As a regression problem, I think ensemble learning in super-resolution may differ from that in a classification problem. Therefore, I decided to read some articles about this topic.
The first paper I read is "Ensemble based deep networks for image super-resolution." This paper discards simple averaging and proposes to find an appropriate weight to fuse different predictions. The method it uses is a sparse coding network. My interest lies in whether simple averaging can performs well. I cannot wait to see the comparison between simple averaging and learnable weighted fusion.
As shown in the experiment, ensemble learning does provide a significant gain for super-resolution, which improves about 0.2 dB. However, the adaptive weighting has little improvement over simple averaging. I am a bit disappointed. This stops me from reading more articles about ensemble learning of super-resolution.