I have a website where people can post Flipbook animations, called Flipnotes. Unfortunately, most of the Popular flipnotes are in the following categories:

  1. MV (Music Video)
  2. AV (Audio Video)
  3. AMV (Animated Music Video)
  4. PV (Preview Video, unfinished content.)

While not popular, I also want to detect:

  1. RP (Roleplay)

As you can see here: http://www.sudomemo.net/browse/?popular - the popular flipnotes are entirely composed of these. Unfortunately, the Most Popular and "Hot" flipnotes are feedbacking into themselves: Increased exposure increases view count and defeats the time decay.

I want to present a more balanced selection of content, and dilute the rankings some. Thankfully, as you can see from the above link, people usually slap "MV" into the thumbnail:

thumbnails of flipnote animations, many of which are labeled "mv"

What I want to be able to do is recognize the presence of "MV","AV","RP", et cetera, so I can sort these Flipnotes better - if people like to watch MVs then those preferences will be tracked - and I can present them with more MVs.

In the end, I want to have a selection more like this: http://www.sudomemo.net/browse/?featured

for the high-ranking animations.

So, I thought, I have my copy of Mathematica on the server - what if I could use it to detect the type of Flipnote it is by the thumbnail frame? People will stick the appropriate acronym in the thumbnail so more people will click on them. While TextRecognize won't work very well at all for this, might there be some way to detect the general shape of the letters in the image? For M:

  • a big stroke up and slightly right

  • a little stroke down and right

  • a little stroke up and right
  • and a big stroke down and slightly right.

How might Mathematica recognize something like this?

  • $\begingroup$ @bobthechemist - I have a Mathematica license, so why not find ways to use Mathematica for things I like to do? edit: there was a comment before this.. $\endgroup$ Jan 26, 2015 at 0:47
  • 3
    $\begingroup$ I have worked in pattern classification for three decades, including problems such as posed. There is much more work involved than the poser thinks, and what needs to be done depends upon the variations in the letter sizes, colors, orientations, thicknesses, etc., and the nature of the clutter. Notice that sometimes the M and V are separated across the frame; A and V and aligned vertically, and so forth. A very simple approach would be to create what is called a "matched template" for each target category (a pixel image of the expected shape of MV, for instance) and use ImageConvolve. $\endgroup$ Jan 26, 2015 at 1:47
  • $\begingroup$ like here $\endgroup$
    – Kuba
    Jan 26, 2015 at 6:54


Browse other questions tagged or ask your own question.