OCR in scar.

Basically needing to use scar to read a captcha of a number 1-9.. everything else is set up along with if no number found guess. I have no idea how to narrow it down. I am completely baffled.

An example of one of the numbers:


Accuracy needs to be at 50+%!!

Complete random guessing = 11%

ANY small thing helps. Ideas? Comments? procedures?

P.s. I attempted to use some things to clean it up with a program i wrote in c++. Here was the results.




Meh.. No idea what good me being able to clean it up a bit in c++ will do.. But just showing i'm not a COMPLETE retard.

----------------------------------------------------------------------

Thought Process:

Maybe a way to predefine a shape? and have it load the pixels into memory rotate them and see if theirs any combination for that shape? Or a cheaper method of just making a "learning" mode. And record the image as a string along with the right answer and if it comes across the same image again it knows. Which.. with captchas is a BIG if. But couple people running through 30-50 thousand of them a day. Could build up a database. Obviously in "learning" mode it would have to check make sure it doesn't alrdy have that image :x. Just thought processes ignore me =P thinking openly right now. If anyone has any ideas or anything.. my msn is [email protected]

 

I don't think you can write an ocr just with a one letter example...

Do they use the same font all the time? Is it always at that weird angle? How are they spaced out?

 

From what I seen they use 2-4 different fonts, and different sizes and different angles, along with the little black blobs are random and AT TIMES overlap the number. The goal isn't to get a 100% accurate (although that would be insane lol) but just get one 55-60%. like

a few examples

- -
And the results of running the first one through my program


A thought process to start narrowing it down could be to search for a black vertical line thats y height

 

Is it a single number at a time?

 

Yep a single number 1-9

 

ya i would make like a bunch of bitmaps 30-40 (i know its alot) and the bitmaps would be individual traits each number.

 

Length of the work doesnt matter. I wrote over 1000 lines for the program to fetch and organize the captcha long with queue the next and a few other things, could someone here show what one bitmap might be as an example?

 

Why would you need an ocr for something that only shows a random number 1-9?

But if it's same angle all the time, you can have (like mentioned above) an image of each one, then compare it until you get the right one.

It would be faster if you just used random numbers 1-9 instead of trying to use an ocr to crack it, although you might have less valid results...

 

Well it can't be random cause it has to have at least a 50% accuracy or it will get locked out really quick. And its not the same angles, and i had thought of what you said but there would probably be thousands and thousands for each number, could scar handle that?

 

Scar should be able to handle it, but it being able to largely depends on your computer being able to handle that many files being compared for fast results.

 

That should not be an issue, Uhm, tips on how to right the procedure? I am not a scar expert, I usually mess more in c++/php and well I have no damn clue how to do it in php and i know scar has some nice procedures setup all ready to do some of it. Also it makes it easier to run with the other program.

Also would that bit taking the whole bmp and putting it as a string? or just parts of it? My program now loads the captchas from the server and temporarily stores them as png to save loading time. Anyway. If you got a moment can I add you on msn?

 

Anyone with any suggestions would be a great help

 

Does anyone know how I can get the weight of each blob? (amount of pixels making up each blob) with that I could solve it.