You can find Original article on www.cs.sfu.ca/~mori/, and don’t confuse it with MoneyAisle as source.
Resume
This website illustrates the approach to the Gimpy puzzle, a CAPTCHA test used on Yahoo! to weed out bots. Our method has been successfully tested. Our approach uses common goal algorithms that have been designed for object recognition. The same basic ideas have been applied to finding people in images that match handwritten numbers and recognizing 3D objects.
News
Man or computer? Please try the following tests, “New York Times”, December 10, 2002.
Challenge: Scientists’ computers solve AI-based puzzles, SIAM News, November 2002.
Quick Links:
Background
Our Approach
Results
Related Links
CAPTCHAs used on Yahoo
CAPTCHA Image used on Yahoo.
Background
CAPTCHA is a program that can generate and grade tests that:
Most people can pass, but Current computer programs can’t pass
CAPTCHA stands for “Completely Automated Public Turing test to tell Computers and Humans Apart.” See the CAPTCHA site for more details. The CAPTCHA concept was motivated by real problems faced by Internet companies like Yahoo! and AltaVista. These companies offer free email accounts intended for human use. However, they found that numerous Internet stores were using “bots” – computer programs that automatically register thousands of email accounts from
EZ-Gimpy and Gimpy, the variations we broke, are examples of the basic CAPTCHA word. In EZ-Gimpy CAPTCHA, used by Yahoo!, (Shown above), a user is presented with an image of a single word. The image has been distorted and the background textured. There are enough distortions and obstacles to confuse current OCR (Optical Character Recognition) programs. However, using our computer vision technology, we can correctly identify the word.
Gimpy is a more complicated version of the CAPTCHA word. It consists of ten words represented in distortions and obstacles similar to EZ-Gimpy. The words are also overlapped, providing a CAPTCHA test that in some cases can be challenging even for people. A user must correctly name three out of ten words in the image to pass the test. Our algorithm can pass a more complicated test with a success rate of up to 33%.
Our Approach
The main idea is that our approach to solving Gimpy is the same as the one we use for solving general object recognition problems. Our solution to Gimpy CAPTCHA is simply applying the general frameworks we have used for comparing images of household objects, and even identifying and tracking people in videos. The results from these problems have been analogous. Searching
letters “T”, “A”, “M”, “E” in the image, combining them, and reading the word “tame” is similar to searching for limbs, faces, and body parts to find a person. Real images of people and objects contain a large amount of clutter. Learning to deal with real clutter in Gimpy helps us understand the general problem of object recognition.
Our combined work on finding people and typical objects…
If you would like more detailed information, please see our papers from CVPR 2003.
Results
EZ-Gimpy
Below are some examples of images that were analyzed using our method and the word that was found. Correct words are shown in green, incorrect words are shown in red. For EZ-Gimpy, we ran experiments using 191 images. We were able to correctly identify the words in 176 of these images: a success rate of 92%! Our algorithm takes only a few seconds to process 1 image. If you would like to see our results for all 191 images, please click here.
POLISH
SPACE
JOIN
AGAIN
FLAG
CANVAS
PORTER
MINE
HORSE
JEWEL
WEIGHT
SOUND
RICE
SOCK
Gimpy
The more advanced version of Gimpy CAPTCHA consists of images like the ones shown below. There are 10 words (some repeated). The user is required to list three words present in the image to proceed.
With the clutter in these images, real words, and not just random textured backgrounds, it becomes much more challenging. Additionally, we have to find exactly three words, not just one. Our current algorithm can find 3 correct words and pass this Gimpy test with a success rate of 33%. It’s worth noting that even if we were able to find not a single word correctly, we would only expect to get three right words about 0.7 * 0.7 * 0.7 = 34% of the time. Furthermore, considering our 33% success rate, this CAPTCHA will still be highly ineffective in filtering out “bots” as they can bombard the program with thousands of requests.
The algorithm we use is detailed in our paper mentioned above. Check our results as it will be even harder for you to use the Gimpy version.