sirracha / srcnn_satellite

Nearly complete submissions for Super Resolution Convolutional Neural Network (SRCNN) algorithm.

These directories contain different approaches to working with SRCNN.

They include nearly complete submissions for Super Resolution Convolutional Neural Network (SRCNN) algorithm.

SRCNN-Tensorflow is our main focus for completing this application.

SRCNN-COLAB showcases how to go about setting up Tensorflow models in the Google Colab environment.

SRCNN-Keras is a nonworking proof of concept.

Model Architecture

Super Resolution CNN (SRCNN)

The model above is take from user titu1994 who I cite as a source for this project.

I have some differences from the original paper, and from other scientists who have worked on this topic:
[1] Used the Adam optimizer instead of Stochastic Gradient Descent.
[2] Stride is set to 21 instead 14.

My models underperform compared to the results posted in the paper. I was unable to replicate the GPU power and time for running to get my PSNR as high, ie. 32dbs.

I have optimized this model for use on different imagery instead of the usual 91 images from ImageNET.