This repository contains all the work done , observations noted and models worked upon during my internship at Indian Institute of Technology Gandhinagar under Dr. Ravi S. Hegde. My Research focus was on Image SuperResolution using Deep Learning. Most of my work involved exploring existing solutions and analysing and study their approach. The work was concluded with a Progressive Upsacalling methods using RDNSR with VGG Perceptual Loss for generating high resolution images. The Progressive upscalling can upscale to very high resolution given the RDNSR base network.
Report Document For the Inernship submitted in college : Report on Deep Super Resolution
- Most of my scripts follow a common pattern and can be understood from these repositories.
RNSR Keras
DBPN-SR Keras - The networks given here were worked on 3 Nvidia GTX 1070s with data parallelism in Keras multi gpu mode for training.
- Some of the folder may contain observational notes. These were noted during training.
- The Deep Prior network is not adhering to the original paper. I though of creating residual connections between encoder and decoder network layers thinking it would work out well. It did but only for noise reduction purpose. Also I did not realise this was a U-net architecture.