Mixed supervision for surface-defect detection: from weakly to fully supervised learning [Computers in Industry 2021]

Modification of the official PyTorch implementation for "Mixed supervision for surface-defect detection: from weakly to fully supervised learning" published in journal Computers in Industry 2021.

Code and the dataset are licensed under [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License][cc-by-nc-sa]. For comerical use please contact danijel.skocaj@fri.uni-lj.si.

How to run:

Requirements

Code has been tested to work on:

• Python 3.8
• PyTorch 1.6, 1.8
• CUDA 10.0, 10.1
• using additional packages as listed in requirements.txt

Datasets

You will need to download the datasets yourself. For DAGM and Severstal Steel Defect Dataset you will also need a Kaggle account.

• DAGM available here.
• KolektorSDD available here.
• KolektorSDD2 available here.
• Severstal Steel Defect Dataset available here.

For details about data structure refer to README.md in datasets folder.

Cross-validation splits, train/test splits and weakly/fully labeled splits for all datasets are located in splits directory of this repository, alongside the instructions on how to use them.

Using on other data

Refer to README.md in datasets for instructions on how to use the method on other datasets.

Demo - fully supervised learning

To run fully supervised learning and evaluation on all four datasets run:

./DEMO.sh
# or by specifying multiple GPU ids 
./DEMO.sh 0 1 2

Results will be written to ./results folder.

Replicating paper results

To replicate the results published in the paper run:

./EXPERIMENTS_COMIND.sh
# or by specifying multiple GPU ids 
./EXPERIMENTS_COMIND.sh 0 1 2

To replicate the results from ICPR 2020 paper:

@misc{Bozic2020ICPR,
    title={End-to-end training of a two-stage neural network for defect detection},
    author={Jakob Božič and Domen Tabernik and Danijel Skočaj},
    year={2020},
    eprint={2007.07676},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

run:

./EXPERIMENTS_ICPR.sh
# or by specifying multiple GPU ids 
./EXPERIMENTS_ICPR.sh 0 1 2

Results will be written to ./results-comind and ./results-icpr folders.

Usage of training/evaluation code

The following python files are used to train/evaluate the model:

• train_net.py Main entry for training and evaluation
• models.py Model file for network
• data/dataset_catalog.py Contains currently supported datasets

In order to train and evaluate a network you can also use EXPERIMENTS_ROOT.sh, which contains several functions that will make training and evaluation easier for you. For more details see the file EXPERIMENTS_ROOT.sh.

Running code

Simplest way to train and evaluate a network is to use EXPERIMENTS_ROOT.sh, you can see examples of use in EXPERIMENTS_ICPR.sh and in EXPERIMENTS_COMIND.sh

If you wish to do it the other way you can do it by running train_net.py and passing the parameters as keyword arguments. Bellow is an example of how to train a model for a single fold of KSDD dataset.

python -u train_net.py  \
    --GPU=0 \
    --DATASET=KSDD \
    --RUN_NAME=RUN_NAME \
    --DATASET_PATH=/path/to/dataset \
    --RESULTS_PATH=/path/to/save/results \
    --SAVE_IMAGES=True \
    --DILATE=7 \
    --EPOCHS=50 \
    --LEARNING_RATE=1.0 \
    --DELTA_CLS_LOSS=0.01 \
    --BATCH_SIZE=1 \
    --WEIGHTED_SEG_LOSS=True \
    --WEIGHTED_SEG_LOSS_P=2 \
    --WEIGHTED_SEG_LOSS_MAX=1 \
    --DYN_BALANCED_LOSS=True \
    --GRADIENT_ADJUSTMENT=True \
    --FREQUENCY_SAMPLING=True \
    --TRAIN_NUM=33 \
    --NUM_SEGMENTED=33 \
    --FOLD=0

Some of the datasets do not require you to specify --TRAIN_NUM or --FOLD- After training, each model is also evaluated.

For KSDD you need to combine the results of evaluation from all three folds, you can do this by using join_folds_results.py:

python -u join_folds_results.py \
    --RUN_NAME=SAMPLE_RUN \
    --RESULTS_PATH=/path/to/save/results \
    --DATASET=KSDD 

You can use read_results.py to generate a table of results f0r all runs for selected dataset.
Note: The model is sensitive to random initialization and data shuffles during the training and will lead to different performance with different runs unless --REPRODUCIBLE_RUN is set.