Skip to content
Go to file


Failed to load latest commit information.
Latest commit message
Commit time
Aug 15, 2019
Aug 15, 2019
Aug 16, 2019

Image Super Resolution using a Multi-Resolution Deep Laplacian Neural Network

In this repository, I attempted to implement a convolutional neural network that increases the resolution of a photo by 2 and 4 times. The algorithm is based on the one described by Wei-Sheng Lai and colleagues, but I adapted it to suit my computational limit (i.e training using only one GPU on Colab for a few hours).

Training Details:

I use the Microsoft COCO dataset as training data. I resize each image first to 512 by 512, then crop a square patch of size 128, randomly flipping the image horizontally and interpolate by a random amount of scale between 0.5 and 1.0 (using the nearest neighbor mode for speed). I also use the pixel shuffle convolution for upsampling instead of transposed convolution.
I train for only 10 epochs using Adam optimizer with default setting, which means that the results could be much better with more hyperparameter tuning or longer training.


Code for training can be found in, and code for inference can be found in Note that this project made use of my nn-toolbox for various components and training procedures (you might want the experimental branch for latest updates). It also use another repository of mine for the computation of the SSIM metric, generative-models, which is unreleased at the moment.
Pretrained weights can be downloaded here.


The high resolution image is first resized using bicubic interpolation to obtain a lower resolution version. This version is then upsampled by both the neural network and the bicubic algorithm for comparison.

Cat 1:

Original high resolution image:


Super resolution using bicubic (PSNR: 24.6858, SSIM: 0.8688):


Super resolution using neural network (PSNR: 25.2227, SSIM: 0.8738):


Cat 2:

Original high resolution image:


Super resolution using bicubic (PSNR: 27.5483, SSIM: 0.8965):


Super resolution using neural network (PSNR: 27.3232, SSIM: 0.9014):



  1. Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, Ming-Hsuan Yang. Deep Laplacian Pyramid Networks for Fast and Accurate Super-Resolution.

  2. Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, Ming-Hsuan Yang. Fast and Accurate Image Super-Resolution with Deep Laplacian Pyramid Networks.

  3. COCO dataset: Licensed under a Creative Commons Attribution 4.0 License.

  4. The test images are from and


Super resolution of images using neural network



No releases published


No packages published


You can’t perform that action at this time.