PyTorch-Image-Dehazing

PyTorch implementation of some single image dehazing networks.

Currently Implemented: AOD-Net: An extremely lightweight model (< 10 KB). Results are good.

Prerequisites:

  1. Python 3
  2. Pytorch 0.4

Preparation:

  1. Create folder "data".
  2. Download and extract the dataset into "data" from the original author's project page. (https://sites.google.com/site/boyilics/website-builder/project-page).

Training:

  1. Run train.py. The script will automatically dump some validation results into the "samples" folder after every epoch. The model snapshots are dumped in the "snapshots" folder.

Testing:

  1. Run dehaze.py. The script takes images in the "test_images" folder and dumps the dehazed images into the "results" folder. A pre-trained snapshot has been provided in the snapshots folder.

Evaluation: WIP.
Some qualitative results are shown below:

Alt text
Alt text
Alt text
Alt text
Alt text
Alt text