EDN-GTM: A novel Encoder-Decoder Network with Guided Transmission Map for single image dehazing

The official implementation of the paper "A Novel Encoder-Decoder Network with Guided Transmission Map for Single Image Dehazing"

(Presented in International Conference on Industry Science and Computer Sciences Innovation 2022 (iSCSi'22), Porto, Portugal, March 9-11, 2022)

Authors: Le-Anh Tran, Seokyong Moon, Dong-Chul Park

I. Introduction

i. Publication: Procedia Computer Science 204

ii. Blog: Towards Data Science

iii. Results on Papers With Code

• 
• 
• 
• 

iv. Abstract:

A novel Encoder-Decoder Network with Guided Transmission Map (EDN-GTM) for single image dehazing scheme is proposed in this paper. The proposed EDN-GTM takes conventional RGB hazy image in conjunction with its transmission map estimated by adopting dark channel prior as the inputs of the network. The proposed EDN-GTM adopts U-Net for image segmentation as the core network and utilizes various modifications including spatial pyramid pooling module and Swish activation to achieve state-of-the-art dehazing performance. Experiments on benchmark datasets show that the proposed EDN-GTM outperforms most of traditional and deep learning-based image dehazing schemes in terms of PSNR and SSIM metrics. The proposed EDN-GTM furthermore proves its applicability to object detection problems. Specifically, when applied to an image preprocessing tool for driving object detection, the proposed EDN-GTM can efficiently remove haze and significantly improve detection accuracy by 4.73% in terms of mAP measure.

v. Architecture:

II. Requirements

Main dependencies (or equivalent):

• CUDA 10.0
• CUDNN 7.6
• OpenCV
• Tensorflow 1.14.0
• Keras 2.1.3

For other packages, simply run:

$ pip install -r requirements.txt

III. Test using Pre-trained Weights

1. Download Pre-trained Weights

• Download pre-trained weights from GoogleDrive
• There are 4 weight files available for test on I-HAZE, O-HAZE, Dense-HAZE, NH-HAZE datasets (respective to their filenames)
• Make a folder 'weights' to locate downloaded weight files

2. Correct Data Paths in test_on_images.py

• Path to pre-trained weight: weight_path
• Path to output directory: output_dir
• Path to folder containing test images: img_src

3. Run Test Script

$ python test_on_images.py

IV. Train Network

1. Prepare Dataset

• Each image in a clean-hazy image pair must have the same name
• Make Folder 'A' and Folder 'B' containing hazy and clean images, respectively

2. Correct Data Paths in train.py

• Path to folder containing train data: path/to/data
• Note that path/to/data nevigates to the parent directory of 'A' and 'B' like below:

-- path/to/data /
                |- A (containing hazy images)
                |- B (containing clean images)

3. Run Train Script

$ python train.py

V. Results

1. Quantitative Results

(a) I-HAZE & O-HAZE Datasets

Approaches	I-HAZE Dataset		O-HAZE Dataset
	PSNR	SSIM	PSNR	SSIM
DCP (TPAMI’10)	14.43	0.7516	16.78	0.6532
CAP (TIP’15)	12.24	0.6065	16.08	0.5965
MSCNN (ECCV’16)	15.22	0.7545	17.56	0.6495
NLID (CVPR’1)	14.12	0.6537	15.98	0.5849
AOD-Net (ICCV’17)	13.98	0.7323	15.03	0.5385
PPD-Net (CVPRW’18)	22.53	0.8705 *	24.24 *	0.7205
EDN-GTM	22.90 *	0.8270	23.46	0.8198 *

(b) Dense-HAZE & NH-HAZE Datasets

Approaches	Dense-HAZE Dataset		NH-HAZE Dataset
	PSNR	SSIM	PSNR	SSIM
DCP (TPAMI’10)	10.06	0.3856	10.57	0.5196
DehazeNet (TIP’16)	13.84	0.4252	16.62	0.5238
AOD-Net (ICCV’17)	13.14	0.4144	15.40	0.5693
GridDehazeNet (ICCV’19)	13.31	0.3681	13.80	0.5370
FFA-Net (AAAI’20)	14.39	0.4524	19.87	0.6915
MSBDN (CVPR’20)	15.37	0.4858	19.23	0.7056
KDDN (CVPR’20)	14.28	0.4074	17.39	0.5897
AECR-Net (CVPR’21)	15.80 *	0.4660	19.88	0.7173
EDN-GTM	15.43	0.5200 *	20.24 *	0.7178 *

2. Qualitative Results

(a) I-HAZE Dataset

(b) O-HAZE Dataset

(c) Dense-HAZE Dataset

(d) NH-HAZE Dataset

3. Application to Object Detection

(a) Dehazing in Driving Scenes

Visual results on synthesized hazy driving scenes (left: synthesized hazy image, right: dehazed image).

(b) Object Detection Performances

Object detection performances on two sets of hazy (left) and dehazed (right) images (red: ground-truth box, green: predicted box, blue: zoom-in region).

VI. Citation

@article{Tran_2022,	
year = 2022,	
publisher = {Elsevier {BV}},	
volume = {204},	
pages = {682--689},	
author = {Le-Anh Tran and Seokyong Moon and Dong-Chul Park},	
title = {A novel encoder-decoder network with guided transmission map for single image dehazing},	
journal = {Procedia Computer Science}}

Have fun!

LA Tran

Dec. 2021