This is the code for Deep Learning part of the paper.

mmdetection

The deep learning part is based on the mmdetection framework, we recommend read the mmdetection documentation for further detail: https://mmdetection.readthedocs.io/en/latest/2_new_data_model.html.

All model configurations and training/testing pipeline configurations are defined in mmdetection config file. config file is saved in folder configs/. To learn more about mmdetection config, please read https://mmdetection.readthedocs.io/en/latest/tutorials/config.html.

1. Put all training images in folder dataset/images/, put all annotation files in folder dataset/annotations/. The annotations are COCO format.
2. Config training setting through mmdetection config file in folder configs/.
3. Run train.py to start training, the trainer could be configured in the script. Config file could be selected by the statement config=....
4. In order to do 2 Stage training, you can firstly do the first stage training, then save the pretrained model in checkpoints/, then in the second stage, load the pretrained model by load_from=... in config.
5. A scaled dataset can be created by hacking the annotation file by multiplying the image and annotation object. e.g. real_defect_train_scaled.json

1. Put all testing images in folder dataset/images/, put all annotation files in folder dataset/annotations/.
2. Config testing setting through mmdetection config file in folder configs/.
3. Run test.py to start testing, the tester could be configured in the script. Config file could be selected by the statement args.config=... and the checkpoint could be selected by the statement args.checkpoint=....