Repository of "Learning for super resolution" project 2018
The purpose of this project was to compare two approaches to deep learning based super resolution, one based on wavelet transforms and the other on the spatial domain. Models were compared with respect to each other as well as to bicubic interpolation through measures such as PSNR, SSIM and RMSE. It was shown that networks trained on the frequency and spatial domain outperformed bicubic interpolation and the two had very similar performance with wavelets achieving a slightly higher performance.
We trainned two networks in wavelet and spatial domain using residual netkorks and keras.
All image processing methods are in the file "srPreprocessing.py" Networks architecture is implemented in srcnn.py and wavelet_cnn.py in spatial and wavelets domain respectively
Two notebooks for each model pipeline from the high definition image, trainning the models to predicting results
SRCNN_notebook.ipynb
SRCNN_spatial_notebook.ipynb
Comparison.ipynb
Motivations, discussion and results are "report" folder
Learning a Deep Convolutional Network for Image Super-Resolution, Chao Dong, Chen Change Loy, Kaiming He, Xiaoou Tang
Accurate Image Super-Resolution Using Very Deep Convolutional Networks Jiwon Kim, Jung Kwon Lee and Kyoung Mu Lee
Deep Wavelet Prediction for Image Super-resolution Tiantong Guo, Hojjat Seyed Mousavi, Tiep Huu Vu, Vishal Monga
J. Simpkins, R.L. Stevenson, "An Introduction to Super-Resolution Imaging." Mathematical Optics: Classical, Quantum, and Computational Methods, Ed. V. Lakshminarayanan, M. Calvo, and T. Alieva. CRC Press, 2012. 539-564.