CRL-SemCom-VidCI

This is the official implementation for 'Compression Ratio Learning and Semantic Communications for Video Imaging' paper

Update 22/07/2024 We have uploaded the complete training/testing codes for compression ratio learning.

Compression Ratio Learning

Test:

For testing learned-ratio methods, run experiment_scripts/project_test_x.py.
You can test different models by changing line 72 and 73 of project_test_x.py.

For example, when testing models under logs/24-03-08/24-03-08-MST/MST_adaptive/v_34/checkpoints/model_epoch_0006.pth, line 72/73 are,

file = [f for f in os.listdir(f'{dir_name}/v_{34}/checkpoints') if 'model_epoch_0006.pth' in f][0] 

fname = f'{dir_name}/v_{34}/checkpoints/{file}'

The average ratio for v_30, v_19, v_10, v_11, v_34 are 0.87, 1.18, 1.3, 1.44, 1.69 respectively.
For testing fixed-ratio methods, also run experiment_scripts/project_test_x.py. but pls make the following changes:

1. Change line 198 from parser.add_argument('--exp_name', type=str, default='MST_adaptive') to parser.add_argument('--exp_name', type=str, default='MST_fixed')
2. still change line 72/73 to test models under logs/24-03-08/24-03-08-MST/MST_fixed/v_1 or v_2 or v_3 or v_4
3. open line 366 in shutters/shutters_adaptive5_nomask.py and for testing v_1, v_2, v_3, v_4, action=1*, 2*, 3*, 4* torch.ones_like(action), respectively.

Besides, we made a mistake when plotting the figure for the fixed ratio method when the average ratio is 1 and mask B is not used. The PSNR should be 32.3 not 29.54. 29.54 is the performance when mask B is used.

Train:
For training learned-ratio methods, run experiment_scripts/train_x.py.
When training learned-ratio methods, we recover the video reconstruction network in fixed-ratio methods with action=4 by default.
To adjust the average sampling rate, adjust the parameter in line 177 of train_x.py. As a reference, when 0.05 is used, the average ratio is about 3.25; when 1 is used, the average ratio is about 1.82.

Additional information

If you find the source code is useful for your research, please cite our paper:
@ARTICLE{10539255, author={Zhang, Bowen and Qin, Zhijin and Li, Geoffrey Ye}, journal={IEEE Journal of Selected Topics in Signal Processing}, title={Compression Ratio Learning and Semantic Communications for Video Imaging}, year={2024}, volume={}, number={}, pages={1-13}, keywords={Image coding;Semantics;Sensors;Imaging;Time measurement;Optics;Image reconstruction}, doi={10.1109/JSTSP.2024.3405853}}