SCFNet: Lightweight Steel Defect Detection Network Based on Spatial Channel Reorganization and Weighted Jump Fusion

⭐ This code has been completely released ⭐

⭐ our article ⭐

If our code is helpful to you, please cite:

@article{li2024scfnet,
  title={SCFNet: Lightweight Steel Defect Detection Network Based on Spatial Channel Reorganization and Weighted Jump Fusion},
  author={Li, Hongli and Yi, Zhiqi and Mei, Liye and Duan, Jia and Sun, Kaimin and Li, Mengcheng and Yang, Wei and Wang, Ying},
  journal={Processes},
  volume={12},
  number={5},
  pages={931},
  year={2024},
  publisher={MDPI}
}

## Requirements

```python
pip install -r requirements.txt

Train

1. Prepare training data

• The download link for the NEU-DET data set is here.
• The download link for the GC10-DET data set is here.

SCFNet
├── NEU-DET
│   ├── images
│   │   ├── train
│   │   │   ├── crazing_1.jpg
│   │   │   ├── crazing_2.jpg
│   │   │   ├── .....
│   │   ├── val
│   │   ├── test
│   ├── labels
│   │   ├── train
│   │   │   ├── crazing_1.txt
│   │   │   ├── crazing_2.txt
│   │   │   ├── .....
│   │   ├── val
│   │   ├── test

• After downloading the data set, modify the paths in path, train, val and test in the data.yaml file.

2. Begin to train

python train.py

Test

1. Begin to test

python val.py

Results

Methods	P	R	mAP50	mAP50:95	GFLOPs $\downarrow$	Params/M $\downarrow$
Faster R-CNN	0.615	0.865	0.76	0.377	135	41.75
CenterNet	0.712	0.749	0.764	0.412	123	32.12
YOLOv5n-7.0	0.694	0.694	0.746	0.422	4.2	1.77
YOLOv5s-7.0	0.745	0.719	0.761	0.429	15.8	7.03
YOLOv7-tiny	0.645	0.775	0.753	0.399	13.1	6.02
YOLOv8s	0.768	0.726	0.795	0.467	28.4	11.13
YOLOX-tiny	0.746	0.768	0.76	0.357	7.58	5.03
MRF-YOLO	0.761	0.707	0.768	-	29.7	14.9
YOLOv5s-FCC	-	-	0.795	-	-	13.35
WFRE-YOLOv8s	0.759	0.736	0.794	0.425	32.6	13.78
CG-Net	0.734	0.687	0.759	0.399	6.5	2.3
ACD-YOLO	-	-	0.793	-	21.3	-
YOLOv5-ESS	-	0.764	0.788	-	-	7.07
PMSA-DyTr	-	-	0.812	-	-	-
MED-YOLO	-	-	0.731	0.376	18	9.54
MAR-YOLO	-	-	0.785	-	20.1	-
SCFNet	0.786	0.715	0.812	0.469	5.9	2

• Bold indicates first or second best performance.

Time

2024.4.25 open the val.py

2024.5.16 update train.py

2024.5.16 update ScConv module.

Visualization of results

Acknowledgements

This code is built on ultralytics (PyTorch). We thank the authors for sharing the codes.

Contact

If you have any questions, please contact me by email ([email protected]).