This is the code repository for our work ``DRACO-DehazeNet: An Efficient Image Dehazing Network Combining Detail Recovery and a Novel Contrastive Learning Paradigm" by Gao Yu Lee, Tanmoy Dam, Md. Meftahul Ferdaus, Daniel Puiu Poenar, and Vu Duong, currently appeared as an ArXiV preprint.
Abstract: Image dehazing is crucial for clarifying images obscured by haze or fog, but current learning-based approaches is dependent on large volumes of training data and hence consumed significant computational power. Additionally, their performance is often inadequate under non-uniform or heavy haze. To address these challenges, we developed the Detail Recovery And Contrastive DehazeNet, which facilitates efficient and effective dehazing via a dense dilated inverted residual block and an attention-based detail recovery network that tailors enhancements to specific dehazed scene contexts. A major innovation is its ability to train effectively with limited data, achieved through a novel quadruplet loss-based contrastive dehazing paradigm. This approach distinctly separates hazy and clear image features while also distinguish lower-quality and higher-quality dehazed images obtained from each sub-modules of our network, thereby refining the dehazing process to a larger extent. Extensive tests on a variety of benchmarked haze datasets demonstrated the superiority of our approach.
All codes that is and will be shown here are presented in Tensorflow Keras format.
The link to our paper can be found at https://arxiv.org/abs/2410.14595
(This repo may be updated at various times.)
We evalauted our approach on 4 benchmarked dataset, namely RESIDE [1], NH-HAZE [2], DENSE-HAZE [3], and O-HAZE [4].
NH-HAZE and DENSE-HAZE are comprised of generated real haze that are of non-homogeneous and dense nature, while O-HAZE is comprised of generated real haze that ensures homogenity in the image captured. For RESIDE, we used the SOTS subset for evaluation.
| Datasets | PSNR | SSIM |
|---|---|---|
| RESIDE (SOTS) [1] | 38.08 | 0.9906 |
| O-HAZE [4] | 22.94 | 0.9000 |
| NH-HAZE [2] | 20.82 | 0.7582 |
| DENSE-HAZE [3] | 14.25 | 0.6028 |
RESIDE:
NH-HAZE:
O-HAZE:
DENSE-HAZE:
The DDIRB and ATTDRN architecture:
The overall DRACO-DehazeNet architecture:
- Run the data_loading.py for loading any haze datasets of your choice. Mote that the train, valid and test images needs to be added manually into the arrays from the loaded images of your selected directory before execution of the code.
- Run DDIRB.py, which contains the function for the Dense Dilated Inverted Residual Block for direct dehazing.
- Run ATTDRN.py, which contains the function for the Attention Detail Recovert Network for detail recovery of the intermediate DDIRB dehazed outputs.
- Run DRACO-DehazeNet.py, which also contains the codes for the quadruplet network-based contrastive learning architecture, as well as the main overall DRACO-DehazeNet architecture.
- Run VGG19.py, which contains the feature extraction procedure using VGG19 intermediate layer necessary for the quadruplet contrastive loss function contained in the main training codes.
- Run model training via contrastive_training.py.
- Run evaluate.py, which comprised of the model prediction on the test set, as well as the PSNR and SSIM metric evaluation codes.
- Finally, to compute the flops required, run net_flops.py and use the syntax net_flops(model,table=True).
Weights (json, h5) for the benchmark datasets are found in the weights file folder.
Please cite the following preprint if it is useful in your research:
G. Y. Lee, T. Dam, M. M. Ferdaus, D. P. Poenar, and V. Duong, “Draco-dehazenet: An efficient image dehazing network combining detail recovery and a novel contrastive learning paradigm,” arXiv preprint arXiv:2410.14595,2024
@article{lee2024draco,
title={DRACO-DehazeNet: An Efficient Image Dehazing Network Combining Detail Recovery and a Novel Contrastive Learning Paradigm},
author={Lee, Gao Yu and Dam, Tanmoy and Ferdaus, Md Meftahul and Poenar, Daniel Puiu and Duong, Vu},
journal={arXiv preprint arXiv:2410.14595},
year={2024}
}
[1] B. Li, W. Ren, D. Fu, D. Tao, D. Feng, W. Zeng, Z. Wang, Benchmarking single-image dehazing and beyond, IEEE Transactions on Image Processing 28 (1) (2018) 492–505.
[2] C. O. Ancuti, C. Ancuti, R. Timofte, Nh-haze: An image dehazing benchmark with non-homogeneous hazy and haze-free images, in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2020, pp. 444–445.
[3] C. O. Ancuti, C. Ancuti, M. Sbert, R. Timofte, Dense-haze: A benchmark for image dehazing with dense-haze and haze-free images, in 2019 IEEE international conference on image processing (ICIP), IEEE, 2019, pp. 1014–1018.
[4] C. O. Ancuti, C. Ancuti, R. Timofte, C. De Vleeschouwer, O-haze: a dehazing benchmark with real hazy and haze-free outdoor images, in Proceedings of the IEEE conference on computer vision and pattern recognition workshops, 2018, pp. 754–762.