Screw Detection & Classification

End-to-end industrial fastener inspection in a single notebook: screw_detection.ipynb.

Combines a classical OpenCV pipeline (CLAHE → Otsu → morphology → contour features + Hu moments → rule-based family classifier) with a deep-learning classifier (ConvNeXt-Tiny, ImageNet-pretrained, fine-tuned for 25 epochs with AdamW + cosine warm-up, AMP, RandAugment, RandomErasing, horizontal-flip TTA). Includes Grad-CAM explanations.

Runs identically on Google Colab (auto-mounts Drive and locates the project) and locally.

Repository layout

.
├── screw_detection.ipynb   # the entire pipeline
├── README.md               # this file
├── report.md               # methodology + results write-up
└── data/raw/
    ├── Ecrou NOK/    ├── Ecrou OK/
    ├── Rondelle NOK/ ├── Rondelle OK/
    └── Vis NOK/      └── Vis OK/

Running on Google Colab

1. Upload this repository to Google Drive as MyDrive/screw_project/ (the notebook also accepts MyDrive/screw_detection/ or any folder containing data/raw/).
2. Open screw_detection.ipynb in Colab.
3. Runtime → Change runtime type → T4 GPU (recommended for 25 epochs).
4. Run All. Cell 1 auto-detects Colab, mounts Drive, and locates the project root. Cell 2 installs any missing dependency (timm, opencv-python, etc.). Everything downstream is automatic.

Running locally

pip install opencv-python numpy pandas matplotlib seaborn scikit-image \
            scikit-learn timm tqdm Pillow torch torchvision jupyter
jupyter notebook screw_detection.ipynb

Then Run All.

Modes

Edit MODE at the top of the notebook:

Mode	Epochs	Use case
debug	1	smoke-test the pipeline
mini	5	fast verification on a laptop
medium	15	local GPU
full	25	best results on Colab GPU (default)

What the notebook produces

results/
├── metrics.csv                 # per-class precision / recall / F1
├── run_summary.json            # machine-readable run summary
├── checkpoints/
│   └── convnext_tiny_best.pth
├── classical/
│   └── features.csv            # contour + Hu features for sampled images
├── figures/
│   ├── class_distribution.png  ├── sample_grid.png
│   ├── classical_stages.png    ├── training_curves.png
│   ├── confusion_matrix.png    └── failure_cases.png
└── gradcam/
    └── gradcam_grid.png
data/splits/
├── train.csv  ├── val.csv  ├── test.csv  └── class_mapping.json
report.md      # a results section is appended after each run

Headline numbers

• Macro-F1 on the test split is the headline metric (accuracy alone is not reported as a primary number — class balance can be uneven).
• Classical family-classifier accuracy is reported alongside, as a transparent baseline for the family (Ecrou / Rondelle / Vis) decision.

Design decisions

Decision	Choice	Why
Backbone (6-class)	ConvNeXt-Tiny	best 6-class screw discriminator, ~28M params
Optimizer	AdamW + cosine warm-up	reliable, AMP-friendly
Mixed precision	torch.cuda.amp	~2× speed-up on T4
Augmentation	RandAugment(2, 9) + RandomErasing	strong but consistent
TTA	horizontal-flip averaging	small free boost at test time
Loss	cross-entropy, label-smoothing 0.05	calibration
Classical baseline	OpenCV-only, no GT masks	dataset has no pixel labels — kept as family baseline