Image Superresolution using a CNN

See the Code on Github

Abstract

This was one of my favorite class projects, for “Mathematical Foundations of Computer Graphics”, to train a CNN to upsample images to 2x resolution.

The nice part is that students were given the recommended model architecture, but no starter code, so I wrote everything from scratch, which turned out to be a very good learning process.

Results

The model is passed the downsampled image at 240x240 with bilinear interpolation, and it outputs a 2x upsampled version.

The model output is on the left and the original is on the right.

A 10 layer CNN with lr 1e-4. I find it very hard to see any difference.

Ground Truth

Again, a 10 layer CNN with lr 1e-4. You can see a bit of detail is lost in the middle of the castle, and on the stone wall at the bottom.

Ground Truth

CNN with residual connections. This does even better.

Ground Truth