cpwan/steel-defect-detection

Course project for steel defect detection (https://www.kaggle.com/c/severstal-steel-defect-detection/)

Steel defect detection with U-net and CAN

Report

Our report can be found here

References

U-net: Convolutional networks for biomedical image segmentation

Dilated Residual Networks

Code structure modified on top of https://github.com/fyu/drn

Proprocess

1. Download the data from kaggle and put it in the ./data directory, unzip the train_images.zip to ./data/img

2. Run in bash:

cd data
python ./preprocess.py
./create_list.sh
python ./produce_info_json.py

This would create a directory ./data/mask. The labels are stored in this directory.

Also, there will be 4 text files in ./data, each containing the path to images/labels.

The produce_info_json.py writes the mean and standard deviation of the images to info.json. This is already provided.

Training/ eavaluating

Refer to script.md