Zhiyuan Li, Hao Wei, Chenyang Ge
We propose a real-time image super-resolution method called RVSR, which is inspired by previous work RepViT and FMEN. Our method leverages the efficient architectural designs of lightweight ViTs and the re-parameterization technique to achieve superior performance in real-time super-resolution tasks. RVSR first applies a 3×3 convolution to convert the channel of feature map to the target size (16). Then, RVSR employs 8 stacked RepViT blocks to perform deep feature extraction. As shown in Fig. (a), the RepViT blocks integrate the efficient architectural designs of lightweight ViTs. Additionally, RVSR employs the RepConv module to improve the SR performance while maintaining low complexity, as shown in Fig. (b).
(a) Detailed architecture of RVSR
(b) The RepConv module
pip install -r requirements.txt- Modify the configuration file
configs/conf.yamlaccordingly.
pretrained_path: "pretrained/RVSR_rep.pth"
network:
target: models.inference_arch.RVSR
data:
val:
lr_path: ["Path to your low resolution images"]
hr_path: ["Path to your high resolution images"]- Inference on your own datasets.
python3 test.pyWe provide the training codes for RVSR.
- Modify the configuration file
configs/conf.yamlaccordingly.
network:
target: models.train_arch.RVSR
data:
train:
lr_path: ["Path to your low resolution images"]
hr_path: ["Path to your high resolution images"]
val:
lr_path: ["Path to your low resolution images"]
hr_path: ["Path to your high resolution images"]- Training on your own datasets.
python3 train.pyAfter training, you can utilize re-parameterization techniques to streamline complexity without compromising SR performance.
python3 reparameterize.py --pretrained_path path_to_your_model- Release code and pretrained models
This work is based on RepViT, FMEN and Bicubic++, thanks to their invaluable contributions.