Perturbed Progressive Learning for Semisupervised Defect Segmentation [TNNLS 2023]

This is the official PyTorch implementation of our work.

Perturbed Progressive Learning for Semisupervised Defect Segmentation
Yao Wu, Mingwei Xing, Yachao Zhang, Yuan Xie, Zongze Wu, Yanyun Qu
Published in: IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2023.

Overview

Getting Started

Installation

The implementation runs on

• python 3.8
• torch 1.9.0
• torchvision 0.10.0
• numpy
• pillow

Preparation

Dataset

• MSDD-3 (soon) | MTD | NEU-Seg | DAGM2007

Pre-trained Model

• resnet18 | resnet50

File Organization

├── [./dataset/splits]
    ├── MSDD-3
    ├── magnetic_tile
    ├── neu_seg
    └── DAGM
├── [./pretrained]
    ├── resnet18.pth
    └── resnet50.pth

Training and Testing

Example of MTD dateset:

CUDA_VISIBLE_DEVICES=0 python -W ignore train.py 
    --dataset magnetic_tile 
    --data-root /path/dataset/defect/magnetic_tile 
    --backbone resnet18 
    --model deeplabv3p 
    --labeled-id-path dataset/splits/magnetic_tile/1_8/labeled.txt 
    --unlabeled-id-path dataset/splits/magnetic_tile/1_8/unlabeled.txt 
    --pseudo-mask-path outdir/pseudo_masks/magnetic_tile 
    --save-path outdir/models/magnetic_tile 
    --use_PPL 
    --reliable-id-path outdir/reliable_ids/magnetic_tile

CUDA_VISIBLE_DEVICES=0 python -W ignore eval.py 
    --dataset magnetic_tile 
    --data-root /path/dataset/defect/magnetic_tile 
    --backbone resnet18 
    --model deeplabv3p 
    --best-model-path outdir/models/magnetic_tile/best_model.pth 
    --vis_path outdir/vis_result/magnetic_tile 
    --vis_save

Acknowledgements

Code is built based on ST++. Reproduced comparisons are based on CPS.

Citation

If you find this project useful, please consider citing:

@article{wu2023perturbed,
  title={Perturbed Progressive Learning for Semisupervised Defect Segmentation},
  author={Wu, Yao and Xing, Mingwei and Zhang, Yachao and Xie, Yuan and Wu, Zongze and Qu, Yanyun},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2023},
  publisher={IEEE}
}