ISTA-Net: Interpretable Optimization-Inspired Deep Network for Image Compressive Sensing [PyTorch version]
This repository is for ISTA-Net and ISTA-Net+ introduced in the following paper
Jian Zhang, Bernard Ghanem , "ISTA-Net: Interpretable Optimization-Inspired Deep Network for Image Compressive Sensing", CVPR 2018, [pdf] [Supp]
The code is built on PyTorch and tested on Ubuntu 16.04/18.04 and Windows 10 environment (Python3.x, PyTorch>=0.4) with 1080Ti GPU.
With the aim of developing a fast yet accurate algorithm for compressive sensing (CS) reconstruction of natural images, we combine in this paper the merits of two existing categories of CS methods: the structure insights of traditional optimization-based methods and the speed of recent network-based ones. Specifically, we propose a novel structured deep network, dubbed ISTA-Net, which is inspired by the Iterative Shrinkage-Thresholding Algorithm (ISTA) for optimizing a general L1 norm CS reconstruction model. To cast ISTA into deep network form, we develop an effective strategy to solve the proximal mapping associated with the sparsity-inducing regularizer using nonlinear transforms. All the parameters in ISTA-Net (\eg nonlinear transforms, shrinkage thresholds, step sizes, etc.) are learned end-to-end, rather than being hand-crafted. Moreover, considering that the residuals of natural images are more compressible, an enhanced version of ISTA-Net in the residual domain, dubbed ISTA-Net+, is derived to further improve CS reconstruction. Extensive CS experiments demonstrate that the proposed ISTA-Nets outperform existing state-of-the-art optimization-based and network-based CS methods by large margins, while maintaining fast computational speed.
Figure 1. Illustration of the proposed ISTA-Net framework.
-
All models for our paper have been put in './model'.
-
Run the following scripts to test ISTA-Net models.
You can use scripts in file 'TEST_ISTA_Net_scripts.sh' to produce results for our paper.
# test scripts python TEST_CS_ISTA_Net.py --epoch_num 200 --cs_ratio 1 --layer_num 9 python TEST_CS_ISTA_Net.py --epoch_num 200 --cs_ratio 4 --layer_num 9 python TEST_CS_ISTA_Net.py --epoch_num 200 --cs_ratio 10 --layer_num 9 python TEST_CS_ISTA_Net.py --epoch_num 200 --cs_ratio 25 --layer_num 9 python TEST_CS_ISTA_Net.py --epoch_num 200 --cs_ratio 30 --layer_num 9 python TEST_CS_ISTA_Net.py --epoch_num 200 --cs_ratio 40 --layer_num 9 python TEST_CS_ISTA_Net.py --epoch_num 200 --cs_ratio 50 --layer_num 9
-
Run the following scripts to test ISTA-Net+ models.
You can use scripts in file 'TEST_ISTA_Net_plus_scripts.sh' to produce results for our paper.
# test scripts python TEST_CS_ISTA_Net_plus.py --epoch_num 200 --cs_ratio 1 --layer_num 9 python TEST_CS_ISTA_Net_plus.py --epoch_num 200 --cs_ratio 4 --layer_num 9 python TEST_CS_ISTA_Net_plus.py --epoch_num 200 --cs_ratio 10 --layer_num 9 python TEST_CS_ISTA_Net_plus.py --epoch_num 200 --cs_ratio 25 --layer_num 9 python TEST_CS_ISTA_Net_plus.py --epoch_num 200 --cs_ratio 30 --layer_num 9 python TEST_CS_ISTA_Net_plus.py --epoch_num 200 --cs_ratio 40 --layer_num 9 python TEST_CS_ISTA_Net_plus.py --epoch_num 200 --cs_ratio 50 --layer_num 9
-
Prepare test data.
The original test set11 is in './data'
-
Run the test scripts.
See Quick start
-
Check the results in './result'.
-
Trainding data (Training_Data.mat including 88912 image blocks) is in './data'. If not, please download it from GoogleDrive or BaiduPan [code: xy52].
-
Place Training_Data.mat in './data' directory
-
run the following scripts to train ISTA-Net models.
You can use scripts in file 'Train_ISTA_Net_scripts.sh' to train models for our paper.
# CS ratio 1, 4, 10, 25, 30, 40, 50 # train scripts python Train_CS_ISTA_Net.py --cs_ratio 10 --layer_num 9 python Train_CS_ISTA_Net.py --cs_ratio 25 --layer_num 9 python Train_CS_ISTA_Net.py --cs_ratio 50 --layer_num 9 python Train_CS_ISTA_Net.py --cs_ratio 1 --layer_num 9 python Train_CS_ISTA_Net.py --cs_ratio 4 --layer_num 9 python Train_CS_ISTA_Net.py --cs_ratio 30 --layer_num 9 python Train_CS_ISTA_Net.py --cs_ratio 40 --layer_num 9
We found that the re-trained ISTA-Net models may get a bit higher performance than the results reported in our paper.
-
run the following scripts to train ISTA-Net+ models.
You can use scripts in file 'Train_ISTA_Net_plus_scripts.sh' to train models for our paper.
# CS ratio 1, 4, 10, 25, 30, 40, 50 # train scripts python Train_CS_ISTA_Net_plus.py --cs_ratio 10 --layer_num 9 python Train_CS_ISTA_Net_plus.py --cs_ratio 25 --layer_num 9 python Train_CS_ISTA_Net_plus.py --cs_ratio 50 --layer_num 9 python Train_CS_ISTA_Net_plus.py --cs_ratio 1 --layer_num 9 python Train_CS_ISTA_Net_plus.py --cs_ratio 4 --layer_num 9 python Train_CS_ISTA_Net_plus.py --cs_ratio 30 --layer_num 9 python Train_CS_ISTA_Net_plus.py --cs_ratio 40 --layer_num 9
-
All models for our paper have been put in './model'.
-
Run the following scripts to test ISTA-Net+ models.
# test scripts python TEST_MRI_CS_ISTA_Net_plus.py --epoch_num 200 --cs_ratio 20 --layer_num 9 python TEST_MRI_CS_ISTA_Net_plus.py --epoch_num 200 --cs_ratio 30 --layer_num 9 python TEST_MRI_CS_ISTA_Net_plus.py --epoch_num 200 --cs_ratio 40 --layer_num 9 python TEST_MRI_CS_ISTA_Net_plus.py --epoch_num 200 --cs_ratio 50 --layer_num 9
-
Prepare test data.
The original test BrainImages_test is in './data'
-
Run the test scripts.
See Quick start
-
Check the results in './result'.
-
Trainding data (Training_BrainImages_256x256_100.mat including 88912 image blocks) is in './data'. If not, please download it from GoogleDrive.
-
Place Training_BrainImages_256x256_100.mat in './data' directory
-
run the following scripts to train ISTA-Net+ models.
You can use scripts in file 'Train_ISTA_Net_plus_scripts.sh' to train models for our paper.
# train scripts python Train_MRI_CS_ISTA_Net_plus.py --cs_ratio 20 --layer_num 9 python Train_MRI_CS_ISTA_Net_plus.py --cs_ratio 30 --layer_num 9 python Train_MRI_CS_ISTA_Net_plus.py --cs_ratio 40 --layer_num 9 python Train_MRI_CS_ISTA_Net_plus.py --cs_ratio 50 --layer_num 9
If you find the code helpful in your resarch or work, please cite the following papers.
@inproceedings{zhang2018ista,
title={ISTA-Net: Interpretable optimization-inspired deep network for image compressive sensing},
author={Zhang, Jian and Ghanem, Bernard},
booktitle={CVPR},
pages={1828--1837},
year={2018}
}