Super Resolution CNN's are used to increase the resolution of an image. These architectures are not used to predict classes or to detect objects. They are an image processing technique.
There are several approaches but usually, in a first step, the image is upsampled using Bicubic interpolation. Afterwards, the quality of the upsampled image is enhanced using a CNN. The second step yields additional improvements in image quality since Bicubic interpolation simply interpolates and a CNN, on the other hand, is able to recognize structures.
(Note: Bicubic interpolation is not a definitive requirement. This is just an example to explain why a CNN is used for resolution improvement.)
The image above is from a paper of the authors of SRCNN. As you can see, a low-resolution image is used as input for the Super-Resolution CNN. The SRCNN then generates a high-resolution image.
This short video has additional information about Super Resolution CNN's and should give a good intuitive understanding.