Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
MVREC: A General Few-shot Defect Classification Model Using Multi-View Region-Context
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MVREC, a few-shot defect multi-classification framework that combines AlphaCLIP region-context features (mask prompts plus multi-view patch augmentation) with training-free and fine-tuned Zip-Adapter classifiers, and introduces the MVTec-FS benchmark.
Problem
Few-shot defect multi-classification (FSDMC) lacks generalizable methods and multi-category benchmarks, and standard cropping-and-resizing discards the surrounding context that defect classification needs.
Approach
Use AlphaCLIP with defect masks as region-context input, augment each defect with multi-scale and offset patches to form multi-view features, and classify with a cache-based Zip-Adapter (training-free) or Zip-Adapter-F (fine-tuned with cross-entropy plus triplet loss).
Key results
- Zip-Adapter-F (MVREC) reaches 89.4% average accuracy on MVTec-FS in the 5-shot setting, versus 82.5% for the best baseline (CLIP-LinearProb).
- Zip-Adapter (MVREC) achieves 71.6% in 1-shot on MVTec-FS, versus 61.5% for CLIP-Adapter.
- Adding MVREC features improves CLIP-Adapter by +11.6 points at 1-shot, and by +8.2 points at 5-shot.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MVREC features instead of the unaugmented representation (ablation) | few-shot classifiers on MVTec-FS · MVTec-FS, 1-shot | accuracy: +11.3 | Helped |
| Use fixed-size crops with masked AlphaCLIP region context rather than removing the mask context (ablation) | Zip-Adapter on MVTec-FS · MVTec-FS, 1-shot | accuracy: 61.8 -> 71.6 | Helped |
| Use fixed-size crops with masked AlphaCLIP region context rather than removing the mask context (ablation) | Zip-Adapter-F on MVTec-FS · MVTec-FS, 1-shot | accuracy: 66.3 -> 73.7 | Helped |
| Add scale augmentation alone (ablation) | Zip-Adapter on MVTec-FS · MVTec-FS, 1-shot | accuracy: 60.6 -> 68.1 | Helped |
| Add rotation augmentation alone (ablation) | Zip-Adapter on MVTec-FS · MVTec-FS, 1-shot | accuracy: 60.6 -> 68.5 | Helped |
| Add flip augmentation alone (ablation) | Zip-Adapter on MVTec-FS · MVTec-FS, 1-shot | accuracy: 60.6 -> 55.3 | Hurt |
| Add offset augmentation alone (ablation) | Zip-Adapter on MVTec-FS · MVTec-FS, 1-shot | accuracy: 60.6 -> 70.7 | Helped |
| Combine scale and offset augmentation (ablation) | Zip-Adapter on MVTec-FS · MVTec-FS, 1-shot | accuracy: 60.6 -> 71.6 | Helped |
| Train both the support features and ZIP module (ablation) | Zip-Adapter-F on MVTec-FS · MVTec-FS, 1-shot | accuracy: 72.67 -> 73.74 | Helped |
| Use Zip-Adapter-F (MVREC) rather than the strongest listed baseline, CLIP-LinearProb | MVTec-FS · MVTec-FS, 5-shot | accuracy: 82.5 -> 89.4 | Helped |
Surprising findings
- Vanilla CLIP with defect-size cropping gives the worst results, below fixed-size cropping, underscoring the value of context.
- Flip augmentation hurts accuracy substantially (e.g. 55.3% vs 60.6% baseline at 1-shot for Zip-Adapter), contrary to typical augmentation intuitions.
Limitations
- Does not explore a unified model trained once across different defect datasets.
- Uses only CLIP image features and does not exploit the text encoder for multimodal research.
Experimental setup
Datasets: MVTec-FS NEU-DET PCB Defect Dataset MTD AITEX
Benchmarks: MVTec-FS NEU-DET PCB Defect Dataset MTD AITEX
Metrics: classification accuracy
Baselines: CLIP-ZeroShot CLIP-Adapter CLIP-ProtoNet CLIP-KNN CLIP-LinearProb Tip-Adapter Tip-Adapter-F
Builds on: AlphaCLIP Tip-Adapter CLIP-Adapter Prototypical Networks
Closest papers by meaning
- Towards Open-Vocabulary Industrial Defect Understanding with a Large-Scale Multimodal Dataset (CVPR 2026 · similarity 0.87)
- MuSc: Zero-Shot Industrial Anomaly Classification and Segmentation with Mutual Scoring of the Unlabeled Images (ICLR 2024 · similarity 0.86)
- Zero-Shot Learning in Industrial Scenarios: New Large-Scale Benchmark, Challenges and Baseline (AAAI 2025 · similarity 0.85)
- Wavelet and Prototype Augmented Query-based Transformer for Pixel-level Surface Defect Detection (CVPR 2025 · similarity 0.85)
- Automated Defect Report Generation for Enhanced Industrial Quality Control (AAAI 2024 · similarity 0.84)
- 3CAD: A Large-Scale Real-World 3C Product Dataset for Unsupervised Anomaly Detection (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.