GBH-YOLOv5 implementation in Pytorch.

---

Contents

1. Project description
2. Required environment
3. Training
4. Testing
5. Acknowledgements

Project description

This project is based on the improved and optimized model of YOLOv5s, and its task is to detect defects on the surface of photovoltaic panels. In this study, the YOLOv5 model was improved to achieve 97.8% performance on PV Multi-Defect dataset.

Required environment

Pytorch == 1.8.1
python == 3.8
Cuda == 11.1

Install the project environment

pip install -r requirements.txt

Check whether the dependent environment is normal.

python detect.py --source ./data/images/ --weights weights/yolov5s.pt

Training

Data preparation

Download PV Multi-Defect images (train, val) and labels. Link: https://github.com/houhou34/PV-Multi-Defect-Datasets.
Download weights. Link: https://pan.baidu.com/s/1ApCScpr1CZ_AeVJH7crgCw Extraction code: lmn8

Start training

python train.py

Testing

The optimal weight obtained from the training was used for the test.

python detect.py --source ./testfiles/img1.jpg --weights runs/train/exp/weights/best.pt

Acknowledgements

https://github.com/ultralytics/yolov5
https://github.com/robintzeng/Pytorch-CSPNet