This is the dlib/image_processing/generic_image.h file.  Then I changed all the
image processing functions so that they use this interface.  All the changes
are very minor, but there are just a lot of them.

Any user code that was using array2d objects to represent images will still
work.  However, this change makes it so that users can use their own custom
image objects with dlib by simply implementing a few global functions for their
image object.

In particular, array2d::set_size() now simply allows any row and
column sizes so long as they are non-negative.

byte.  This makes their serialization format the same as the format for std::string.  I had
to revert the previous mod for this because it didn't work in visual studio.  This change
fixes that.  The code should still be able to read all previously serialized data correctly.
Moreover, any char arrays that aren't null terminated will still be read and written in
their entirety.

serialized using the same format as std::string.  Previously, the
trailing 0 was also saved so if you deserialized them into a std::string
you ended up with a trailing 0 in the std::string which isn't what you
want.

code.  Also made the initial bracketing step a little more efficient.

find_min_single_variable() and find_max_single_variable() routines.  Also made
these functions a little more robust to unusual objective functions.

that happened when using find_max_single_variable().

BOBYQA for model selection.

you to convert between a readonly and write lock but still use RAII.

duplicate copies of the training data since doing so uses a lot of RAM for
large datasets.

aliasing violations in user code, even when things get inlined in unfavorable
ways.

opening files in binary mode.  Oops.  This is now fixed.

documented that you couldn't change the padding, cell size, or window size
parameters between calls to load() and detect().