LCNet: Lightweight Cycle Network Driven by Physical and Deep Prior for Compressed Sensing [PyTorch]

This repository provides an implementation of the model proposed in the following paper

Datasets

For training, we use 400 images from the training set and test set of the BSDS500 dataset. The training images are cropped to 89600 96*96 pixel sub-images with data augmentation. For testing, we utilize three widely-used benchmark datasets, including Set11, BSDS68 and Urban100. Users can download the pre-processed training set from GoogleDrive. Training sets and test sets need to be placed under ./DataSets/.

Test

1. All models for our paper have been put in './save_temp'.
2. Run the following scripts to test LCNet model.

   # test scripts
   python test.py  --sensing-rate 0.1 --test_name Set11

3. You can change the sensing-rate and test_name to to get test results for different sensing rates in different datasets.

Train

1. Run the following scripts to train LCNet model.

   # train scripts
   python train.py --sensing_rate 0.1 --layer_num 12 --learning_rate 2e-4 --start_epoch 0 --epochs 100 --batch_size 32

2. You can change the sensing_rate to train models for different sensing rates.

Citation

If you find the code helpful in your research or work, please cite our papers.

@article{yang2025lcnet,
  title={LCNet: Lightweight Cycle Network Driven by Physical and Deep Prior for Compressed Sensing},
  author={Yang, Shuowen and P{\'e}rez-Bueno, Fernando and Qin, Hanlin and Molina, Rafael and Katsaggelos, Aggelos K},
  journal={IEEE Transactions on Multimedia},
  year={2025},
  publisher={IEEE}
}