CACS-YOLO: A Lightweight Model for Insulator Defect Detection based on improved YOLOv8m

​ Fig. 1: Model architecture diagram of CACS-YOLO

Highlights

1) We proposed a new model CACS-YOLO by adding CAM based on feature reuse and feature CSO, which achieved good results in both detection accuracy and lightweight.

2) We proposed a new dataset by using the synthetic weather algorithms, which solved the problem of insufficient data and helped the model to further improve the detection accuracy.

3) The insulator dataset SFRID containing 20,567 insulator images, including foggy and rainy days, was constructed and released, so that the model can cope well with different weather conditions.

Getting started

1. Get codes

git clone https://github.com/CACS-YOLO.git cd CACS-YOLO

2. Installation of dependent libraries

pip install -r ./requirements.txt

3. Train model

To train the model, create a new python file and run the following code:

``` from ultralytics import YOLO

Load CACS-YOLO

model = YOLO(model = './ultralytics/cfg/models/CACS-YOLO.yaml')

Load YOLOv8m

model = YOLO(model = './ultralytics/cfg/models/YOLOv8m.yaml')

Load YOLOv8m-CAM

model = YOLO(model = './ultralytics/cfg/models/YOLOv8m-CAM.yaml')

Load YOLOv8m-CSO

model = YOLO(model = './ultralytics/cfg/models/YOLOv8m-CSO.yaml')

Training with UPID

model.train(data = './ultralytics/cfg/datasets/UPID.yaml', device = 0, epochs = 500)

Training with SFID

model.train(data = './ultralytics/cfg/datasets/SFID.yaml', device = 0, epochs = 500)

Training with SRID

model.train(data = './ultralytics/cfg/datasets/SRID.yaml', device = 0, epochs = 500)

Training with SFRID

model.train(data = './ultralytics/cfg/datasets/SFRID.yaml', device = 0, epochs = 600)

```

4. Test model

To test the model, create a new python file and run the following code:

``` from ultralytics import YOLO

Load the best model

model = YOLO('./runs/detect/train/weights/best.pt')

Testing with UPID

model.val(data = './ultralytics/cfg/datasets/UPID.yaml', device = 0)

Testing with SFID

model.val(data = './ultralytics/cfg/datasets/SFID.yaml', device = 0)

Testing with SRID

model.val(data = './ultralytics/cfg/datasets/SRID.yaml', device = 0)

Testing with SFRID

model.val(data = './ultralytics/cfg/datasets/SFRID.yaml', device = 0)

```

5. Inference

To infer the model, create a new python file and run the following code:

``` from ultralytics import YOLO

Load the best model

model = YOLO("./runs/detect/train/weights/best.pt", task = "detect")

Infer the model

result = model(source = "./data_example/000050.jpg", save = True) ```

6. Synthetic fog generation

You can use the imagesyntheticfog.py to generate foggy images.

``` python ./imagesyntheticfog.py

Example

python imagesyntheticfog.py ./dataexample/000050.jpg 0.9 0.05 ./dataexample/synthesiseffect/foggy000050.jpg

```

7. Synthetic rain generation

You can use the imagesyntheticrain.py to generate rainy images.

``` python ./ImageSyntheticRain.py

Example

python imagesyntheticrain.py ./dataexample/000050.jpg -30 600 5 60 ./dataexample/synthesiseffect/rainy000050.jpg

```

8. Citing CACS-YOLO

If you use CACS-YOLO in your research, please use the following BibTeX entry. 📣 Thank you!

@article{cao2024cacs, title={CACS-YOLO: A Lightweight Model for Insulator Defect Detection based on Improved YOLOv8m}, author={Cao, Zhong and Chen, Kaihong and Chen, Junzuo and Chen, Zhaohui and Zhang, Man}, journal={IEEE Transactions on Instrumentation and Measurement}, year={2024}, publisher={IEEE} }