SRCNN in Tensorflow

Tensorflow implementation of Image Super-Resolution Using Deep Convolutional Networks.

GT | Bicubic | SRCNN :---: | :---: | :---: | <img src = 'figs/intro1.png'> | <img src = 'figs/intro2.png'> | <img src = 'figs/intro3.png'>

Implementation Details

Our implementation used TensorFlow to train SRCNN. We used almost same methods as described in the paper. We trained the network with 91-image dataset and validated with Set5 dataset while training. At test time, to get the same result size for up-scaling factor 2, 3 and 4, we cropped the test image with 12, which is the Least Common Multiple. Also, we padded border of the ground truth and bicubic interpolated test image by 6 to make same size with SRCNN result.

According to the paper, the best performance on Set5 with upscaling factor 3 is the average PSNR value of 32.75dB with filter size 9-5-5 and ImageNet training dataset, but we were aim to 32.39dB which is the demonstrated average PSNR value when the model is trained with 91-image dataset, 9-1-5 filter size and Y only. After training 12,500 epoch, we got the same value to the paper, 32.39dB.

Pretrained-model with 91-image training dataset and up-scaling factor 3 is given.

Note that we trained and tested with Y-channel. If you want to train and test with 3-channels (YCbCr or RGB), you may add or implement some type-casting code.

Installation

git clone https://github.com/jinsuyoo/SRCNN-Tensorflow.git

Requirements

You will need the following to run the above:

• Tensorflow-gpu
• Python3, Numpy, Pillow, h5py, tqdm

To install quickly, use requirements.txt. Example usage:

pip install -r requirements.txt

Note that we run the code with Windows 10, Tensorflow-gpu 1.13.1, CUDA 10.0, cuDNN v7.6.0

Documentation

To pre-process the train and test dataset, you need to execute the Matlab code.

For those of who cannot execute the Matlab code,

Click here to download the pre-processed training data with 91 dataset. Put the file under SRCNN-Tensorflow directory.

The pre-processed test data with Set5 and Set14 is provided.

Training SRCNN

Use main.py to train the network. Run python main.py to view the training process. Training takes 7-8 hours on a NVIDIA GeForce GTX 1050. Example usage:

# Quick training
python main.py

# Example usage
python main.py --use_pretrained=False \
    --epoch=1000 \
    --scale=4 \

Testing SRCNN

Also use main.py to test the network. Pretrained-model with 91-image training dataset and up-scaling factor 3 is given. Example usage:

# Quick testing
python main.py --is_training=False \
    --use_pretrained=True

# Example usage
python main.py --is_training=False \
    --use_pretrained=True \
    --test_dataset=YOUR_DATASET \
    --scale=4

Please note that if you want to train or test with your own dataset, you need to execute the Matlab code with your own dataset first :)

Results

The average results of PSNR (dB) trained with up-scale factor 3

Code | Dataset | Scale | Bicubic | SRCNN :---: | :---: | :---: | :---: | :---: | SRCNN | Set5 | 3x | 30.39dB | 32.39dB SRCNN-Tensorflow| Set5 | 3x | 30.38dB | 32.39dB

Some of the result images

GT | Bicubic | SRCNN :---: | :---: | :---: | <img src = 'figs/result1_gt.png'> | <img src = 'figs/result1_bicubic.png'> | <img src = 'figs/result1_srcnn.png'> <img src = 'figs/result2_gt.png'> | <img src = 'figs/result2_bicubic.png'> | <img src = 'figs/result2_srcnn.png'> <img src = 'figs/result3_gt.png'> | <img src = 'figs/result3_bicubic.png'> | <img src = 'figs/result3_srcnn.png'> <img src = 'figs/result4_gt.png'> | <img src = 'figs/result4_bicubic.png'> | <img src = 'figs/result4_srcnn.png'>

References

• Official Website

  • We referred to the original Matlab and Caffe code.

• tegg89/SRCNN-Tensorflow

  • We highly followed the structure of this repository.