MCFD-Net

Heping Song, Jingyao Gong, Hongying Meng, Yuping Lai.

Multi-Cross Sampling and Frequency-Division Reconstruction for Image Compressed Sensing.

Our paper can be found in: MCFD_Net(AAAI-2024)

Abstract

Existing CS methods use simple convolutional downsampling and only refer to low-level information in the reconstruction optimization in the information domain, making it difﬁcult to utilize high-level features of the original signal. In deep reconstruction,they ignore the different importance of distinguishing high- and low-frequency information. To address the above issues, we propose MCFD Net, which employs a series of clever methods through experiments show that MCFD-Net signiﬁcantly outperforms state of-the-art CS methods in multiple benchmark datasets while achieving better noise robustness.

Overview

Our poster in the conference of AAAI 2024:

Network Architecture

Requirements (recommend)

Python >= 3.9
Pytorch >= 1.13
Numpy >= 1.16

Datasets

Train data: Train Dataset from COCO 2014 dataset.
Test data: Test Dataset from Set5, Set11, BSDS200.
You should decompress and place these datasets in the "./dataset/" directory.

Pre training model

We provide pre-trained weights for convenient evaluation. It contains six sampling rates.

How To Run

Test

Ensure the test dataset path in data_processor.py is correct
Run the following scripts to test MCFD-Net model:

python val.py --sensing-rate <0.015625/0.3125/0.0625/0.125/0.25/0.5>

The default parameter configuration will evaluate the model with a perception rate of 0.5:

python val.py

Train

Ensure the train dataset path (train_40k) in data_processor.py is correct
Run the following scripts to train MCFD-Net model:

python train_mcfd.py --sensing-rate <0.015625/0.3125/0.0625/0.125/0.25/0.5> --epochs 100 --batch-size 4

The default parameters for training are as follows: learning rate = 0.5, batch size = 6, and number of epochs = 200:

python train_mcfd.py

Results

Quantitative Results

Retest on July 2, 2024, (method *) compared to the newly added method

Comparison of noise resistance performance

Contact

If you have any question, please email gongjy@stmail.ujs.edu.cn.

Citation

If you find the code helpful in your research or work, please cite the following paper:

@article{song2024multi,
  title={Multi-Cross Sampling and Frequency-Division Reconstruction for Image Compressed Sensing},
  author={Song, Heping and Gong, Jingyao and Meng, Hongying and Lai, Yuping},
  year={2024},
  publisher={Association for the Advancement of Artificial Intelligence}
}