SSNet2: a novel communication wideband signal detection network with start-stop points, which can be trained without special candidate anchors.
This project implemented SSNet2 for Wideband signal detection and tested on simulated datasets and actual received wideband signals.
Relevant training data sets and test data sets will be released in the near future.
(./figures/center_heatmap.jpeg)
*Predicted heatmap of object center points on an time frequency spectrum of broadband signal
*Our trained model is placed on ./snapshot_best folder
├── SSNet
│ ├── dataset.py
│ ├── DLAnet.py
│ ├── loss.py
│ ├── predict.py
│ ├── train.py
│ └── utils.py
├── dataset_split
│ ├── train.txt
│ └── val.txt
├── environment.yml
This repository was developed and tested in PyTorch 1.5.
- Intall required dependencies as listed in environment.yml
- Modify signal dataset directory in centernet-vanilla/dataset.py
- Run train.py for training and predict.py for inference
(./figures/predict_results.pdf)
Compare evaluation results of our implementation to the original CenterNet on all datasets.
(./figures/train_loss.png)
An example image : The network loss function with epochs.
(./figures/network.pdf)
The pipeline of our method for wideband signal detection and classification.
(./figures/pr_curve.png)
*Wideband signal detection Recall vs. Precision curve at GFSK signal
We used the DLA-34 network, loss functions and some other functions from this R-CenterNet repo.