Maptcha: An Experiment in Map-Based CAPTCHA

Beyond that, the examples use extremely hard to see letters... that completely defeats the purpose. There's no need if the challenging part for the computer is detecting satellite imagery.

There are multiple letters on some of the images, sometimes hovering between two images. Again, why? The computer is going to have no problem divvying up the images, but the ambiguity is confusing for humans.

Make the letters as easy to read as possible please. If that breaks your captcha, find a new captcha method.

Having said that, there are a couple of issues: 1) This is really not a hard problem for a computer to solve. Detecting different image types is pretty trivial. You made it a lot easier by making the letters sit in the boxes in a regular manner. 2) A high percentage of captchas are now broken by humans, instead of robots. This is not a problem you will likely encounter in a small site, though.

(1) Is an easier problem to solve. You could do "input the letters near the cat", or "input the letters near the map of florida".

(2) Is harder, but most of the people who break captchas live in 3rd world countries, so you could possibly arbitrage cultural knowledge (see florida example above). They will catch on pretty damn quick, though.

On the upside, you can usually type a wrong number into the ReCAPTCHA. ;)

Technically it has two weaknesses:

- non-map images are chosen from a small set. Good implementation will need equally large set of non-map (and presumably non-offensive and deceptively map-like) images.

- satellite photos don't seem very hard to identify, e.g. even using histogram as an input to a simple classification algorithm could give decent accuracy (gray/brown/dark green color scheme dominates) and tree shadows, cars and rooftops may be easy features to detect with fancier algorithms.

Couple of notes: 1. The ultimate intention is to either pull images (both satellite and decoy) from the web as they are needed or create a database of images and update it regularly. Thus identifying images as real or decoy will only work on a short-term basis; there will always be new images entering the fold.

2. If people like this concept well enough to keep at it, I'm planning to automate some kind of color-matching scheme so that it's not possible to just look at the histograms of the images and identify the satellite images thusly.

3. I'm only just starting to learn about OCR, so at this moment I have no idea whether it's worth bothering to disguise letters or even whether it's worth using letters at all. I prefer having a two-tiered approach; having to ID a satellite image AND identify characters. If the bots are too good at identifying the characters and/or it's too hard for humans, that just leaves the satellite question and from what I've read that's not a high enough level of entropy.

4. I don't have anything lined up for an audio solution just yet. I'm fooling around with a few concepts, but nothing worth mentioning just yet.

5. Thanks to all those who wrote in suggestions/solutions and signed up!

$('.trueletterimage').each(function () {$('#userdata').val($('#userdata').val() + $(this).attr('src').split('.')[0].split('/')[1])});

Copy paste into the console and enjoy.

Ok, obviously the above is mostly in jest. I always enjoy seeing new captchas, despite most having major flaws. I do have to echo other comments that mention the user difficulty issue. Using this in production would require more explanatory text than your average captcha, and it does seem to take a little longer to figure out than, say, a recaptcha image. That said, likely the bigger issue is one of image classification allowing bots to solve most of them.

This however, is difficult to read for a human and probably easy to write a bot for.

My guess is the contiguous nature of the satellite images might make the problem a bit easier, sort of like Battleship.

Finally a good way to keep potential Customers from using your product!

http://techcrunch.com/2012/03/29/google-now-using-recaptcha-...