PyTorch implementation of "Light Field Image Super-Resolution with Transformers", IEEE SPL 2022. [pdf].
- We make the first attempt to adapt Transformers to LF image processing, and propose a Transformer-based network for LF image SR.
- We propose a novel paradigm (i.e., angular and spatial Transformers) to incorporate angular and spatial information in an LF.
- With a small model size and low computational cost, our LFT achieves superior SR performance than other state-of-the-art methods.
- PyTorch 1.3.0, torchvision 0.4.1. The code is tested with python=3.6, cuda=9.0.
- Matlab (For training/test data generation and performance evaluation)
We used the EPFL, HCInew, HCIold, INRIA and STFgantry datasets for both training and test. Please first download our dataset via Baidu Drive (key:7nzy) or OneDrive, and place the 5 datasets to the folder ./datasets/
.
- Run
Generate_Data_for_Training.m
to generate training data. The generated data will be saved in./data_for_train/
(SR_5x5_2x, SR_5x5_4x). - Run
train.py
to perform network training. Example for training LFT on 5x5 angular resolution for 4x/2xSR:$ python train.py --model_name LFT --angRes 5 --scale_factor 4 --batch_size 4 $ python train.py --model_name LFT --angRes 5 --scale_factor 2 --batch_size 8
- Checkpoint will be saved to
./log/
.
- Run
Generate_Data_for_Test.m
to generate test data. The generated data will be saved in./data_for_test/
(SR_5x5_2x, SR_5x5_4x). - Run
test.py
to perform network inference. Example for test LFT on 5x5 angular resolution for 4x/2xSR:python test.py --model_name LFT --angRes 5 --scale_factor 4 \ --use_pre_pth True --path_pre_pth './pth/LFT_5x5_4x_epoch_50_model.pth' python test.py --model_name LFT --angRes 5 --scale_factor 2 \ --use_pre_pth True --path_pre_pth './pth/LFT_5x5_2x_epoch_50_model.pth'
- The PSNR and SSIM values of each dataset will be saved to
./log/
.
- Quantitative Results
- Efficiency
- Visual Comparisons
- Angular Consistency
- Spatial-Aware Angular Modeling
If you find this work helpful, please consider citing:
@Article{LFT,
author = {Liang, Zhengyu and Wang, Yingqian and Wang, Longguang and Yang, Jungang and Zhou, Shilin},
title = {Light Field Image Super-Resolution with Transformers},
journal = {IEEE Signal Processing Letters},
year = {2022},
}
Any question regarding this work can be addressed to zyliang@nudt.edu.cn.