A Lightweight Convolutional Network for Few-Shot and Multi-Class Detection of Tiny Aluminum Defects
The original paper can be found here.
This project addresses two key challenges in the field of industrial aluminum sheet surface defect detection:
- Limited Training Data: The difficulty in capturing small defect features with limited training data.
- Industrial Deployment Needs: The demand for high efficiency and lightweight structures in industrial applications.
To tackle the above challenges, we propose a solution based on attention mechanisms and lightweight architecture. Our approach involves designing an efficient encoder-decoder network that utilizes:
- Channel-wise convolutions and point convolutions to reduce computational costs.
- Spatial and channel attention modules embedded in skip connections between encoder and decoder modules to enhance the recognition of subtle defects.
We compared our model with traditional segmentation models, and the results are showcased below:
The experimental results indicate that our method achieves high precision (73.54% mIoU) on a small sample aluminum sheet defect dataset. It also demonstrates extremely fast inference speed (314.55 FPS) on a single V100 GPU, along with a small model size (0.114M parameters).
| Package | Version |
|---|---|
| torch | 2.0.1+cu118 |
| torchvision | 0.15.2+cu118 |
First, use the tools in the dataset_toolbox folder to prepare the dataset. Then run 'python train.py' to start training.
python train.py