The official implementation of the paper "Soft Knowledge-based Distilled Dehazing Networks".
Authors: Le-Anh Tran, Dong-Chul Park
The completion of this project is currently in progress, please stay tuned for updates!
- Updating results for real-world hazy scenes
- Results on benchmarks (I-HAZE, O-HAZE, Dense-HAZE, NH-HAZE, SOTS-Outdoor, HSTS)
- Pre-trained weights (Hugging Face)
- Inference
- Training
- Create environment & install required packages
- Download pre-trained weights from Hugging Face
- Prepare a folder containing test images
python test_dehaze.py --model_path {weight_path} \
--image_path {input_folder_path} \
--output_path {output_folder_path} \
--ch_mul {channel_multiplier} \
--image_size {input_size}
python test_dehaze.py --model_path weights/sddn_ihaze_180_11.h5 \
--image_path ihaze \
--output_path ihaze_dehazed \
--ch_mul 0.25 \
--image_size 512
python test_dehaze.py --model_path weights/sddn_ihaze_180_11.h5 --image_path ihaze --output_path ihaze_dehazed
- Create environment & install required packages
- Prepare dataset folder (a parent directory containing two sub-folders 'A' and 'B' like below):
.../path/to/data
| A (containing hazy images)
| B (containing clean images)
*** Note: a pair of hazy-clean images must have the same name
- Configure training parameters in train.py, the default settings are as below:
# Train Parameters:
n_images = 4
batch_size = 1
epoch_num = 200
critic_updates = 5
learning_rate = 1E-4
loss_balance_weights = [10, 5, 5, 1]
teacher_weight_path = 'path_to_pretrained_teacher_model'
dataset_path = 'path_to_dataset'
python train.py
Please cite our work if you use the data in this repo.
@article{tran2024soft,
title={Soft Knowledge-based Distilled Dehazing Networks},
author={Tran, Le-Anh and Park, Dong-Chul},
journal={Authorea Preprints},
year={2024},
publisher={Authorea}
}
LA Tran