In this project, we investigate the possibility of replacing the L2 loss with perceptually derived loss functions (SSIM, MS-SSIM, etc.) in training an end-to-end dehazing neural network. Objective experimental results suggest that by merely changing the loss function we can obtain significantly higher PSNR and SSIM scores on the SOTS set in the RESIDE dataset, compared with a state-of-the-art end-to-end dehazing neural network (AOD-Net) that uses the L2 loss. The best PSNR we obtained was 23.50 (4.2% relative improvement), and the best SSIM we obtained was 0.8747 (2.3% relative improvement.) For more details, please read this report.

• Ubuntu 16.04 64 bit; other OSs were not tested
• nvidia-caffe with CUDA 8 and cuDNN v7: branch caffe-0.15, commit 4b8d54d892116b9cb6822a917065a616f56b1292; the original BLVC caffe did not support the training scripts very well, but for testing purpose, the BVLC caffe should work
• PyCaffe has to be installed and included in your python search path. For example, run export PYTHONPATH=$PATH_TO_CAFFE/python:$PYTHONPATH, where $PATH_TO_CAFFE is the caffe root dir
• PyCaffe may need a lot of other dependencies, you can install anaconda to resolve most of them
• Matlab 2017a and up; older versions should work but not tested; if you don't have the parallel computing toolbox, just change all the parfor in evaluate.m to for
• Include ./src/psnr_633.m and ./src/ssim_633.m in your Matlab's search path

• Copy loss.py to $PATH_TO_CAFFE/python/
• Rename it to pyloss.py

• Open file run.sh
• Change sots_dir to where you put the RESIDE SOTS set. Please use absolute path. You can obtain this testing set here. We used the 1,000-image version of SOTS that contains 500 indoor and 500 outdoor images, and we assume that they are all put into sots_dir.
• Change gt_dir to where you put the groundtruth images. Please use absolute path. Again, we assume that you have put all the groundtruth images in the same folder
• Change dehaze_dir to the desired output directory. Please use absolute path.
• Run ./run.sh
• The script may run for a while (<30 min) and may open your Matlab
• When it finishes, you can find the dehazed images in dehaze_dir and the average PSNR and SSIM printed on the screen. You can also find the PSNR and SSIM for each test sample in a .mat file named as result_per_image_test.mat under your current directory

• data10k/solver_msssimL2_10k_fine_tune_iter_9000.caffemodel
  • PSNR: 23.43
  • SSIM: 0.8747
• data10k/solver_msssimL2_10k_fine_tune_0.7_iter_9000.caffemodel
  • PSNR: 23.50
  • SSIM: 0.8676

• Guanlong Zhao ([email protected]), Department of Computer Science and Engineering, Texas A&M University
• Yu Liu ([email protected]), Department of Electrical and Computer Engineering, Texas A&M University