Tip:
Highlight text to annotate it
X
Our today´s world is full of text. We meet with text on daily basis. It a standard of
our communication.
But imagine that we build a robot. He can easily lost himself around all of the text.
So scientists, about 80 years ago, described process of Optical character recognition,
the OCR.
Nowadays OCR programs are mostly used for digitization of old books, but to be accurate,
images have to be high resolution. And that is problem because of high data bandwidth
and big latencies.
Imagine a robot that is moving today´s street. To figure out what is around him, he has to
recognize text around him, so he take photos, but letters are from big distance and bad
angle. And this is the same problem as reCAPTCHA challenge gives us.
So, my project beCAPTCHA is about recognition of low-resolution images with distortion.
This is flowchart of algorithm that I designed for decoding CAPTCHA. Firstly, it splits image
into to images that every image represents one word. Next step is that it load a Bézier
curves and computes typographic baseline. Then it balances a text and sends it to tesseract
recognition core. Output text is considered if it has sense. If not it uses a Hough transformation
to get skewness, it rotates a balances a text. The same thing with another word.
With this algorithm I was able to achieve accuracy of 5%, but I wanted to create artificial
intelligence, so I created a plan how to integrate neural network.
I did not achieved presupposed accuracy of 10%, but by harvesting data for 3 months I
found security vulnerability. As you can see I found a way how to figure out which word
is recognition and which is verification.
So, I found a way how to effectively solve about 5% of all non-recognizable words.
Thanks for watching.