Deep Laplacian Pyramid Networks for Fast and Accurate Super-Resolution (CVPR 2017)

Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang

IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017

Introduction
Citation
Requirements and Dependencies
Installation
Test Pre-trained Models
Training LapSRN
Training MS-LapSRN
Third-Party Implementation

Introduction

The Laplacian Pyramid Super-Resolution Network (LapSRN) is a progressive super-resolution model that super-resolves an low-resolution images in a coarse-to-fine Laplacian pyramid framework. Our method is fast and achieves state-of-the-art performance on five benchmark datasets for 4x and 8x SR. For more details and evaluation results, please check out our project webpage and paper.

Citation

If you find the code and datasets useful in your research, please cite:

@inproceedings{LapSRN,
    author    = {Lai, Wei-Sheng and Huang, Jia-Bin and Ahuja, Narendra and Yang, Ming-Hsuan}, 
    title     = {Deep Laplacian Pyramid Networks for Fast and Accurate Super-Resolution}, 
    booktitle = {IEEE Conferene on Computer Vision and Pattern Recognition},
    year      = {2017}
}

Requirements and Dependencies

MATLAB (we test with MATLAB R2017a on Ubuntu 16.04 and Windows 7)
Cuda & Cudnn (we test with Cuda 8.0 and Cudnn 5.1)

Installation

Download repository:

$ git clone https://github.com/phoenix104104/LapSRN.git

Run install.m in MATLAB to compile MatConvNet:

# Start MATLAB
$ matlab
>> install

If you install MatConvNet in your own path, you need to change the corresponding path in install.m, train_LapSRN.m and test_LapSRN.m.

Test Pre-trained Models

To test LapSRN / MS-LapSRN on a single-image:

>> demo_LapSRN
>> demo_MSLapSRN

This script will load the pretrained LapSRN / MS-LapSRN model and apply SR on emma.jpg.

To test LapSRN / MS-LapSRN on benchmark datasets, first download the testing datasets:

$ cd datasets
$ wget http://vllab1.ucmerced.edu/~wlai24/LapSRN/results/SR_testing_datasets.zip
$ unzip SR_testing_datasets.zip
$ cd ..

Then choose the evaluated dataset and upsampling scale in evaluate_LapSRN_dataset.m and evaluate_MSLapSRN_dataset.m, and run:

>> evaluate_LapSRN_dataset
>> evaluate_MSLapSRN_dataset

which can reproduce the results in our paper.

Training LapSRN

To train LapSRN from scratch, first download the training datasets:

$ cd datasets
$ wget http://vllab1.ucmerced.edu/~wlai24/LapSRN/results/SR_training_datasets.zip
$ unzip SR_train_datasets.zip
$ cd ..

or use the provided bash script to download all datasets and unzip at once:

$ cd datasets
$ ./download_SR_datasets.sh
$ cd ..

Then, setup training options in init_LapSRN_opts.m, and run train_LapSRN(scale, depth, gpuID). For example, to train LapSRN with depth = 10 for 4x SR using GPU ID = 1:

>> train_LapSRN(4, 10, 1)

Note that we only test our code on single-GPU mode. MatConvNet supports training with multiple GPUs but you may need to modify our script and options (e.g., opts.gpu).

To test your trained LapSRN model, use test_LapSRN(model_name, epoch, dataset, test_scale, gpu). For example, test LapSRN with depth = 10, scale = 4, epoch = 10 on Set5:

>> test_LapSRN('LapSRN_x4_depth10_L1_train_T91_BSDS200_pw128_lr1e-05_step50_drop0.5_min1e-06_bs64', 10, 'Set5', 4, 1)

which will report the PSNR and SSIM.

Training MS-LapSRN

Setup training options in init_MSLapSRN_opts.m, and run train_MSLapSRN(scales, depth, recursive, gpuID), where scales should be a vector, e.g., [2, 4, 8]. For example, to train MS-LapSRN with D = 5, R = 2 for 2x, 4x and 8x SR:

>> train_MSLapSRN([2, 4, 8], 5, 2, 1)

To test your trained MS-LapSRN model, use test_MS-LapSRN(model_name, model_scale, epoch, dataset, test_scale, gpu), where model_scale is used to define the number of pyramid levels. test_scale could be different from model_scale. For example, test MS-LapSRN-D5R2 with two pyramid levels (model_scale = 4), epoch = 10, on Set5 for 3x SR:

>> test_MSLapSRN('MSLapSRN_x248_SS_D5_R2_fn64_L1_train_T91_BSDS200_pw128_lr5e-06_step100_drop0.5_min1e-06_bs64', 4, 10, 'Set5', 3, 1)