conformal-clip-data 0.1.0

Data bundle for conformal-clip examples and tests

Textile Image Dataset

This package provides the simulated textile images used in
Megahed et al., 2025 for few-shot CLIP classification experiments.

These images were originally generated using the R script below and were previously released under an MIT License in our repository:

• Image generation script:
  https://raw.githubusercontent.com/fmegahed/qe_genai/refs/heads/main/data/textile_images/extract_textile_images_from_r\_textile_pkg.R

• Original dataset release:
  https://github.com/fmegahed/qe_genai/tree/main/data/textile_images

Dataset Summary

To systematically evaluate CLIP's performance on STS image classification, we used the spc4sts R package to create a controlled dataset of simulated textile fabric textures. This approach allowed us to precisely model both nominal and defective weave structures and to control defect type and severity.

Our dataset contains:

Class Description Count

---

Nominal Standard textile weave patterns 1,000 Local defects Localized disruptions in the weave 500 Global defects Systematic shifts in weave parameters 500

Each image is 250 × 250 px, generated using spc4sts recommended parameters:

• Nominal images:
  Spatial autoregressive parameters ϕ₁ = 0.6, ϕ₂ = 0.35

• Global defects:
  Both parameters reduced by 5%

• Local defects:
  Generated using the package's defect-insertion functions

Citation

If you use this dataset, please cite:

Megahed et al., 2025
Adapting OpenAI's CLIP Model for Few-Shot Image Inspection in Manufacturing Quality Control: An Expository Case Study with Multiple Application Examples
arXiv:2501.12596

**Bui, A. T., & Apley, D. W. (2020)*. *spc4sts: Statistical process control for stochastic textured surfaces in R. Journal of Quality Technology, 53(3), 219–242.

License

These images were generated by the authors and are released under the MIT License.