well before two years I've been here so something handwriting recognition written cellwriter the program presented and cut out some links. At least since the time (but actually a bit longer) I am working now and again with the topic and at some point it occurred to me even in the shower a small factor that is reasonably suitable for individual written characters (glyphs) to a previously trained sample ( and thus a character / characters) assigned. More specifically, it is a matter On-line Recognition and like cellwriter to individual characters, with no segmentation. I can already claim to have much rumprobiert in recent years: starting with a small character array in php-gtk , the single svg -pictures threw out, in between (because the signature does not work out) some tests with characters from a ttf converted to bmp, svg, here is a change because even a new (improvement) idea (again, under the shower), and finally then xournal as a canvas whose xoj files I translated in svg, at some point, even tried HMM s and into the corner thrown over CRM114 gestoplert, calculated a few distributions with R over SVM s thought and finally (again) ended up with Bayes . Tinkering, tinkering, very bad!
now have I with ruby and gtk knitted together a small program to test my criterion. For the Bayesian classification I use the classifier by Lucas Carlson, my little (beta, beta, beta!) Surface serves both to create and store the samples and for classification of characters. And what can I say? For a more sophisticated analysis (such as in cellwriter) stinks my little considered in isolation from natural criterion, but - given the right samples - so already we can classify the odd scribbled letters.
It is here indeed not a complete handwriting recognition, but only about my To test idea - and against the very, very sobering results from yesteryear makes this thing today, making quite a bit. As you can see in the picture (poor) (see the left sidebar at the bottom of the status bar or in the right window of the character that is in the list at the top), the written orders my program after all the correct character glyph (a) - and that in a sample population of 26 lower case letters. The limits of my analysis criterion I are fairly clear: ä will thus not be such a difference from one. For a proper handwriting recognition one has to still try a few other criteria. My (now old) idea is certainly first implemented and for me to return a small portion in terms of "tinkering to handwriting recognition" was completed.
