LWDN
Exploration of Lightweight Single Image Denoising with Transformers and Truly Fair Training (ICMR 2023, ACM International Conference on Multimedia Retrieval)
Haram Choi*, Cheolwoong Na, Jinseop Kim, and Jihoon Yang+
*: This work has been done during my 3rd semester of Master Course in Sogang University.
+: Corresponding author.
- Proposes seven lightweight image denosing (LWDN) Transformer baselines.
- Three come from lightweight super-resolution and four are downsized versions of the current best large denoising models.
- Lightweight Super-Resolution Models: SwinIR-light (ICCVW21), ELAN-light (ECCV22), NGswin (CVPR23)
- Large Denosing Models: Restormer (CVPR22), Uformer (CVPR22), CAT (NeurIPS22), ART (ICLR23)
- Trains and compares the baselines in a truly fair manner.
- Identical randomly cropped patches are selectd every iteration in each epoch.
- Records and provides cropped areas (starting point of height (top) and width (left) to be cropped) and random data augmentation (horizontal flip and angle of rotation)
- Robust to random seed in terms of reproducibility.
- Emperically analyzes various aspects of our baselines.
- Hierarchical Structure.
- Spatial vs. Channel Self-Attention.
- Excessive Weight Sharing.
- Still Useful CNN.
News
May 03, 2023: Visual results shared
Apr 04, 2023: Codes released publicly (will be re-available soon due to several issues.)
Apr 02, 2023: Our paper accepted at ICMR 2023
Visual Results
Comparison of Our Lightweight Baselines (Please Click)
Comparison with Large Models (Please Click)
Main Results
Comparison of Our Lightweight Baselines for Color Gaussian Blind Denoising (Please Click)
Comparison of Our Lightweight Baselines for Grayscale Gaussian Blind Denoising (Please Click)
Requirements
Libraries
- Python 3.6.9
- PyTorch >= 1.10.1+cu102
- timm >= 0.6.1
- torchvision >= 0.11.2+cu102
- einops 0.3.0
- numpy 1.19.5
- OpenCV 4.6.0
- tqdm 4.61.2
Datasets (names and path)
TBD
Testing with pre-trained models
You can get the results in the Tables of our paper.
If you have multi gpus that can be used for Distributed Data Parallel (DDP), follow the commands below.
Please properly edit the first five arguments to work on your devices.
TBD
Training from scratch: x2 task
with DDP
- NOTE: argument
batch_sizemeans the size of mini-batch assigned per gpu
TBD
Training by warm-start: x3, x4 tasks
with DDP
TBD
Citation
(preferred)
@inproceedings{choi2023exploration,
title={Exploration of Lightweight Single Image Denoising with Transformers and Truly Fair Training},
author={Choi, Haram and Na, Cheolwoong and Kim, Jinseop and Yang, Jihoon},
booktitle={Proceedings of the 2023 International Conference on Multimedia Retrieval},
year={2023}
}
@article{choi2023exploration,
title={Exploration of Lightweight Single Image Denoising with Transformers and Truly Fair Training},
author={Choi, Haram and Na, Cheolwoong and Kim, Jinseop and Yang, Jihoon},
journal={arXiv preprint arXiv:2304.01805},
year={2023}
}