AI-enabled defect detection in industrial products: A comprehensive survey, key insights and future research challenges

Abstract

Artificial Intelligence (AI) techniques, especially Machine Learning (ML) and Deep Learning (DL), are increasingly used for automated defect inspection in industries like metals, ceramics, glass, and textiles. These methods process high-quality images to detect and localise defects. Inspection approaches range from manual and semi-automated (classical ML) to fully automated (DL), with DL preferred for eliminating manual feature engineering. DL models are typically supervised or unsupervised—supervised models offer high accuracy but require large labelled datasets, while unsupervised models work without labels but often yield lower performance. Emerging alternatives like semi-supervised, weakly supervised, and self-supervised learning show promising results in data-scarce settings. This paper presents a comprehensive taxonomy of DL models for defect classification, detection, and segmentation. For classification, models such as Convolutional Neural Networks (CNNs), transfer learning, and hybrid architectures are discussed. For segmentation tasks, encoder–decoder models, pyramid networks, and attention mechanisms are reviewed. In defect detection, the evolution of YOLO architectures is examined, particularly combinations with pyramid networks and attention modules to enhance feature extraction. Unsupervised approaches, such as Generative Adversarial Networks (GANs) and diffusion models, are explored for their capacity to detect anomalies without extensive labelled datasets. Emerging strategies — including transformer-based models, few-shot learning, large language models (LLMs), and foundation models — are also covered. Despite recent advancements, challenges remain in ensuring model reliability due to overconfidence and poor calibration. We address this by discussing model trustworthiness, human-in-the-loop frameworks, and reliability assessment techniques critical for industrial deployment. This review analyses the current strengths and limitations of existing approaches, identifies ongoing challenges, and highlights future directions in ML and DL-based defect inspection across various materials and defect types. • Provides a comprehensive review of recent progress in defect detection. • Discusses various image processing techniques for defect identification and localisation. • Highlights the importance of human-in-the-loop analytics in industrial model deployment. • Presents key challenges in model training, detection accuracy, and calibration.

Topics & Concepts