SurfaceGuard models
Industrial surface anomaly segmentation (4 defect types, 128x800 images).
Models: U-Net (ResNet34, and optionally ResNet50) with a classification head, ImageNet-pretrained, input up-scaled x2. Final prediction = average of the models + flip test-time augmentation; per-defect pixel threshold, minimum area and classification gate are stored in ensemble_params.json.
W&B project: https://wandb.ai/axiomloops-indian-institute-of-technology-madras/24f3002363-t32026
Validation results (all runs)
run score dice f1_macro pixel_acc
5-unet-resnet34-bce 0.662196 0.524287 0.543751 0.995615
v2-ensemble-tuned 0.648289 0.517237 0.503822 NaN
4b-unet-resnet34-bce_dice-8ep 0.637148 0.512918 0.390207 0.992535
4b-unet-resnet34-bce_dice-8ep 0.635925 0.508267 0.426215 0.993109
4b-unet-resnet34-bce_dice-8ep 0.633143 0.515953 0.364868 0.992597
v2-B-unet-resnet50-up2 0.632614 0.497271 0.446302 0.992958
5-unet-resnet34-bce-8ep 0.626108 0.471110 0.414089 0.995324
5-unet-resnet34-bce-8ep 0.623543 0.467652 0.411681 0.995115
5-unet-resnet34-bce-8ep 0.623243 0.467995 0.405734 0.994973
5-unet-resnet34-bce-8ep 0.620405 0.463052 0.406536 0.995233
v2-A-unet-resnet34-up2 0.616166 0.478025 0.418710 0.992578
4b-unet-resnet34-bce_dice-8ep 0.611900 0.477212 0.387218 0.992032
2-unet-resnet34-pretrained-bce_dice 0.606762 0.465905 0.385044 0.990593
4-unet-resnet34-gray-bce_dice 0.598955 0.460546 0.370157 0.991895
2-unet-resnet34-pretrained-bce_dice 0.597914 0.457613 0.358274 0.991153
2-unet-resnet34-pretrained-bce_dice 0.597914 0.457613 0.358274 0.991153
2-unet-resnet34-pretrained-bce_dice 0.597914 0.457613 0.358274 0.991153
2-unet-resnet34-pretrained-bce_dice 0.597914 0.457613 0.358274 0.991153
2-unet-resnet34-pretrained-bce_dice 0.597235 0.456631 0.368650 0.992312
4-unet-resnet34-gray-bce_dice 0.596804 0.455599 0.380052 0.991866
2-unet-resnet34-pretrained-bce_dice 0.591355 0.450079 0.368339 0.988875
5-unet-resnet34-bce 0.578050 0.403332 0.385418 0.994412
2-unet-resnet34-pretrained-bce_dice 0.577385 0.423682 0.289705 0.988298
4-unet-resnet34-gray-bce_dice 0.575830 0.413089 0.351766 0.991443
4-unet-resnet34-gray-bce_dice 0.572914 0.406271 0.340572 0.991654
4-unet-resnet34-gray-bce_dice 0.567381 0.403728 0.310122 0.992001
2-unet-resnet34-pretrained-bce_dice 0.559764 0.387148 0.321792 0.993135
1-scratch-unet-bce_dice 0.549552 0.382638 0.293855 0.989955
1-scratch-unet-bce_dice 0.549552 0.382638 0.293855 0.989955
1-scratch-unet-bce_dice 0.549552 0.382638 0.293855 0.989955
1-scratch-unet-bce_dice 0.549552 0.382638 0.293855 0.989955
1-scratch-unet-bce_dice 0.549552 0.382638 0.293855 0.989955
1-scratch-unet-bce_dice 0.549552 0.382638 0.293855 0.989955
1-scratch-unet-bce_dice 0.549552 0.382638 0.293855 0.989955
1-scratch-unet-bce_dice 0.549552 0.382638 0.293855 0.989955
1-scratch-unet-bce_dice 0.544902 0.379547 0.278391 0.990264
1-scratch-unet-bce_dice 0.541374 0.376980 0.265756 0.988413
1-scratch-unet-bce_dice 0.537879 0.369913 0.286081 0.990016
1-scratch-unet-bce_dice 0.537879 0.369913 0.286081 0.990016
1-scratch-unet-bce_dice 0.536731 0.366589 0.272715 0.990901
1-scratch-unet-bce_dice 0.530941 0.373951 0.235222 0.984088
1-scratch-unet-bce_dice 0.522173 0.339898 0.281727 0.992067
1-scratch-unet-bce_dice 0.519787 0.339004 0.286286 0.990402
6-unet-resnet34-dice-8ep 0.519504 0.444031 0.303561 0.906634
6-unet-resnet34-dice-8ep 0.513731 0.444995 0.308442 0.901961
1-scratch-unet-bce_dice 0.506446 0.325270 0.263331 0.989588
6-unet-resnet34-dice 0.505180 0.436173 0.293895 0.883640
3-fpn-resnet34-pretrained-bce_dice 0.501997 0.290444 0.342990 0.994151
3-fpn-resnet34-pretrained-bce_dice 0.493750 0.285252 0.350334 0.991522
3-fpn-resnet34-pretrained-bce_dice 0.485390 0.271712 0.343232 0.991897
6-unet-resnet34-dice-8ep 0.482450 0.409540 0.300612 0.874386
3-fpn-resnet34-pretrained-bce_dice 0.480894 0.263623 0.337458 0.992590
3-fpn-resnet34-pretrained-bce_dice 0.462410 0.239116 0.287899 0.991534
3-fpn-resnet34-pretrained-bce_dice 0.461029 0.231932 0.300279 0.993647
6-unet-resnet34-dice-8ep 0.444116 0.365618 0.252267 0.872620