Single-image super resolution (ISR) addresses the problem of reconstructing high-resolution images given their low-resolution (LR) counterparts. ISR finds use in various computer vision applications: from security and surveillance imaging, satellite imaging, medical imaging to object recognition. This ill-posed problem has multiple solutions for any LR input. Deep learning approaches, specifically convolutional neural networks (CNN) have proven to be able to achieve better results than the classic interpolation based methods. At idealo.de we are using this technology to ensure a 4K image for each product in our catalog.

In this talk I'll briefly introduce some of the recent literature, I would also touch upon the main parts of our project, which includes a Keras implementation and training of a combination of some of the SOTA architectures for ISR.

Specifically, we will go through the different training steps, analyzing the results from each of them and derive insights that ultimately helped us improve the model performance:

• at first, we will look at the results of the ISR network trained in a standard settings and the pitfalls of this approach;
• I would then go on to show how we tried to remove compression artifact using the same network;
• to improve upon these results, a form of perceptual loss is introduced using deep features from a pre-trained popular classification network, the VGG19;
• then, after a few tricks to obtain better results, we will finally introduce a discriminator network and further train the previous model in a GAN fashion;
• finally, we'll have a look at how model averaging can help ISR to achieve a wide range of results by only training 2, or a few, models.

To conclude: I would like to present some ideas for the next steps, share some of the lessons learned and, in hindsight, comment on what I would have done differently.