Implementation of SRCNN model in Image Super-Resolution using Deep Convolutional Network paper by Tensorflow, Keras. We used Adam with optimize tuned hyperparameters instead of SGD + Momentum.
- Python 3.8 or 3.9
- Tensorflow 2.5.0
- Numpy 1.19.1
- Matplotlib 3.4.3
- Pandas 1.3.4
- OpenCV 4.5.3
You MUST generate data first:
python GenerateData.py --dataset_path="dataset/" --input_size=33 --chanels=3
- dataset_path: path to dataset directory.
- input_size: size of input subimages (does not affect to input size of model).
- chanels: Number of color chanels.
After generating data, you can run this command to begin the training:
python Train.py --epoch=2500 --batch_size=128
NOTE: if you want to train a new model, you can delete all files in checkpoint directory. Your checkpoint will be saved when above command finishs and can be used for next times, so you can train this model on Colab without taking care of GPU limit.
After Training, you can test the model with this command and see the results in test directory:
python Test.py
We trained 60000 epochs and evaluated model with Set5 and Set14 dataset by PSNR:
| Methods | Set5 x2 | Set5 x3 | Set5 x4 | Set14 x2 | Set14 x3 | Set14 x4 |
|---|---|---|---|---|---|---|
| Bicubic Interpolation | 35.7642 | 33.4815 | 31.5064 | 32.7529 | 30.4050 | 28.8816 |
| Resize + Smoothing | 35.3901 | 33.0289 | 31.2894 | 31.9534 | 30.0626 | 28.7333 |
| SRCNN | 36.7126 | 34.5102 | 32.0289 | 33.0441 | 31.0303 | 29.2228 |
You can put your images to demo/images directory first and run this command to see the results in demo/results directory:
python Demo.py --scale=2
We tested with some images from following links and got some results:
- https://www.pixiv.net/en/artworks/67083950
- https://mocah.org/306004-anime-girl-pink-hair-yae-sakura-honkai-impact-3rd-4k.html
- https://images.app.goo.gl/eB7PazUsSfYamCES7
Bicubic x3 (left), resize+smoothing x3 (center), SRCNN x3 (right).
Bicubic x3 (left), resize+smoothing x3 (center), SRCNN x3 (right).
Bicubic x3 (left), resize+smoothing x3 (center), SRCNN x3 (right).
- Image Super-Resolution Using Deep Convolutional Networks: https://arxiv.org/pdf/1501.00092.pdf
- SRCNN Matlab code: http://mmlab.ie.cuhk.edu.hk/projects/SRCNN.html
- T91 and BSDS100 dataset: http://vllab.ucmerced.edu/wlai24/LapSRN/
- Set5 and Set14 dataset: https://github.com/jbhuang0604/SelfExSR#comparison-with-the-state-of-the-art
- YeongHyeon/Super-Resolution_CNN: https://github.com/YeongHyeon/Super-Resolution_CNN
- aditya9211/Super-Resolution-CNN: https://github.com/aditya9211/Super-Resolution-CNN