVM and QBoost) against classical counterparts on real-world, unbalanced image data.
• To assess the performance of quantum vision systems under realistic constraints, including dimensionality reduction and hyperparameter tuning.
• To explore the practical deployability of quantum-trained models in industrial settings using current NISQ hardware.
• To determine whether quantum advantages in precision and inference efficiency can be achieved with today's noisy quantum devices.

Proposed method

• Employed a quantum Support Vector Machine (QSVM) on a universal gate-based quantum computer for binary image classification of defective vs. non-defective car parts.
• Applied QBoost, a quantum-enhanced boosting algorithm, on a D-Wave quantum annealer to train a classifier using weak estimators and quantum annealing for optimization.
• Preprocessed images via dimensionality reduction (PCA) and contrast enhancement to improve signal-to-noise ratio and reduce computational load.
• Tuned hyperparameters in QBoost, including regularization strength and initial number of weak classifiers (Decision Tree Classifiers), to balance performance and inference speed.
• Used classical counterparts (SVM, AdaBoost) as baselines for direct comparison on the same dataset and preprocessing pipeline.
• Evaluated models using standard metrics: precision, recall, F1-score, and inference time per image.

Experimental results

Research questions

• RQ1 Can quantum machine learning algorithms achieve higher classification precision than classical models for defect detection in manufacturing?
• RQ2 To what extent do quantum algorithms like QSVM and QBoost outperform classical SVM and AdaBoost on unbalanced, real-world industrial image data?
• RQ3 How does dimensionality reduction and contrast enhancement affect the performance of quantum vision systems on NISQ devices?
• RQ4 Can QBoost-trained models be deployed efficiently on classical hardware after quantum training, enabling practical industrial use?
• RQ5 What is the trade-off between inference time and model performance in QBoost as a function of hyperparameters like regularization and number of weak classifiers?

Key findings

• The QBoost algorithm on a D-Wave quantum annealer achieved higher precision than classical AdaBoost, with a best F1-score of 0.94 using 50 initial weak classifiers and regularization parameter 0.
• Even after aggressive dimensionality reduction (PCA), QBoost maintained high performance, demonstrating robustness to data compression.
• Inference time for QBoost remained stable across hyperparameter variations, averaging 2.94 milliseconds per image, with no significant increase despite higher classifier counts.
• The QBoost model trained on quantum hardware could be deployed and executed on classical systems, enabling practical industrial deployment without requiring full quantum infrastructure.
• QSVM showed strong performance but required full quantum execution for both training and inference, limiting its current deployability compared to QBoost.
• The study demonstrates, to the best of the authors' knowledge, the first implementation of quantum computer vision for an industrially relevant defect detection task on real manufacturing data.