Case study
InkjetOOD: Transferring Generative OOD Detection to Industrial Print Quality Control
A public YOLO + conditional diffusion pipeline for industrial print-quality control evaluated on FTI_Zer0P.
Problem
Industrial quality control in inkjet printing on composite building components deals with micro-scale defects, severe class imbalance, and strict latency requirements, requiring rigorous cross-validation.
Role
Machine vision researcher: dataset annotation in Roboflow, YOLOv8 feature integration, conditional diffusion modeling, and 5-fold cross-validation benchmarking.
Architecture & Pipeline Summary
YOLOv8PyTorchRoboflowOpenCVHydraFTI_Zer0P
1. Industrial Surface Imaging
High-resolution thermographic and optical frame capture of composite building components.
2. YOLOv8 Feature Extraction
Localizing regions of interest and extracting structured defect feature representations.
3. 5-Fold Cross-Validation
Strict 5-fold CV protocol on FTI_Zer0P establishing an auditable 0.8673 ± 0.0230 baseline.
Approach
- Extract structured visual features from industrial prints using YOLOv8 backbone representations.
- Condition a diffusion model on print metadata and feature embeddings for quality verification.
- Run strict 5-fold cross-validation on FTI_Zer0P and compare separation loss settings under statistical correction.
Results
- 0.8673 ± 0.0230 AUROC baseline achieved on the public FTI_Zer0P benchmark under strict 5-fold cross-validation.
- Separation loss showed null effect on small industrial textures, confirming that natural-image gains do not automatically transfer to manufacturing surfaces.
- Clarified domain boundary conditions and documented negative results openly to save future engineering and research cycles.
Lessons
- Industrial transfer requires re-evaluating core assumptions, not just reapplying code.
- Feature heterogeneity and texture scale profoundly change generative classifier dynamics.
- A rigorously documented null result is a valuable scientific contribution that prevents wasted engineering time.