This repository contains an op-for-op PyTorch reimplementation of Accelerating the Super-Resolution Convolutional Neural Network.
- About Accelerating the Super-Resolution Convolutional Neural Network
- Installation
- Test
- Train
- Contributing
- Credit
If you're new to FSRCNN, here's an abstract straight from the paper:
As a successful deep model applied in image super-resolution (SR), the Super-Resolution Convolutional Neural Network (SRCNN) has demonstrated superior performance to the previous hand-crafted models either in speed and restoration quality. However, the high computational cost still hinders it from practical usage that demands real-time performance (24 fps). In this paper, we aim at accelerating the current SRCNN, and propose a compact hourglass-shape CNN structure for faster and better SR. We re-design the SRCNN structure mainly in three aspects. First, we introduce a deconvolution layer at the end of the network, then the mapping is learned directly from the original low-resolution image (without interpolation) to the high-resolution one. Second, we reformulate the mapping layer by shrinking the input feature dimension before mapping and expanding back afterwards. Third, we adopt smaller filter sizes but more mapping layers. The proposed model achieves a speed up of more than 40 times with even superior restoration quality. Further, we present the parameter settings that can achieve real-time performance on a generic CPU while still maintaining good performance. A corresponding transfer strategy is also proposed for fast training and testing across different upscaling factors.
git clone https://github.com/Lornatang/FSRCNN-PyTorch.git
cd FSRCNN-PyTorch/
pip install -r requirements.txtcd weights/
bash download_weights.shcd data/
bash download_dataset.shEvaluate the overall performance of the network.
usage: test.py [-h] [--dataroot DATAROOT] [--weights WEIGHTS] [--cuda]
[--scale-factor {2,3,4}]
Fast Super Resolution CNN.
optional arguments:
-h, --help show this help message and exit
--dataroot DATAROOT The directory address where the image needs to be
processed. (default: `./data/Set5`).
--weights WEIGHTS Generator model name. (default:`weights/fsrcnn_4x.pth`)
--cuda Enables cuda
--scale-factor {2,3,4}
Image scaling ratio. (default: `4`).
# Example
python test.py --dataroot ./data/Set5 --weights ./weights/fsrcnn_4x.pth --scale-factor 4 --cudaEvaluate the benchmark of validation data set in the network
usage: test_benchmark.py [-h] [--dataroot DATAROOT] [--image-size IMAGE_SIZE]
[-j N] [--scale-factor {2,3,4}] [--cuda] --weights
WEIGHTS
Fast Super Resolution CNN.
optional arguments:
-h, --help show this help message and exit
--dataroot DATAROOT Path to datasets. (default:`./data/DIV2K`)
--image-size IMAGE_SIZE
Size of the data crop (squared assumed). (default:256)
-j N, --workers N Number of data loading workers. (default:0)
--scale-factor {2,3,4}
Low to high resolution scaling factor. (default:4).
--cuda Enables cuda
--weights WEIGHTS Path to weights.
# Example
python test_benchmark.py --dataroot ./data/DIV2K --weights ./weights/fsrcnn_4x.pth --scale-factor 4 --cudaTest single picture
usage: test_image.py [-h] [--file FILE] [--weights WEIGHTS] [--cuda]
[--scale-factor {2,3,4}]
Fast Super Resolution CNN.
optional arguments:
-h, --help show this help message and exit
--file FILE Test low resolution image name.
(default:`./assets/baby.png`)
--weights WEIGHTS Generator model name. (default:`weights/fsrcnn_4x.pth`)
--cuda Enables cuda
--scale-factor {2,3,4}
Super resolution upscale factor.
# Example
python test_image.py --file ./assets/baby.png --weights ./weights/fsrcnn_4x.pth --scale-factor 4 --cudaTest single video
usage: test_video.py [-h] --file FILE --weights WEIGHTS --scale-factor {2,3,4}
[--view] [--cuda]
FSRCNN algorithm is applied to video files.
optional arguments:
-h, --help show this help message and exit
--file FILE Test low resolution video name.
--weights WEIGHTS Generator model name.
--scale-factor {2,3,4}
Super resolution upscale factor.
--view Super resolution real time to show.
--cuda Enables cuda
# Example
python test_video.py --file ./video/1.mp4 --weights ./weights/fsrcnn_4x.pth --scale-factor 4 --view --cudaLow resolution / Recovered High Resolution / Ground Truth
usage: train.py [-h] [--dataroot DATAROOT] [-j N] [--epochs N]
[--image-size IMAGE_SIZE] [-b N] [--lr LR]
[--scale-factor {2,3,4}] [-p N] [--cuda] [--weights WEIGHTS]
[--manualSeed MANUALSEED]
Fast Super Resolution CNN.
optional arguments:
-h, --help show this help message and exit
--dataroot DATAROOT Path to datasets. (default:`./data/DIV2K`)
-j N, --workers N Number of data loading workers. (default:0)
--epochs N Number of total epochs to run. (default:200)
--image-size IMAGE_SIZE
Size of the data crop (squared assumed). (default:256)
-b N, --batch-size N mini-batch size (default: 16), this is the total batch
size of all GPUs on the current node when using Data
Parallel or Distributed Data Parallel.
--lr LR Learning rate. (default:0.0001)
--scale-factor {2,3,4}
Low to high resolution scaling factor. (default:4).
-p N, --print-freq N Print frequency. (default:5)
--cuda Enables cuda
--weights WEIGHTS Path to weights (to continue training).
--manualSeed MANUALSEED
Seed for initializing training. (default:0)python train.py --dataroot ./data/DIV2K --scale-factor 4 --cudaIf you want to load weights that you've trained before, run the following command.
python train.py --dataroot ./data/DIV2K --scale-factor 4 --weights ./weights/fsrcnn_4x_epoch_100.pth --cudaIf you find a bug, create a GitHub issue, or even better, submit a pull request. Similarly, if you have questions, simply post them as GitHub issues.
I look forward to seeing what the community does with these models!
Chao Dong, Chen Change Loy, Xiaoou Tang
Abstract
As a successful deep model applied in image super-resolution (SR), the Super-Resolution Convolutional Neural Network (SRCNN) has demonstrated superior performance to the previous hand-crafted models either in speed and restoration quality. However, the high computational cost still hinders it from practical usage that demands real-time performance (24 fps). In this paper, we aim at accelerating the current SRCNN, and propose a compact hourglass-shape CNN structure for faster and better SR. We re-design the SRCNN structure mainly in three aspects. First, we introduce a deconvolution layer at the end of the network, then the mapping is learned directly from the original low-resolution image (without interpolation) to the high-resolution one. Second, we reformulate the mapping layer by shrinking the input feature dimension before mapping and expanding back afterwards. Third, we adopt smaller filter sizes but more mapping layers. The proposed model achieves a speed up of more than 40 times with even superior restoration quality. Further, we present the parameter settings that can achieve real-time performance on a generic CPU while still maintaining good performance. A corresponding transfer strategy is also proposed for fast training and testing across different upscaling factors.
@misc{dong2016accelerating,
title={Accelerating the Super-Resolution Convolutional Neural Network},
author={Chao Dong and Chen Change Loy and Xiaoou Tang},
year={2016},
eprint={1608.00367},
archivePrefix={arXiv},
primaryClass={cs.CV}
}