Research map / Image generation / Autoregressive and latent generation
Automated Defect Report Generation for Enhanced Industrial Quality Control
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.
Introduces defect detection report generation (DDRG), builds Easy/Hard report datasets (NEU-r, MVTec-r) over 16 materials, and proposes a knowledge-aware report generation (KRG) baseline that uses ChatGPT-retrieved defect knowledge.
Problem
Industrial defect detection yields only bounding boxes and defect types, which give limited information for downstream processing; no dataset or method exists for generating textual defect reports including effects, causes, and solutions.
Approach
Frame defect detection output as text reports combining templated defect descriptions with human-written comprehensive information (effects, causes, solutions), and augment a VisualBERT-based generator with LLM-retrieved defect knowledge filtered by a knowledge-aware attention encoder feeding an LSTM decoder.
Key results
- On MVTec-r Hard, KRG reaches BLEU-4 19.43 vs. 14.02 for the best baseline (VisualBERT), a 38% relative gain.
- On NEU-r Hard, KRG reaches BLEU-4 16.88 vs. 9.82 for VisualBERT (about 53% relative improvement) and 70.75% accuracy vs. 65.54%.
- Human evaluation on comprehensive information (C) gives KRG 0.58 on NEU-r Hard and 0.49 on MVTec-r Hard, versus 0.31 and 0.30 for the best baselines, though well below human-written reports (0.87 and 0.83).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add knowledge retrieval and knowledge-aware encoding to the VisualBERT-based report generator | KRG · NEU-r Hard | BLEU-4: 9.82 -> 16.88 | Helped |
| Add knowledge retrieval and knowledge-aware encoding to the VisualBERT-based report generator | KRG · MVTec-r Hard | BLEU-4: 14.02 -> 19.43 | Helped |
| Use KRG for zero-shot report generation rather than VisualBERT | KRG · NEU-r Zero-shot | BLEU-4: 7.11 -> 9.42 | Helped |
| Use KRG for zero-shot report generation rather than VisualBERT | KRG · MVTec-r Zero-shot | BLEU-4: 9.98 -> 11.26 | Helped |
| Add knowledge retrieval and knowledge-aware encoding to the report generator | KRG · NEU-r Hard | Comprehensive Information: 0.31 -> 0.58 | Helped |
| Add knowledge retrieval and knowledge-aware encoding to the report generator | KRG · MVTec-r Hard | Comprehensive Information: 0.19 -> 0.49 | Helped |
Surprising findings
- Hard reports yield higher defect-type classification accuracy than Easy reports, suggesting comprehensive-information generation can improve defect description.
- All models perform poorly in the zero-shot setting (train on one dataset, test on the other), with KRG reaching only about 10% and 24% accuracy.
Limitations
- Easy reports are template-generated with roughly 50% duplication, limiting diversity.
- Comprehensive information is manually annotated and the zero-shot generalization across materials remains poor for all models.
- Generated comprehensive information remains far below human quality, and classification accuracy is moderate compared with supervised classifiers.
Experimental setup
Datasets: NEU MVTec
Benchmarks: NEU-r MVTec-r
Metrics: BLEU-1 BLEU-2 BLEU-3 BLEU-4 METEOR ROUGE-L CIDEr classification accuracy human fluency/description/comprehensive scores
Baselines: GRNN CLIP VisualBERT
Builds on: VisualBERT CLIP GRNN ConceptNet
Closest papers by meaning
- Towards Open-Vocabulary Industrial Defect Understanding with a Large-Scale Multimodal Dataset (CVPR 2026 · similarity 0.87)
- Text-Guided Variational Image Generation for Industrial Anomaly Detection and Segmentation (CVPR 2024 · similarity 0.85)
- MVREC: A General Few-shot Defect Classification Model Using Multi-View Region-Context (AAAI 2025 · similarity 0.84)
- DefectFill: Realistic Defect Generation with Inpainting Diffusion Model for Visual Inspection (CVPR 2025 · similarity 0.83)
- Wavelet and Prototype Augmented Query-based Transformer for Pixel-level Surface Defect Detection (CVPR 2025 · similarity 0.82)
- 3CAD: A Large-Scale Real-World 3C Product Dataset for Unsupervised Anomaly Detection (AAAI 2025 · similarity 0.82)
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.