SurfaceGuard: surface defect segmentation
Four defect types on 128x800 grayscale strips. Final option: U-Net ResNet-34 + FPN EfficientNet-B3, flip TTA, per-class post-processing tuned on a held-out validation fold. Validation score 0.7297 (threshold 0.5: 0.7121).
| option | raw_score | score | fg_dice | fp_score | empty_acc | params_M |
|---|---|---|---|---|---|---|
| Scratch U-Net + U-Net ResNet-34 + FPN EfficientNet-B3 | 0.7099 | 0.7306 | 0.6632 | 0.8861 | 0.8911 | 44.7598 |
| U-Net ResNet-34 + FPN EfficientNet-B3 | 0.7121 | 0.7297 | 0.6578 | 0.8966 | 0.8995 | 36.9086 |
| FPN EfficientNet-B3 | 0.6882 | 0.7280 | 0.6561 | 0.8947 | 0.8981 | 12.4760 |
| Scratch U-Net + FPN EfficientNet-B3 | 0.6965 | 0.7271 | 0.6444 | 0.9211 | 0.9173 | 20.3272 |
| Scratch U-Net + U-Net ResNet-34 | 0.6914 | 0.7108 | 0.6119 | 0.9467 | 0.9319 | 32.2838 |
| U-Net ResNet-34 | 0.6859 | 0.7084 | 0.6165 | 0.9261 | 0.9155 | 24.4326 |
| Scratch U-Net | 0.6259 | 0.6759 | 0.5976 | 0.8571 | 0.8620 | 7.8512 |
Files: weights/*.pt (state dict + config), pp_params.json, selection.json. Input: 1x1x128x800 in [0, 1].
IIT Madras BS Deep Learning & GenAI project (Atul Kumar, 24f2003045). Demo: https://huggingface.co/spaces/Sawarn/surfaceguard-demo