Code of the paper "Learning a Patch Quality Comparator for Single Image Dehazing"

Project Page

Input	Dehzed

Dependency

• For Running
  • Python 2
  • keras (with any backend)
  • scikit-image
  • scipy
  • scikit-sparse
  • numpy

Both scipy and scikit-sparse are used for sparse matrix computation. The scikit-sparse has been used for solving linear equation with sparse matrix. Although the same thing can be achieved using scikit only, scikit-sparse is faster as it uses Cholesky decomposition using CHOLMOD library.

• Additional dependency for training
  • opencv (required for clustering)

Running

Running python dehaze_im_bin_search.py dehazes all the images present in haze_image folder and stores the output in out folder.

Files

.
├── cluster_data.py                             # extract patches from images and runs k-means
├── data                                        # training data. need to download separately
│   ├── README.md                               # the details are given in this README.md
│   └── training_images
│       └── filelist.txt                        # files used for clustering
├── data_gen.py                                 # comparator training data generator
├── dehaze_im_binsearch_no_sksparse.py          # dehaze image with a trained comparator without using sksparse
├── dehaze_im_binsearch.py                      # dehaze image with a trained comparator
├── haze_image                                  # hazy images
│   └── 2230089563_06d4982122_z.jpg
├── lib                                         # helper functions
│   ├── __init__.py
│   ├── lib.py
├── LICENSE
├── model                                       # trained comparator model
│   └── comp_c_tpartition_30comp_2A.h5
├── out                                         # output obtained from the hazy images
│   ├── 2230089563_06d4982122_z_out.png         # dehazed output
│   ├── 2230089563_06d4982122_z_t_est.png       # estimated transmittance before smoothing
│   └── 2230089563_06d4982122_z_t.png           # smoothed and interpolated transmittance
├── README.md
├── train_comp_model.py                         # trains the comparator with the generated data
└── visualize_cluster_centers.py                # for visualizing the generated cluster centers

For dehazing an image running the dehaze_im_binsearch.py is sufficient. To train a new comparator the following steps need to be followed.

1. Some fog-free images needs to be gathered. We have used the fog-free files given by Choi et al [1]. The details can be found in data/README.md.
2. Then running cluster_data.py will extract some patches and cluster them.
3. Now to generate the training data for the comparator, data_gen.py needs to be run.
4. After this calling train_comp_model.py trains the comparator with the generated data.

Publication

Sanchayan Santra, Ranjan Mondal, and Bhabatosh Chanda. "Learning a Patch Quality Comparator for Single Image Dehazing." IEEE Transactions on Image Processing 27, no. 9 (2018).

Reference

1. L. K. Choi, J. You, and A. C. Bovik, "LIVE Image Defogging Database," Online: http://live.ece.utexas.edu/research/fog/fade_defade.html, 2015.