Hybrid Transformer-CNN for Real Image Denoising

Description

We propose a hybrid denoising model based on Transformer Encoder and Convolutional Decoder Network (TECDNet), which can achieved great denosing performance while maintaining a relatively low computational complexity

Network Architecture

Preparation

$ conda create -n venv python=3.8
$ conda activate venv
$ pip install -r requirements.txt

Training

python main_train.py \
	--arch "RBF_TECDNet_S"  \
	--pth_dir "./experiments/TECDNet" \ 
	--data_dir "[your train data dir]" \ 
	--log_dir "./runs" \
	--is_warmup True \ 
	--augment True \
	--img_size 128 \
	--batch_size 32 \
	--n_epochs 250

Test

python main_test.py \
	--arch "RBF_TECDNet_S" \
	--pth_path "[the weights file dir]" \
	--data_path "[test images path]" \
	--device "cuda:0"

Models

Our pre-trained models can be downloaded as following:

	PSNR on SIDD	SSIM on SIDD	Weights Link
TECDNet-S	39.788	0.970	Link

Citation

@Article{9779501,
  author  = {Zhao, Mo and Cao, Gang and Huang, Xianglin and Yang, Lifang},
  journal = {IEEE Signal Processing Letters}, 
  title   = {Hybrid Transformer-CNN for Real Image Denoising}, 
  year    = {2022},
  volume  = {29},
  pages   = {1252-1256},
  doi     = {10.1109/LSP.2022.3176486}
}

Contact

If you have any questions, please contact me(koblod@163.com).