An U-net model is build for classification and segmentation of steel defects.
Exploratory data analysis shows that training data is highly un-balanced among defect classes and there is no defects on many sample images.
The model is trained with augmentation (on both images and masks).
A few computer vision techniques are used in post-processing to further improve the detection.
Dice coefficient of 0.79 is achieved.
A few ideas are proposed for future work.