This project is based on ultralytics/yolov3.

LF-YOLO (Lighter and Faster YOLO) is used to detect defect of X-ray weld image. The related paper is available here.

Download

$ git clone https://github.com/lmomoy/LF-YOLO

Train

We provide multiple versions of LF-YOLO with different widths.

$ python train.py --data coco.yaml --cfg LF-YOLO.yaml      --weights '' --batch-size 1
                                         LF-YOLO-1.25.yaml                           1
                                         LF-YOLO-0.75.yaml                           1
                                         LF-YOLO-0.5.yaml                            1

Results

We test LF-YOLO on our weld defect image dataset. Other methods are trained and tested based on MMDetection.

Model	size (pixels)	mAP50test	params (M)	FLOPS (B)
Cascasde-RCNN (ResNet50)	(1333, 800)	90.0	68.9	243.2
Cascasde-RCNN (ResNet101)	(1333, 800)	90.7	87.9	323.1
Faster-RCNN (ResNet50)	(1333, 800)	90.1	41.1	215.4
Faster-RCNN (ResNet101)	(1333, 800)	92.2	60.1	295.3
Dynamic-RCNN (ResNet50)	(1333, 800)	90.3	41.1	215.4
RetinaNet (ResNet50)	(1333, 800)	80.0	36.2	205.2
VFNet (ResNet50)	(1333, 800)	87.0	32.5	197.8
VFNet (ResNet101)	(1333, 800)	87.2	51.5	277.7
Reppoints (ResNet101)	(1333, 800)	82.7	36.6	199.0
SSD300 (VGGNet)	300	88.1	24.0	30.6
YOLOv3 (Darknet52)	416	91.0	62.0	33.1
SSD (MobileNet v2)	320	82.3	3.1	0.7
YOLOv3 (MobileNet v2)	320	90.2	3.7	1.6
LF-YOLO-0.5	320	90.7	1.8	1.1
LF-YOLO	320	92.9	7.3	4.0

We test our model on public dataset MS COCO, and it also achieves competitive results.

Model	size (pixels)	mAP50test	params (M)	FLOPS (B)
YOLOv3-tiny	640	34.8	8.8	13.2
YOLOv3	320	51.5	39.0	61.9
SSD	300	41.2	35.2	34.3
SSD	512	46.5	99.5	34.3
Faster R-CNN (VGG16)	shorter size: 800	43.9	-	278.0
R-FCN (ResNet50)	shorter size: 800	49.0	-	133.0
R-FCN (ResNet101)	shorter size: 800	52.9	-	206.0
LF-YOLO	640	47.8	7.4	17.1

Requirements

Python 3.8 or later with all requirements.txt dependencies installed, including torch>=1.7. To install run:

$ pip install -r requirements.txt

Inference

$ python detect.py --source data/images --weights LF-YOLO.pt --conf 0.25

Citation

Please consider citing my work as follows if it is helpful for you.

@article{liu2023lf,
  title={LF-YOLO: A lighter and faster yolo for weld defect detection of X-ray image},
  author={Liu, Moyun and Chen, Youping and Xie, Jingming and He, Lei and Zhang, Yang},
  journal={IEEE Sensors Journal},
  volume={23},
  number={7},
  pages={7430--7439},
  year={2023},
  publisher={IEEE}
}

@article{liu2021lf,
  title={LF-YOLO: A Lighter and Faster YOLO for Weld Defect Detection of X-ray Image},
  author={Liu, Moyun and Chen, Youping and He, Lei and Zhang, Yang and Xie, Jingming},
  journal={arXiv preprint arXiv:2110.15045},
  year={2021}
}