The short answer
Industrial visual defect detection is shifting from “classical CV + CNN everywhere” to a cooperative pattern: traditional models handle the main line, foundation models backstop the edge cases. In 2026, high-speed production lines use YOLO-family detectors, diverse low-volume production leans on zero-shot VLM inspection, and interpretability requirements keep threshold-based methods alive. No single route covers all scenarios — the first rule of selection is to know your defect types, line cadence and budget before picking anything.
Why industrial QC became AI vision’s fastest landing zone
Quality inspection is one of the earliest AI vision use cases to scale, driven by three weaknesses of manual inspection: inconsistent judgment, fatigue-driven miss rates, and chronic labor shortages. On September 24, 2026, Siemens and Procter & Gamble announced they were extending their co-developed AI visual inspection system from pilot lines to multiple plants worldwide, aiming at consistent defect detection on high-cadence lines. That is a “pilot to scale” milestone, and it matches Liu Gong’s long-held observation: inspection is the manufacturing use case with the clearest ROI math and the hardest budget to veto.
Route one: classical machine vision and CNN — mature, with a visible ceiling
Classical approaches combine thresholding, morphology and edge operators with a classifier such as ResNet. The strengths are real: millisecond inference, minimal training data, and pixel-level explainability. The weakness is generalization — a change in lighting or a new material batch means re-tuning parameters, and unseen defect classes are simply out of scope. Liu Gong has spent years around production lines and sees the same pattern everywhere: these systems are stable, cheap and maintainable, but “writing a new rule for every new defect” is their biggest hidden cost.
Route two: YOLO-family detectors — the production-line workhorse
Single-stage detectors, most notably YOLO variants, are the de facto standard on high-speed lines today. Industry deployment data shows AI vision systems built on YOLO variants reach 95-99% detection accuracy in production, with pilot costs typically ranging from $50,000 to $250,000 and enterprise rollouts from $500,000 to $2 million. The model balances speed, accuracy and deployment complexity, and with labeled data you can train it on exactly the defects you care about. The trade-offs are equally clear: it needs sufficient annotated samples, continuous iteration, and it cannot respond to defect classes it has never seen.
Route three: VLMs and foundation models — the new paradigm
Since 2024, vision-language models have pushed defect detection into a new stage, with zero- and few-shot generalization as the headline promise: no more retraining for every new defect. Public progress in 2026 has been dense:
- AD-Copilot reached 82.3% accuracy on the MMAD benchmark, near ordinary human level, improving localization accuracy 3.35x over a Qwen2.5-VL baseline on MMAD-BBox, outperforming both proprietary and open models;
- DCS (a Grounding DINO + CLIP + SAM combination) topped AUROC, F1-max and AP on MVTec AD and VisA, gaining 21.4% AUROC over the CLIP-SAM method;
- EIAD and similar frameworks deliver zero-shot defect segmentation and localization across MVTec AD, ViSA and MVTec 3D-AD;
- GenAU unifies image-level detection, pixel-level segmentation and multi-type defect identification in one instruction-following model, while Myriad attempts to have a large model orchestrate classical vision-expert models.
This route solves the “retrain on every new scene” pain point, but it carries real costs: inference overhead, deployment complexity, and data-compliance questions — sending production images to a cloud VLM is unacceptable to many factories.
Key benchmarks: which numbers not to be fooled by
Four benchmarks dominate the field: MVTec AD (mainstream 2D anomaly detection), VisA (complex-background anomalies), MVTec 3D-AD (3D defects) and MMAD (multimodal industrial anomalies, an emerging 2026 benchmark). Liu Gong’s advice is to watch three numbers: AUROC (separating normal from defective), F1-max (the best precision-recall balance), and above all zero-shot performance on new defect classes. Leaderboard rankings change quickly year over year; what production actually cares about is whether the system holds up when a defect it has never seen arrives.
Cost and deployment: how to run the numbers
| Dimension | Classical CV + CNN | YOLO-family | VLM / foundation models |
|---|---|---|---|
| Production accuracy | 85-95% by scene | 95-99% | benchmark-leading, TBD on lines |
| New-defect adaptation | rewrite rules | re-annotate and retrain | zero / few-shot |
| Data dependency | low | medium-high (labels) | low (transfer) |
| Per-image inference cost | lowest | low | high (especially cloud) |
| Explainability | strong (pixel-level) | medium (box + class) | medium (explanatory text) |
| Deployment cost | lowest | $50K-250K pilot | highest compute + compliance |
| Typical scenarios | stable lighting, single defect | real-time high-speed lines | diverse low-volume, many defect types |
Liu Gong’s take
Liu Gong’s judgment is that the mainstream pattern for the next two to three years will not be “foundation models replace classical ones” but a cooperative architecture — YOLO-family and CNN models run the real-time main inspection on the line, while VLMs only backstop low-confidence samples and unseen defects. The reasons are practical: line cadence cannot afford a heavy model on every frame; factories are extremely sensitive to inference cost; and quality-compliance demands traceable, explainable results that a pure black box cannot pass.
Liu Gong believes most small and mid-sized factories should still choose YOLO-family plus classical operators in 2026, spending budget on labeling quality and lighting fixtures rather than chasing new models. Only high-variety, low-volume scenarios with frequently changing defect classes — such as non-standard 3C parts and new-material surfaces — justify paying for VLM zero-shot capability. The real bottleneck for VLM adoption is not model capability but deployment cost and data compliance; until those are solved, foundation models in industrial inspection remain an evaluation exercise rather than a wholesale replacement.
Selection advice: match the scenario
- High-speed line, stable defect types: YOLO-family plus classical operators, balancing cadence and cost;
- Diverse low-volume, ever-changing defects: start with VLM zero-shot validation before investing;
- Explainability and compliance first: classical thresholding plus CNN with pixel-level output;
- Tight pilot budget: start from open-source YOLO; a five-figure budget can validate ROI.
Sources and further reading
Sources: Acta Automatica Sinica survey “Visual Industrial Defect Detection: Evolution and Synergy from Non-Foundation to Foundation Models” (2026); arXiv AD-Copilot (MMAD 82.3%, 3.35x MMAD-BBox); DCS zero-shot anomaly detection (MVTec-AD AUROC +21.4%); Jidoka deployment-cost report (95-99% accuracy, $50K-250K pilots); Siemens & P&G AI inspection expansion (2026-09-24). Figures reflect the original sources as of October 2026.
Further reading: 2026 Frontier LLM Comparison Guide ()
Daily aggregation of overseas AI developments with Liu Gong’s commentary. Bookmark this site; return to the homepage for more.