The review synthesizes 55 studies published between 2015 and 2025, spanning textile, steel, wood, ceramic, and electronics manufacturing. What sets it apart from earlier surveys is its methodology. The authors borrowed the PICOS framework—Population, Intervention, Comparator, Outcomes, Study design—from evidence-based medicine, where it is standard practice, and applied it to computer vision literature, where structured review protocols remain exceedingly rare. They extended it with a sixth dimension, Evidence Source, to assess dataset accessibility and reproducibility. Starting from 1,875 records identified in Scopus, IEEE Xplore, ScienceDirect, and SpringerLink, the team filtered down through PRISMA-guided screening to their final corpus of 55 empirical studies, each coded for material type, architecture family, baseline comparisons, performance metrics, and whether the system could actually translate detection outputs into industrial quality grades.

The architectural story the review tells is one of steady escalation. In the early period, convolutional neural networks dominated, establishing that learned representations outperform hand-crafted features in production settings. Pixel-level segmentation networks improved accuracy on microdefects, while multi-stream CNNs used parallel processing paths to capture complementary local and global features. Cascade pyramid architectures and multi-level feature fusion then made systems more resistant to variations in defect size, a persistent headache on production lines where flaws range from pinholes to sprawling dye irregularities. These convolutional foundations remain relevant today, particularly in the YOLO family of real-time detectors.

The paradigm shift arrived in 2024 and 2025 with the systematic integration of Transformer architectures. Unlike CNNs, which process local receptive fields hierarchically, Transformers employ self-attention to capture global context and long-range dependencies across entire images. The review highlights several landmark systems: a Swin Transformer-based defect detector that replaces conventional CNN backbones in Faster R-CNN, achieving 81.05 percent mean average precision on the NEU-DET steel benchmark and 99 percent on printed circuit board data; a global dual-attention Transformer that reaches an F-beta score of 0.836 on the Severstal steel dataset; and an adaptive cross-Transformer with self-supervised contrastive learning that enables few-shot adaptation to novel defect types, addressing the chronic scarcity of labeled industrial data. Meanwhile, hybrid designs flourished, pairing CNN feature extractors with Transformer attention modules to balance computational efficiency against global sensitivity.

Yet the review is refreshingly candid about the costs. Transformers carry a computational overhead of three to ten times that of optimized CNNs, a serious barrier for edge devices on resource-constrained production lines. Only two of the twelve Transformer studies the authors examined reported edge-device deployment results, raising questions about industrial readiness beyond laboratory validation. Convolutional engineering fought back impressively: YOLOv10 achieved end-to-end detection without non-maximum suppression post-processing, cutting latency by 40 percent, while YOLO-World extended detection to open vocabularies through vision-language pre-training, enabling zero-shot detection across more than 1,200 categories at 52 frames per second. The emerging consensus, the authors argue, favors hybrid architectures that synergize global modeling with convolutional inductive biases.

The review’s most original conceptual contribution is the notion of grade-readiness: the capability of a detection system to translate raw model outputs—bounding boxes, segmentation masks, confidence scores—into industrial quality classifications aligned with production standards such as ISO 12945-1 for textile pilling, ISO 21920 for metal surface texture, and ASTM specifications for stainless steel grading. The pipeline the authors envision runs from an inspection image through a detection model to multi-dimensional defect characterization—type, location, size, severity, and count—and finally to an actionable grade: Pass or Grade A, Grade B, Grade C, or Fail. The sobering finding is how rarely this bridge is actually built. Although 87.5 percent of the reviewed studies discuss industrial quality integration in their introductions, only seven of 55—about 12.5 percent—operationalize or demonstrate grade-mapping workflows. Detection metrics like mean average precision routinely exceed 0.90 in controlled benchmarks, but the translation into quality decisions that buyers and regulators recognize remains largely aspirational.

Cross-domain analysis reveals material-specific patterns that practitioners will find immediately useful. Textiles demand multi-scale architectures to accommodate diverse fabric textures and suffer from small-defect and data-imbalance problems; an enhanced YOLOv7-tiny there achieved 84 percent mAP at 90 frames per second on a proprietary dataset. Steel inspection, the dominant domain at 37.5 percent of the corpus, prioritizes real-time processing for production-line integration, with systems like IDP-Net reaching 98.7 percent mAP at 50 frames per second. Wood surfaces benefit from texture-aware features that distinguish genuine defects from natural grain variations, while particleboard grading systems have demonstrated full detection-to-grade pipelines. Ceramics and electronics remain comparatively underexplored. The catch, the authors stress, is that cross-domain comparisons are often impossible: textiles report full detection metrics while metals emphasize segmentation scores, and evaluation protocols differ so widely that state-of-the-art claims frequently cannot be tested against one another.

Reproducibility emerges as the field’s uncomfortable Achilles heel. Only about 47 percent of the datasets used across the corpus are fully public, while 14 percent are proprietary and 5 percent undisclosed. The authors’ risk-of-bias stratification sorted the 55 studies into three clusters: 78.2 percent were high quality with low bias, but 14.5 percent were high-quality studies with serious reproducibility concerns—no code repositories, no random seeds, proprietary test sets that may inflate accuracy claims—and 7.3 percent showed high risk across all six assessed bias domains and were downweighted in the synthesis. Sixty percent of studies report accuracy without any computational efficiency figures, and a quarter apply custom train-test splits to public benchmarks without disclosing split ratios or seeds. The authors recommend a minimum reporting checklist requiring both accuracy and efficiency metrics, transparent split documentation, and routine release of inference code and trained weights—practices that require no new data collection, only more complete reporting of experiments already being run.

Looking forward, the review identifies three converging directions. Self-supervised learning could reduce dependence on large annotated defect datasets by learning transferable representations from unlabeled normal-surface imagery. Standardized multimodal fusion protocols could extend the nascent RGB-depth and multi-sensor work already showing promise. And foundation models—promptable segmentation systems like SAM, vision-language models like CLIP, and industrial anomaly adaptations like AnomalyGPT—remain conspicuously absent as genuine tools in the reviewed corpus, even though they could enable zero-shot detection of defect categories no one has labeled. The authors’ verdict is that surface defect detection has matured from a research curiosity into a strategic instrument of global quality assurance, but that its promise will only be fulfilled when detection, grading, and standardized assessment are connected into a coherent, verifiable ecosystem. Fabric Quality 4.0, they conclude, must move beyond rhetoric to become an industrial reality—and this review provides both the map and the measuring stick for getting there.

Subject of Research: Automated surface defect detection and quality grading using deep learning across manufacturing industries

Article Title: Conceptualizing automated surface quality: A cross-domain evidence synthesis from CNNs to transformers

Article References: Dharma, F. P., Singgih, M. L., & Prastyo, D. D. (2026). Conceptualizing automated surface quality: A cross-domain evidence synthesis from CNNs to transformers. Machine Learning with Applications, 26, Article 100989. https://doi.org/10.1016/j.mlwa.2026.100989

Image Credits: AI Generated

DOI: 10.1016/j.mlwa.2026.100989

Keywords: surface defect detection, deep learning, convolutional neural networks, vision transformers, quality grading, textile inspection, steel inspection, systematic review, PICOS framework, YOLO, industrial automation, reproducibility