Title: Multi-objective detection and instance segmentation of shield tunnel diseases based on deep learning
We utilize yolov5-7.0 to perform the object detection and instance segmentation tasks for shield tunnel diseases, for example, cracks, damages, and seepage.
Install
Clone repo and install requirements.txt in a Python>=3.7.0 environment, including PyTorch>=1.7.
git clone https://github.com/JiwenJ/Graduation-project.git # clone
cd yolov5
pip install -r requirements.txt # installDataset preparation
ddTrain from scratch
Train with transfer learning
Inference
Weights
According to our experiments, we conclude the object detection results in the following tables:
| 编号 | 加入注意力机制 | 修改特征金字塔 | 修改损失函数 | mAP@0.5 | mAP@0.5:0.95 | 权重下载 |
|---|---|---|---|---|---|---|
| 1 | ✗ | ✗ | ✗ | 72.5 | 43.1 | [Google Drive] |
| 2 | ✓ | ✗ | ✗ | 71.4 | 35.3 | [Google Drive] |
| 3 | ✗ | ✓ | ✗ | 74.6 | 40.6 | [Google Drive] |
| 4 | ✗ | ✗ | ✓ | 78.9 | 47.8 | [Google Drive] |
| 5 | ✓ | ✓ | ✓ | 80.1 | 48.0 | [Google Drive] |
1: What we refer to is YOLOv5s-7.0.
The best combination for object detection in our experiments is YOLOv5 with BiFPN, SIoU and AM. You can download the weight from [Google Drive].
- E-mail: 1546631808@qq.com
Dataset is credited to NJU.
Computation resource is credited to NJU and CASIA.
Computer vision tutorial is credited to online resources and RDD competition.