This project was to improve algorithms for localizing and classifying surface defects on a steel sheet. Models were built with sample base CNN to dense NET to VGG for classification, and with different UNET combined with data augmentation to further boost the performance for segmentation. Algorithms are tested on Kaggle.