Image dumps are available from
http://download.wikimedia.org/images/
but are to be used "at your own risk" - as explained in
http://download.wikimedia.org/images/README_ABOUT_COPYRIGHT.txt
Using the appropriate database dumps you could probably select out a
reasonable subset using the template-and-category based labelling
system - i.e. delete those with no such tags, those marked as "fair
use", and any other licenses likely to be incompatible with your
mirror.
This is the sort of information that should be part of a "best
practices" Wikipedia: namespace topic.