Multi-Frame Image Super Resolution using Convolutional Neural Networks

Australian National University, 2017

Download Paper

Citation: Henderson, J., Russell, J. (2017). "Multi-Frame Image Super Resolution using Convolutional Neural Networks." Unpublished undergraduate Thesis, Australian National University.

Abstract:
We explore the application of CNNs to the problem of image super resolution, focussing on SRCNN. We replicate the implementation and confirm the performance of SRCNN, showing that simple, 3-layered, fully convolutional network is able to measurably enhance the level of detail in an image. We extend this network to a multi-frame case, which utilises the temporal relationship of subsequent video frames to add additional context to the network. Our experiments demonstrate a clear improvement of the multi-frame network over the original SRCNN. We also perform a cursory exploration of image alignment and re-introduction of colour channels with mixed results.