Industrial Visual Anomaly and Defect Detection
Finding defects on products and equipment by learning what normal looks like, often with few or no defect examples.
Tier Grades
Tier A: Frontier AI labs
stretch
This is applied work; frontier labs rarely lead it, though Amazon published PatchCore.
Tier B: Indonesian AI labs
feasible
Methods need only normal images and modest compute, which suits an Indonesian lab.
Tier C: Institution units such as ITB
feasible
A 6-author team set the reference result of 99.6% AUROC on MVTec AD; a university unit can work on it.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Training on normal images only fits client factories where defects are rare.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeTraining compute is not in the sources read.
- DataMVTec AD benchmark, where PatchCore reaches up to 99.6% image-level AUROC.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 6 authors on the PatchCore paper (arXiv 2106.08265).
- Time horizonMonths to a year.
- Shelf-life signalsZero-shot anomaly detection with vision-language models, MVTec AD saturates and new benchmarks replace it
- Pipeline roleTier C and industry publish methods; Tier B and Tier C build local benchmarks; Tier D runs inspection pilots on client lines.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
EfficientADvisionexemplarAnomalibvisionexemplar
PatchCorevisionexemplar
PaDiMvisionexemplar
Dinomaly2visionproduced byINP-Formervisionproduced by
SuperSimpleNetvisionproduced by
GLASSvisionproduced by
Dinomalyvisionproduced by
AnomalyDINOvisionproduced byWinCLIPvisionproduced by
U-Flowvisionproduced by
CFAvisionproduced by
Reverse Distillationvisionproduced byFastFlowvisionproduced by
CS-Flowvisionproduced byDRAEMvisionproduced by
C-Flow (CFLOW-AD)visionproduced bySTFPMvisionproduced by
GANomalyvisionproduced by
Kunlun Large Modelvisionproduced by
Pangu CVvisionproduced by
Intel weld porosity detectionvisionproduced by
Optical Inspection and PCB Defect Classificationvisionproduced byLinks
R&D TopicIndustrial Visual Anomaly and Defect Detection
ModelsAnomalibvisionPatchCoreAmazon, visionPaDiMCEA (French Alternative Energies and Atomic Energy Commission), visionEfficientADMVTec Software GmbH, vision+22 more
PatchCoreAmazon, visionPaDiMCEA (French Alternative Energies and Atomic Energy Commission), visionEfficientADMVTec Software GmbH, vision+22 more
Models 26
Dinomaly2Tsinghua University, Beijing Institute of Technology, Shanghai Jiao Tong University, Fudan University, visionproduced byKunlun Large ModelChina National Petroleum Corporation, visionproduced byPangu CVHuawei (Huawei Cloud), visionproduced by
FOMO-AD (Visual Anomaly Detection)Edge Impulse, TinyMLproduced byCS-FlowLeibniz University Hannover, visionproduced byDinomalyBeijing Institute of Technology, visionproduced byDRAEMUniversity of Ljubljana, visionproduced byEfficientADMVTec Software GmbH, visionexemplarGLASSInstitute of Automation, Chinese Academy of Sciences, visionproduced byINP-FormerTsinghua University, Huazhong University of Science and Technology, visionproduced byIntel weld porosity detectionIntel, visionproduced byOptical Inspection and PCB Defect ClassificationNVIDIA, visionproduced by+14
PaDiMCEA (French Alternative Energies and Atomic Energy Commission), visionexemplarPatchCoreAmazon, visionexemplarSuperSimpleNetUniversity of Ljubljana, visionproduced byU-FlowDigital Sense, Universidad de la Republica, visionproduced byVisual ChangeNetNVIDIA, visionproduced byWinCLIPAmazon, KAIST, visionproduced byAnomalyDINORuhr University Bochum, visionproduced byC-Flow (CFLOW-AD)Panasonic, visionproduced byCFAInha University, visionproduced byGANomalyDurham University, visionproduced byReverse DistillationUniversity of Alberta, visionproduced byAnomalibvisionexemplarFastFlowvisionproduced bySTFPMvisionproduced byPapers
- Uninformed Students: Student-Teacher Anomaly Detection with Discriminative Latent EmbeddingsPaul Bergmann et al., CVPR 2020 (per arXiv comment), 2019 founding
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language ModelsZhaopeng Gu et al., AAAI 2024 (per arXiv comment), 2023 state of the art
- Towards Total Recall in Industrial Anomaly DetectionKarsten Roth et al., CVPR 2022 (per arXiv comment), 2021 state of the art
- PaDiM: a Patch Distribution Modeling Framework for Anomaly Detection and LocalizationThomas Defard et al., arXiv preprint, 2020 state of the art
- Deep Industrial Image Anomaly Detection: A SurveyJiaqi Liu et al., Machine Intelligence Research vol. 21, no. 1, pp. 104-135, 2024, 2023 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2106.08265 https://arxiv.org/abs/2106.08265
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Industrial Visual Anomaly and Defect Detection", https://aurion.semai.id/topics/industrial-visual-anomaly-detection/, records of October 2026.