A deep learning project that automatically detects and classifies surface defects in industrial and aerospace components using a hybrid CNN + attention + transfer learning approach.
Accuracy: 93.61% on the NEU Surface Defect Dataset Hardware: Runs on standard CPU — no GPU needed
Hybrid Deep Learning Framework for Automated Industrial Surface Defect Detection and Classification with Potential Aerospace Applications
Jewel Christanand Pisse, Shekhar R. Alliance University, Bengaluru, India — 2025
| Metric | Value |
|---|---|
| Overall Accuracy | 93.61% |
| Macro Precision | 0.95 |
| Macro Recall | 0.94 |
| Macro F1-Score | 0.94 |
NEU Surface Defect Dataset — 1800 images, 6 defect classes: Crazing, Inclusion, Patches, Pitted Surface, Rolled-in Scale, Scratches
Download: http://faculty.neu.edu.cn/yunhyan/NEU_surface_defect_database.html
Step 1 — Install dependencies
pip install torch torchvision scikit-learn matplotlib
Step 2 — Open the notebook
Open HDLF_runnable.ipynb in Jupyter Notebook
Step 3 — Update your dataset paths
In Cell 2, change these two lines to match where you saved the dataset on your computer:
train_path = "your/path/to/NEU-DET/train/images"
val_path = "your/path/to/NEU-DET/validation/images"
Step 4 — Run all cells top to bottom
| File | Description |
|---|---|
HDLF_runnable.ipynb |
Main project notebook — run this |
Final.pdf |
Conference paper |
README.md |
This file |
Jewel Christanand Pisse PG Student — Alliance University, Bengaluru cjewelMTECH25@stu.alliance.edu.in
Supervisor: Prof. Shekhar R., Alliance University