Surface defect segmentation (DL GenAI T3 2026)
Predicts masks for 4 defect types on 128 x 800 grayscale images.
Three models are saved here: model1_tiny_unet (from scratch), model2_unet_resnet34 and model3_fpn_convnext_small (both pretrained encoders). Each model has a .pth state dict and a config json with the tuned threshold and min area per class.
Best model: model3_fpn_convnext_small
Validation results (score = 0.7 * dice on defects + 0.3 * empty accuracy):
model val_score val_f1 val_acc val_dice val_empty_acc
model3_fpn_convnext_small 0.762043 0.792306 0.937683 0.682348 0.948000 model2_unet_resnet34 0.725394 0.740050 0.916179 0.640575 0.923304 model1_tiny_unet 0.640919 0.597876 0.844353 0.551499 0.849565