Computer vision (CV) systems driven by artificial intelligence (AI) are increasingly replacing manual and conventional rule-based inspection procedures in industrial quality inspection, enabling automated, real-time, and data-driven decision-making based on visual data. Conventional inspection procedures are usually limited in terms of scalability, human-related errors, and high operational costs, which drives the increasing reliance on smart vision-based technologies. A practical, practice-oriented review of AI-based computer vision systems for industrial quality control is provided in this paper, with emphasis on real-world deployment issues and performance aspects. Two representative industrial case studies are examined. The first investigates real-time extrusion monitoring in robotic building construction, where geometric deviations, bead-width variation, surface irregularities, and process inconsistencies are detected during material deposition using vision-based monitoring and image-processing pipelines. The second case study focuses on automated inspection of bolts and screws in manufacturing lines, addressing presence detection, orientation recognition, and defect classification under high-speed production conditions. In both cases, widely adopted vision and AI techniques, including image-processing pipelines, convolutional neural networks, and edge-computing hardware, are discussed and compared. The analysis shows that AI-enabled computer vision systems can outperform traditional rule-based or manual solutions in terms of inspection accuracy, consistency, and throughput when they are supported by reliable acquisition, representative data, and robust industrial integration. Nevertheless, challenges related to dataset quality, model generalization, lighting variability, and real-time computational constraints remain critical in industrial environments. In conclusion, AI-based computer vision plays a central enabling role in intelligent quality inspection within the context of Industry 5.0. Future research should focus on adaptive model capabilities, tighter integration with cyber-physical systems, and scalable deployment strategies to achieve reliable and autonomous inspection across diverse industrial sectors.
Industrial quality inspection determines whether products satisfy dimensional, visual, structural, and functional requirements before they move to the customer or to the next production stage. Manual inspection, sampling, rule-based image processing, and post-process measurement remain useful, but they become less sufficient when production rates increase, product variants multiply, and defects are small, rare, or difficult to express through fixed thresholds. Conventional rule-based Machine Vision (MV) is fast and repeatable, but it usually requires stable lighting, consistent object presentation, and manually tuned thresholds. Modern factories, therefore, need machine vision systems (MVSs) that can handle visual variability, support real-time decisions, and provide traceable quality evidence
[1][2][1,2].
The motivation for Artificial Intelligence (AI)-powered Computer Vision (CV) is not simply to replace human inspectors. The stronger objective is to build closed-loop quality systems in which visual data are captured, interpreted, stored, and connected to production action. In this view, the camera is part of a cyber-physical quality architecture that supports defect detection, process correction, root-cause analysis, product genealogy, and quality assurance. This makes AI-powered inspection closely connected to Industry 4.0, Quality 4.0, and Industry 5.0, where digitalization, traceability, human-centeredness, resilience, and sustainability are central requirements
[3][4][5][6][3,4,5,6].
Despite progress in Deep Learning (DL) and Machine Learning (ML), many inspection deployments remain limited to pilots because industrial inspection is not only a classification problem. A production-line system must coordinate illumination, optics, sensor placement, acquisition timing, calibration, preprocessing, inference, decision logic, actuator triggering, data logging, and quality traceability. A model that performs well on a benchmark can fail on the factory floor because of lighting changes, dust, vibration, reflections, camera aging, product variation, class imbalance, or unseen defects. Recent real-world benchmarks make this gap more visible: they introduce multi-view production images, real defects, changing illumination, and more difficult object conditions that are not well represented by older laboratory datasets
[7][8][9][10][7,8,9,10]. These observations show that the central challenge is not benchmark accuracy alone, but its translation into reliable industrial-quality action
[1][11][1,11].
Existing reviews and case studies often emphasize isolated algorithms, benchmark results, or individual applications. Less attention is usually given to the complete inspection architecture and the deployment conditions that decide whether a system is usable in production. This paper reviews AI-powered CV for industrial quality inspection as an integrated production system, organized from quality-management requirements and production-line control to acquisition hardware, visual processing, model inference, decision logic, validation, and research gaps
[1][2][4][11][1,2,4,11].
The scope is intentionally narrower than an exhaustive survey of all computer-vision applications used for quality assessment. Application domains such as fruit quality assessment and textile defect inspection involve different defect definitions, sensing conditions, datasets, process constraints, and evaluation criteria. This review therefore does not attempt a separate comprehensive survey of each application sector; domain-specific studies are considered only where they provide evidence relevant to the system-level engineering and deployment issues examined here. The focus of this review is therefore the cross-cutting systems-engineering problem: how image acquisition, visual interpretation, real-time computation, industrial decision logic, physical response, traceability, and lifecycle validation must work together for reliable deployment.
The contributions of this paper are as follows:
-
It defines the relationship between AI, ML, DL, CV, MV, and MVSs in industrial quality inspection.
-
It organizes visual inspection from the industrial quality system to the camera, algorithm, edge device, reject actuator, operator interface, and quality database.
-
It critically compares manual inspection, rule-based MV, supervised learning, unsupervised anomaly detection, and edge-deployed DL.
-
It analyzes two illustrative industrial application examples and extracts general lessons about real-time monitoring, embedded deployment, defect taxonomy, and process feedback.
-
It identifies research gaps in robustness, dataset quality, explainability, traceability, edge computing, and human-centered inspection.
This paper is organized from macro to micro.
Section 2 describes the review methodology and systems-engineering evaluation sequence.
Section 3 clarifies core terms.
Section 4 presents the complete inspection architecture from enterprise and quality requirements to acquisition, inference, decision, response, traceability, and lifecycle management.
Section 5 compares major visual interpretation methods.
Section 6 analyzes two representative implementations, while
Section 7 summarizes deployment challenges, standards, and research gaps.
