Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Evaluating Foundation Models For Industrial Surface Defect Detection
2026 (English)Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE creditsStudent thesis
Abstract [en]

Industrial surface defect inspection is important for maintaining product quality, but reliable automatedinspection remains challenging. Defects can be small, subtle, and diverse, and their visual appearancecan be affected by complex surface backgrounds, lighting conditions, and viewpoint changes. In addi-tion, industrial datasets often contain limited labeled defect samples, which makes it difficult to traintask-specific deep learning models from scratch, especially when new products or production conditionsrequire model adaptation. This thesis evaluates the practical potential of pretrained foundation models(FMs) for industrial surface defect inspection, focusing on classification, object detection, and stabilityunder image changes. The study used the GC10-DET dataset and evaluated three pretrained foundationmodels: Contrastive Language–Image Pre-training (CLIP) as a prompt-based vision-language model forzero-shot defect classification, Grounding DINO as a text-guided object detector for zero-shot object de-tection, and DINOv3 as a self-supervised vision foundation model. DINOv3 was used as a pretrainedvisual feature extractor for classification and as the backbone in a Faster R-CNN detector to evaluatehow fine-tuning affected object detection performance. Model stability was assessed under geometrictransformations and changes in brightness and contrast. CLIP showed limited performance for defectclassification and was sensitive to prompt wording, although it performed better in the defective ver-sus non-defective prompt comparison than in identifying the exact defect type. DINOv3 features per-formed well for classification when combined with a lightweight classifier. Grounding DINO showedlimited zero-shot localization performance. In the DINOv3 object detection experiments, performancevaried with the fine-tuning strategy and the way the pretrained features were integrated into the detector.The stability results indicated smaller changes under horizontal flips and larger changes under rotations.Taken together, the findings suggest that foundation models can provide useful starting points for indus-trial inspection, but further evaluation is needed to understand how well these models transfer to differentindustrial conditions.

Place, publisher, year, edition, pages
2026. , p. 56
Keywords [en]
foundation models, industrial surface defect detection, zero-shot, fine-tuning
National Category
Computer graphics and computer vision Engineering and Technology Robotics and automation
Identifiers
URN: urn:nbn:se:hj:diva-71475OAI: oai:DiVA.org:hj-71475DiVA, id: diva2:2063727
External cooperation
RISE Research Institutes of Sweden
Presentation
2026-05-26, 15:05 (English)
Supervisors
Examiners
Available from: 2026-06-17 Created: 2026-05-29 Last updated: 2026-06-17Bibliographically approved

Open Access in DiVA

fulltext(65161 kB)94 downloads
File information
File name FULLTEXT01.pdfFile size 65161 kBChecksum SHA-512
3933f9aca1521e67fdccafa12e2f36166633f311aac6c9910a0c3894b18c49d1b868b5c4f8abdf9aef3e3ff2b924de4c95f2a0fd66e22dd6890d5f760b2935f2
Type fulltextMimetype application/pdf

Computer graphics and computer visionEngineering and TechnologyRobotics and automation

Search outside of DiVA

GoogleGoogle Scholar
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

urn-nbn

Altmetric score

urn-nbn
Total: 239 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf