This repository implements EDSR‑RCNN, which enhances EDSR with an additional Random‑coupled Neural Network (RCNN) channel to inject mid‑/high‑level structural priors into the backbone. RCNN features are fused with the image stream via learnable heads for improved single‑image super‑resolution (SISR).
- Backbone: EDSR [2], plus MDSR and VDSR variants
- RCNN channel: on‑the‑fly ignition maps computed from LR images
- Fusion heads: selectable fusion between image and RCNN channels
- Training: PyTorch with multi‑GPU, background result writers, PSNR logging
- Tuning: Optuna integration for automated hyperparameter search
If you use this code, please cite:
Zuo, X., Liu, H., Liu, M. et al. Super‑Resolution Image Reconstruction based on Random‑coupled Neural Network and EDSR. SIViP 19, 803 (2025). https://doi.org/10.1007/s11760-025-04185-6
Related works: [1], [2].
- Python 3.8+
- PyTorch (CUDA recommended) or MPS (Apple Silicon) on macOS, torchvision
- numpy, scikit‑image, imageio, Pillow
- matplotlib, tqdm
- opencv‑python (for RCNN/data prep)
- thop, fvcore, ptflops (optional FLOPs/params utils)
- optuna, optuna‑dashboard (optional HPO)
Install with pip (example):
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121 # pick the wheel for your CUDA
pip install numpy scikit-image imageio pillow matplotlib tqdm opencv-python thop fvcore ptflops optuna optuna-dashboardEDSR(baseline)EDSR_RCNN(this work)MDSRVDSR
Set --dir_data to the parent directory that contains datasets. For DIV2K it should look like:
<dir_data>/
└── DIV2K/
├── DIV2K_train_HR/
│ ├── 0001.png
│ └── ...
└── DIV2K_train_LR_bicubic/
└── X2/ (or X3/X4 depending on --scale)
├── 0001x2.png
└── ...
When --RCNN_channel on, RCNN ignition maps and merged LR files are generated on first run under the same dataset root:
- No resize:
DIV2K_RCNN_train_LR_bicubic/ - With
--resize on:DIV2K_RCNN_resize_train_LR_bicubic/andDIV2K_RCNN_resize_train_HR/
These are created automatically; you do not need to prepare them manually.
From the src/ directory.
Train EDSR (baseline):
python main.py --model EDSR --scale 2 --dir_data /path/to/datasets \
--data_train DIV2K --data_test DIV2K --data_range 1-790/791-800 \
--n_colors 3 --n_resblocks 32 --n_feats 256 --res_scale 0.1 \
--batch_size 8 --epochs 1000 --save edsr_x2 --cuda cuda:0Train EDSR‑RCNN (with RCNN channel):
python main.py --model EDSR_RCNN --scale 2 --dir_data /path/to/datasets \
--data_train DIV2K --data_test DIV2K --data_range 1-790/791-800 \
--n_colors 3 --n_resblocks 32 --n_feats 290 --res_scale 0.3 \
--RCNN_channel on --model_head SKFusion \
--batch_size 8 --epochs 1000 --save edsr_rcnn_x2 --cuda cuda:0Notes:
- When
--RCNN_channel on, set--model_headto one of:adaptive,SKFusion(case-sensitive). Do not useConvolutionin this case. - For grayscale training, set
--n_colors 1.
Test a trained model:
python main.py --model EDSR --scale 2 --dir_data /path/to/datasets \
--data_test DIV2K --data_range 799-800 \
--n_colors 3 --n_resblocks 32 --n_feats 256 --res_scale 0.1 \
--test_only --pre_train /path/to/model_best.pt --save edsr_x2_test --cuda cuda:0Benchmark datasets are also supported via --data_test Set5+Set14+B100+Urban100.
You can also use the provided script (Linux/macOS):
cd src
sh demo.sh- Experiments:
experiment/<save>/ - Checkpoints:
experiment/<save>/model/ - PSNR curves:
experiment/<save>/test_<DATASET>.pdf - SR/LR/HR images:
experiment/<save>/results-<DATASET>/filename_x<SCALE>_SR.png
--dir_data: dataset root (parent ofDIV2K/,benchmark/, ...)--data_train,--data_test: dataset names (e.g.,DIV2K,Set5)--data_range: train/test split, e.g.,1-790/791-800--scale: upscaling factor(s), e.g.,2or2+3+4--n_colors: 1 for grayscale, 3 for RGB--RCNN_channel:on/offto enable the extra RCNN channel--model_head: when RCNN is on, chooseadaptiveorSKFusion--resize:on/offto pre‑resize images before RCNN processing--cuda: device string, e.g.,cuda:0. On macOS, if MPS is available it is selected automatically; use--cputo force CPU.--test_only: run only evaluation using--pre_train--save: experiment name--n_threads: dataloader workers (set lower on Windows if needed)
Advanced:
- Learning rate and schedule:
--lr,--decay,--gamma - Memory:
--chopfor memory‑efficient forward,--patch_size,--batch_size - Precision:
--precision single|half(half currently applies to both training and testing) - 8‑bit/16‑bit images: set
--rgb_rangeand--ori_rgb_range(e.g.,255for 8‑bit,65536for 16‑bit)
The script src/optuna_utility.py shows how to tune model capacity and training hyperparameters.
cd src
python optuna_utility.pyInside the objective you can add more search dimensions, for example:
lr_star = trial.suggest_float("lr_star", 1e-6, 1e-4, log=True)All trained models, test results, and the DIV2K dataset are hosted on Zenodo: DOI 10.5281/zenodo.13340844.
- DataLoader workers on Windows: try
--n_threads 0–4if you encounter spawn issues. - CUDA OOM: reduce
--batch_size/--patch_size, or enable--chop. - RCNN fusion head: when
--RCNN_channel on, set--model_head adaptiveor--model_head SKFusion. - Paths: ensure
--dir_datapoints to the parent directory that containsDIV2K/.
Codebase derived from EDSR-PyTorch by Lim et al. https://github.com/sanghyun-son/EDSR-PyTorch
If you find this repository helpful, please cite:
@article{zuo2025sr_rcnn_edsr,
title = {Super-Resolution Image Reconstruction based on Random-coupled Neural Network and EDSR},
author = {Zuo, X. and Liu, H. and Liu, M. and others},
journal = {Signal, Image and Video Processing},
volume = {19},
pages = {803},
year = {2025},
doi = {10.1007/s11760-025-04185-6}
}
@article{liu2024rcnn,
title = {Random-coupled Neural Network},
author = {Liu, H. and Xiang, M. and Liu, M. and Li, P. and Zuo, X. and Jiang, X. and Zuo, Z.},
journal = {Electronics},
volume = {13},
number = {21},
pages = {4297},
year = {2024}
}
References:
- [1] Liu H, Xiang M, Liu M, Li P, Zuo X, Jiang X, Zuo Z. Random‑coupled Neural Network. Electronics, 2024.
- [2] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, Kyoung Mu Lee. Enhanced Deep Residual Networks for Single Image Super‑Resolution, CVPR NTIRE 2017.
MIT License (see LICENSE).