Surface Defect Detection
Detecting scratches, dents, and surface imperfections on materials.
Surface defect detection identifies manufacturing flaws on diverse materials — metal, wood, fabric, semiconductors, glass — using computer vision. The challenge is generalizing across materials and defect types, with anomaly detection approaches (PatchCore, EfficientAD) increasingly preferred over task-specific supervised models.
History
DAGM dataset published — synthetic textile defect detection benchmark
Deep learning surpasses traditional machine vision (Haar, HOG) for surface inspection
MVTec AD includes multiple surface texture categories (carpet, grid, leather, tile, wood)
Magnetic tile dataset and KolektorSDD released for specialized surface defect detection
PatchCore and PaDiM achieve >99% detection on MVTec texture categories
Semi-supervised approaches reduce annotation requirements to 5-10% of data
Foundation model features (DINOv2) improve surface defect detection across materials
Multi-material surface inspection systems deployed in electronics manufacturing
Zero-shot defect detection via CLIP/LMM enables rapid deployment on new materials
How Surface Defect Detection Works
Surface Imaging
Controlled lighting (diffuse, directional, structured) captures surface texture with consistent illumination to reveal defects.
Normal Texture Learning
The system learns what defect-free surfaces look like — either through supervised defect examples or unsupervised normal-only training.
Defect Detection
Anomaly scoring identifies regions that deviate from normal texture patterns — scratches, dents, discoloration, inclusions.
Defect Classification
Detected anomalies are classified by type (crack, pit, stain, deformation) if sufficient labeled data exists.
Severity Assessment
Defect size, depth, and location determine severity grading for accept/reject decisions.
Current Landscape
Surface defect detection in 2025 has converged on anomaly detection as the preferred paradigm — learning from normal data rather than requiring labeled defects. This is practically important because defects are rare and diverse. PatchCore-family methods with pretrained features (ImageNet, DINOv2) work across most materials. Supervised approaches remain necessary for severity classification and specific defect typing. The market includes both general-purpose vision platforms (Cognex, Keyence) and specialized AI inspection tools.
Key Challenges
Material diversity — defects look completely different on metal vs. fabric vs. semiconductor wafers
Texture variation — natural materials (wood, leather) have inherent variation that isn't defective
Lighting sensitivity — defect visibility depends heavily on illumination angle and type
Tiny defects — semiconductor and precision manufacturing defects can be <10 pixels in standard images
New product changeover — production lines switch materials frequently, requiring rapid adaptation
Quick Recommendations
General surface inspection
PatchCore / EfficientAD
Best unsupervised approaches — work across materials with only normal training data
Semiconductor wafer inspection
Custom CNN + domain-specific augmentation
Semiconductor defects require specialized high-resolution processing
Rapid new-product deployment
WinCLIP / DINOv2 features + few-shot
Deploy on new materials with 5-10 reference images
Multi-material platform
DINOv2 backbone + material-specific heads
Shared feature extraction with task-specific detection heads
What's Next
The frontier is universal surface inspection — a single model that handles any material with minimal setup. Foundation models pretrained on diverse manufacturing data will enable rapid deployment. Expect 3D surface scanning (structured light, photometric stereo) to complement 2D inspection, and self-supervised learning to reduce dependence on labeled defect data.
Benchmarks & SOTA
No datasets indexed for this task yet.
Contribute on GitHubRelated Tasks
Something wrong or missing?
Help keep Surface Defect Detection benchmarks accurate. Report outdated results, missing benchmarks, or errors.