There exists such a database online under the URL http://IAMMP.org/ 
This was the archive of the Yahoo 'rollscanners' group.  The acronym is
for "International Association of Mechanical Music Preservationists".
Like many on-line things this is pretty much defunct.  Without a
dedicated person to maintain interest, such archives tend to fall in
and out of favor.  Passwords get forgotten, and members move onto other
interests.

The products of the most popular scanning software are three files.
The image file is known as a CIS file.  This name comes from the name
of the scanning array which is typically a 'Contact Image Scan' array.
Most flat-bed image scanners use this type of device.  Such electronics
have been adapted specifically to scan rolls.

There are provisions in the header of the CIS file to track things like
the song name, instrument copyright license and scanning hardware.  In
practice this is seldom updated, as scanning a roll is a time-consuming
process, with good intentions to add this information after the batch
of the rolls are processed.

The program Rollscanconverter.exe takes the CIS file as input and creates
a playable MIDI file and also a e-roll file.  e-roll files typically
contain the additional long name extension "_ScanImageMIDI.mid".
Rollscanconverter.exe uses an XML database to map the holes to MIDI
events.  This does not always produce clean tracker bar images due to
the way the scales are mapped.

MIDI Type 1 files are a side effect of wanting to preserve the database
information from the CIS file.  Rollscanconverter.exe can also be used
to add this information to the file during the conversion process.
Again, it takes time to look up things like composer names and copyright
license dates so this often gets ignored.

Once the files are converted to MIDI, most enthusiasts like to do an
additional processing step by editing the MIDI file with Cakewalk.
This adds an additional point where the song info database can be input
or edited.

It is important to understand that MIDI is just an encoding format
like an ASCII text file.  It is the base level that all hardware can
read and interpret.  MIDI Type 0 files are "single page" documents,
MIDI Type 1 files are "multiple page" documents.

I personally put together a database of most of the Wurlitzer APP rolls
produced between about 1900 to 1967.  The surviving catalogs, auction
sale information and broadsheets, such as produced by the late Ray
Siou, indicate there were about 10,000 song titles offered for sale.
About 3000 of these have survived to be scanned and converted to MIDI
by a number of different people.

There is strong evidence that most of the Wurlitzer APP roll music
produced between 1927 and when T.R.T. took over production around 1939
that the music was licensed from and is identical to QRS arrangements.
All the masters post-1939 still exist and are currently held in the
Hershel Carousel Museum.  Rolls from this catalog were offered for sale
by Doyle Lane.

Most of the popular APP stuff that has survived are show tunes (almost
any song that has won the equivalent of an Academy award or Tony is in
these catalogs).  Many popular songs from the late 1940s, 1950s and
1960s are also represented.  This makes providing a public database
difficult as the license holders of these songs have their own agenda.

Wurlitzer APP is only one roll format, and not a particularly popular
one.  We now have an Ampico piano in the house.  This may be a more
representative format.  Other folk have easier access to Duo-Art
technology, so that forms even a different database.

88-note rolls are ubiquitous; some are more interesting than others.
To sort through these and take the time to index them is more time than
most people have.  Perhaps this is why "Silver Threads Among the Gold"
turns up so often.

While it is often stated that there is a desire for 1920s and 1930s
music (which are mostly under copyright) the real effect is that the
stuff that is royalty free, is the highbrow classical stuff from about
1890 through about 1918.  Anything post 1924 or so is going to be
licensed by someone.

Much depends on when the composer died, or simply the last time the
lyrics were updated or the song arranged.  This can take copyrights
like Happy Birthday (1880) to still be enforced 130 years later as the
rights were renewed by the authors in the 1920s).  According to the
Supreme Court, corporations are persons and immortal.  Once a corporation
holds a copyright they can enforce it.  While not the spirit of the law
this is the effect and evolution of the effect of obfuscating trademarks
and copyrights.

In practice it is much easier to scan the rolls for one's own personal
use.  This is why there is, and probably will be for the foreseeable
future, a lot of duplicated effort.

Julie Porter