GridDehazeNet

This repo contains the official training and testing codes for our paper:

GridDehazeNet: Attention-Based Multi-Scale Network for Image Dehazing

Xiaohong Liu^*, Yongrui Ma^*, Zhihao Shi, Jun Chen

* Equal contribution

Published on 2019 IEEE International Conference on Computer Vision (ICCV)

[Paper]

---

Prerequisites

• Python >= 3.6
• Pytorch >= 1.0
• Torchvision >= 0.2.2
• Pillow >= 5.1.0
• Numpy >= 1.14.3
• Scipy >= 1.1.0

Introduction

• train.py and test.py are the entry codes for training and testing the GridDehazeNet.
• train_data.py and val_data.py are used to load the training and validation/testing datasets.
• model.py defines the model of GridDehazeNet, and residual_dense_block.py builds the RDB block.
• perceptual.py defines the network for perceptual loss.
• utils.py contains all corresponding utilities.
• indoor_haze_best_3_6 and outdoor_haze_best_3_6 are the trained weights for indoor and outdoor in SOTS from RESIDE, where 3 and 6 stand for the network rows and columns (please read our paper for more details).
• The ./trainning_log/indoor_log.txt and ./trainning_log/outdoor_log.txt record the logs.
• The testing hazy images are saved in ./indoor_results/ or ./outdoor_results/ according to the image category.
• The ./data/ folder stores the data for training and testing.

Quick Start

1. Testing

Clone this repo in environment that satisfies the prerequisites

$ git clone https://github.com/proteus1991/GridDehazeNet.git
$ cd GridDehazeNet

Run test.py using default hyper-parameter settings.

$ python3 test.py

If everything goes well, you will see the following messages shown in your bash

--- Hyper-parameters for testing ---
val_batch_size: 1
network_height: 3
network_width: 6
num_dense_layer: 4
growth_rate: 16
lambda_loss: 0.04
category: indoor
--- Testing starts! ---
val_psnr: 32.16, val_ssim: 0.9836
validation time is 113.5568

This is our testing results of SOTS indoor dataset. For SOTS outdoor dataset, run

$ python3 test.py -category outdoor

If you want to change the default settings (e.g. modifying the val_batch_size since you have multiple GPUs), simply run

$ python3 test.py -val_batch_size 2

It is exactly the same way to modify any other hyper-parameters as shown above. For more details about the meaning of each hyper-parameter, please run

$ python3 test.py -h

2. Training

To retrain or fine-tune the GridDehazeNet, first download the ITS (for indoor) and OTS (for outdoor) training datasets from RESIDE. Then, copy hazy and clear folders from downloaded ITS and OTS to ./data/train/indoor/ and ./data/train/outdoor/. Here we provide the indoor and outdoor training list in trainlist.txt for reproduction purpose. Also, we found some hazy images in training set are quite similar to the testing set (use the same ground-truth images but with different parameters to generate hazy images). For fairness, we carefully remove all of them from the training set and write the rest in trainlist.txt.

If you hope to use your own training dataset, please follow the same folder structure in ./data/train/. More details can be found in train_data.py.

After putting the training dataset into the correct path, we can train the GridDehazeNet by simply running train.py using default settings. Similar to the testing step, if there is no error raised, you will see the following messages shown in your bash

--- Hyper-parameters for training ---
learning_rate: 0.001
crop_size: [240, 240]
train_batch_size: 18
val_batch_size: 1
network_height: 3
network_width: 6
num_dense_layer: 4
growth_rate: 16
lambda_loss: 0.04
category: indoor
--- weight loaded ---
Total_params: 958051
old_val_psnr: 32.16, old_val_ssim: 0.9836
Learning rate sets to 0.001.
Epoch: 0, Iteration: 0
Epoch: 0, Iteration: 100
...

Follow the instruction in testing to modify the default settings.

Cite

If you use any part of this code, please kindly cite

@inproceedings{liuICCV2019GridDehazeNet,
    title={GridDehazeNet: Attention-Based Multi-Scale Network for Image Dehazing},
    author={Liu, Xiaohong and Ma, Yongrui and Shi, Zhihao and Chen, Jun},
    booktitle={ICCV},
    year={2019}
}