Add industrial surface defect detection example using NEU dataset

[farquardsolve]

Hi team,

I’m submitting this PR to contribute an end-to-end working example for industrial surface defect detection using MMDetection and the NEU Surface Defect Database.

I work with industrial manufacturing data, and I noticed there isn’t an out-of-the-box config showing how to adapt MMDetection for tasks like inspecting surface flaws on metal sheets or other products. So I’ve put together a simple pipeline that others can pick up quickly.

---

This PR adds:

• A new config under configs/defect/ for training Faster R-CNN on the NEU dataset (6 defect types).
• A convert_annotations.py tool for converting NEU images into COCO format (train/val/test).
• An optional app.py with a Streamlit example to test inference interactively.
• Minor tweaks to base configs where needed.

---

Why this is useful:

Factories and students working on quality inspection often ask how to adapt COCO configs to their own datasets. This example shows how to structure a small industrial dataset in COCO format and train a detector without needing custom code from scratch.

My goal is to help bridge the gap between MMDetection’s research focus and day-to-day practical use cases in manufacturing and visual QA.

---

Tested:

• Trained on a local NEU split.
• Ran inference using the new config and image_demo.py.

---

**Happy to adjust this if you’d like it moved into projects/ or somewhere else — open to any suggestions.

Thanks for maintaining this great framework.**

[CLAassistant]

All committers have signed the CLA.