This repository is for Densely Residual Laplacian Network (DRLN) introduced in the following paper
Saeed Anwar, [Nick Barnes], "Densely Residual Laplacian Super-resolution", [arXiv]
The model is built in PyTorch 1.1.0 and tested on Ubuntu 14.04/16.04 environment (Python3.6, CUDA9.0, cuDNN5.1).
Our DRLN is also available in PyTorch 0.4.0 and 0.4.1. You can download this version from Google Drive or here.
Super-Resolution convolutional neural networks have recently demonstrated high-quality restoration for single images. However, existing algorithms often require very deep architectures and long training times. Furthermore, current convolutional neural networks for super-resolution are unable to exploit features at multiple scales and weigh them equally, limiting their learning capability. In this exposition, we present a compact and accurate super-resolution algorithm namely, Densely Residual Laplacian Network (DRLN). The proposed network employs cascading residual on the residual structure to allow the flow of low-frequency information to focus on learning high and mid-level features. In addition, deep supervision is achieved via the densely concatenated residual blocks settings, which also helps in learning from high-level complex features. Moreover, we propose Laplacian attention to model the crucial features to learn the inter and intra-level dependencies between the feature maps. Furthermore, comprehensive quantitative and qualitative evaluations on low-resolution, noisy low-resolution, and real historical image benchmark datasets illustrate that our DRLN algorithm performs favorably against the state-of-the-art methods visually and accurately.
Sample results on URBAN100 with Bicubic (BI) degradation for 4x on “img 074” and for 8x on “img 040”.
The architecture of our proposed densely residual Laplacian attention network (DRLN) with densely residual laplacian modules (DRLM).
Laplacian attention architecture.
Will be added later
-
Download the trained models for our paper and place them in '/TestCode/TrainedModels'.
All the models (BIX2/3/4/8, BDX3) can be downloaded from Google Drive or Baidu or here. The total size for all models is 737MB.
-
Cd to '/TestCode/code', run the following scripts.
You can use scripts in file 'TestDRLN_All' to produce results for our paper or You can also use individual scripts such as TestDRLN_2x.sh
# No self-ensemble: DRLN # BI degradation model, x2, x3 # x2 CUDA_VISIBLE_DEVICES=0 python main.py --data_test MyImage --scale 2 --model DRLN --n_feats 64 --pre_train ../TrainedModels/DRLN_BIX2/DRLN_BIX2.pt --test_only --save_results --chop --save 'DRLN_Set5' --testpath ../LR/LRBI --testset Set5 CUDA_VISIBLE_DEVICES=0 python main.py --data_test MyImage --scale 2 --model DRLN --n_feats 64 --pre_train ../TrainedModels/DRLN_BIX2/DRLN_BIX2.pt --test_only --save_results --chop --save 'DRLN_Set14' --testpath ../LR/LRBI --testset Set14 # x3 CUDA_VISIBLE_DEVICES=0 python main.py --data_test MyImage --scale 3 --model DRLN --n_feats 64 --pre_train ../TrainedModels/DRLN_BIX3/DRLN_BIX3.pt --test_only --save_results --chop --save 'DRLN_Set5' --testpath ../LR/LRBI --testset Set5 CUDA_VISIBLE_DEVICES=0 python main.py --data_test MyImage --scale 3 --model DRLN --n_feats 64 --pre_train ../TrainedModels/DRLN_BIX3/DRLN_BIX3.pt --test_only --save_results --chop --save 'DRLN_Set14' --testpath ../LR/LRBI --testset Set14 # x3 Blur-downgrade CUDA_VISIBLE_DEVICES=0 python main.py --data_test MyImage --scale 3 --model DRLN --n_feats 64 --pre_train ../TrainedModels/DRLN_BDX3/DRLN_BDX3.pt --test_only --save_results --chop --save 'DRLN_BD_Set5' --testpath ../LR/LRBD --testset Set5
All the results for DRLN can be downloaded from GoogleDrive or here. The size of the results is 2.41GB
The performance of state-of-the-art algorithms on widely used publicly available five datasets (SET5, SET14, BSD100, URBAN100, MANGA109), in terms of PSNR (in dB) and SSIM. The best results are highlighted with red color while the blue color represents the second best super-resolution method.
Quantitative results on blur-down degradations for 3x. The best results are highlighted with red color while the blue color represents the second best.
The plot shows the average PSNR as functions of noise sigma. Our method consistently improves over specific noisy super-resolution methods and CNN for all noise levels.
For more information, please refer to our papar
Visual results with Bicubic (BI) degradation (4x) on "img 076" and "img_044" from URBAN100 as well as YumeiroCooking from MANGA109.
Comparisons on images with fine details for a high upsampling factor of 8x on URBAN100 and MANGA109. The best results are in bold.
Comparison on Blur-Downscale (BD) degraded images with sharp edges and texture, taken from URBAN100 and SET14
datasets for the scale of 3x. The sharpness of the edges on the objects and textures restored by our method is the best.
Noisy SR visual Comparison on BSD100. Textures on the birds are much better reconstructed, and the noise removed by our method as
compared to the IRCNN and RCAN for sigma = 10.
Noisy visual comparison on Llama. Textures on the fur, and on rocks in the background are much better reconstructed in our result as
compared to the conventional BM3D-SR and BM3D-SRNI.
If you find the code helpful in your resarch or work, please cite the following papers.
@article{anwar2019drln,
title={Densely Residual Laplacian Super-Resolution},
author={Anwar, Saeed and Barnes, Nick},
journal={arXiv preprint arXiv:1906.12021},
year={2019}
}
@article{anwar2019deepSR,
title={A Deep Journey into Super-resolution: A survey},
author={Anwar, Saeed and Khan, Salman and Barnes, Nick},
journal={arXiv preprint arXiv:1904.07523},
year={2019}
}
This code is built on RCAN (PyTorch) and EDSR (PyTorch). We thank the authors for sharing their codes.