Implementation of Adaptive ReconNet in Tensorflow, which was proposed in paper Adaptive measurement network for CS image reconstruction.

The Adaptive ReconNet was originated from the ReconNet presented in ReconNet:Non-Iterative Reconstruction of Images from Compressively Sensed Random Measurements, the structure of which is shown below.

1. The structure of ReconNet

The main improvement that the Adaptive ReconNet makes is that it adds a fully-connected layer to replace the random Gaussian measurement part, so that the network should be able to get the measurement matrix all by itself during training. Its corresponding network is shown below.

2. The structure of Adaptive ReconNet

The results have been improved dramatically.

3. Improved results

I worte and tested the code under Linux environment. If you want to run it on Windows, you should modify the code to make it executable.

• Python 2.7.13
• Tensorlow 1.4.0
• h5py 2.7.1
• Scipy 1.0.0

• Source pic

  

• Reconstructed pic (MR=10%)

  

For yor information: The block effect is quite obvious in the comparion, which is because the non-symmetric padding that Tensorflow utilizes for computing convolution. A possible tip is, try to expand the convolution size and then cut the edge might eliminate the block effect.

• Run generate_test.m/ generate_train.m in folder generate_h5 to get yourself the test.h5/train.h5 dataset.
• Simply run the main.py, and you should get the sample results shown above.

• Modify the following part of code (line 71-72) inside funciton prepare_data defined in file utils.py to choose different test set.

# modify the paras if you want to change test pics
data_dir = os.path.join(os.sep, (os.path.join(os.getcwd(), dataset)), "Set5")
data = glob.glob(os.path.join(data_dir, "*.bmp"))

• Modify the following code (line 162) inside function input_setup defined in file utils.py to specify the number of the test pic.

input_init, label_init = preprocess(data[2], config.scale) # decide which pic to restore

• Run main.py to perform prediction on the chosen pic.

• Modify the flags part defined in file main.py as below.

  • Change the "stride" entry from 33 to 14 .
  • Change the "is_train" entry from False to True.
  • Modify the "epoch","batch_size" and "measurement_rate" if you need.

• Run the main.py to start training.

While I was writing this project, I took reference from the following repos:

1. ReconNet, the official implementation of ReconNet written in Matlab, working with Caffe.
2. SRCNN-Tensorflow, the implementation of SRCNN written in Python, working with Tensorflow.

I do appeciate the generous sharing of the open-source code.

This code was finished in August last year(2017). Due to the change to my reserch direction, Compressed Sensing or Reconstructing are not my study topics any more.So, this repo will no longer be under maintanence.

However, if you run into any issue, you're still welcomed to raise it, I will be pleased to answer that for you.