While caching of decoded images happens automatically within MuPDF, it is perhaps worth saying a small amount about it.
Whenever a decoded image is requested, MuPDF searches in the store (see chapter 7 Memory Management and The Store) to see if a suitable pixmap exists there already. If one is found, the store remembers that is has been reused, and returned immediately - no decoding is done.
If no suitable pixmap is found, MuPDF calculates how large the image would be on a rendered page. By comparing this size to the native size of the image, it calculates a log 2 subsampling factor to use. That is, it attempts to avoid decoding the image at full size, when one 1/2 (or 1/4 etc) of the width/height would do.
A log 2 subsampling is used because a) some compression formats such as JPEG can achieve this as part of their decompression run, and b) it is easy to rapidly shrink decompressed pixmaps in this way.
The decoded and subsampled image is then placed into the store so that it will (hopefully) be found the next time a decode of the image is requested.