How can a resizer distinguis nude from non-nude?

December 3, 2007 9:58 AM

How does my photo resizer company know when I post a nude?

I sometimes use a photo resizer for my blog. When the photo is saved and posted, if it is non-nude, all stays as posted. If, however, the photo is of a nude, sometime later in the day, the photo is deleted and the viewer notified that photo deleted. How does the company know there is a nude when so many people use their services? Is this done atomatically or sorted out by humans?

posted by Postroad to Computers & Internet (17 answers total)

Can these be viewed by the public? If so, it's probably users reporting it that causes it to be reviewed by staff.
posted by winston at 10:00 AM on December 3, 2007

By humans would be my guess. I used to have a job working in computer forensics part of which was scanning people's hard drives for porn. You would be surprised how quickly you can page through 30 or so thumbnails at a time and spot the porn.
posted by frieze at 10:01 AM on December 3, 2007

I would guess by human as well, but I think I've heard of software packages that look for large portions of pixels that have a flesh tone. Perhaps a combination, a program that flags potential problems to be checked by a person.
posted by shinynewnick at 10:11 AM on December 3, 2007

frieze, that job description is awesome.
posted by shinynewnick at 10:15 AM on December 3, 2007

Naturally, I tried to visit your website to see there's any porn--I mean "nude" photos posted--but your profile links to a very old blog.

Anyway...

Some cameras, photo editing programs, and search engines now recognize faces, so I don't see why some might be able to recognize nude bodies.
posted by HotPatatta at 10:29 AM on December 3, 2007 [1 favorite]

There has been a lot of success with writing tools that do it automatically as well. Generally they look for certain amounts of skin-tone and shapes. Those will flag potential images for human review.

Never underestimate the power of the web though. Millions of bored, lonely people just looking for something to complain about.
posted by jeffamaphone at 10:30 AM on December 3, 2007

Could it be due to the amount of $colour in the photo?
posted by Rabulah at 10:46 AM on December 3, 2007

I remember reading an article (in Wired perhaps?) about a decade ago that some research group was creating a "Naked People Finder" algorithm for images, so I suspect that it's completely possible.

Ahh yes, here it is.
posted by Wild_Eep at 11:00 AM on December 3, 2007

A couple of years back, I worked as a QA engineer on a spam filter product that had an image recognition module that could heuristically detect nude photos (as attachments or inline images) with a fairly high degree of accuracy. It worked pretty well, about a 99% detection rate with less than .01% false positives. They're probably using something similar.
posted by deadmessenger at 12:15 PM on December 3, 2007

If it's "sometime later in the day", humans are probably involved in the process.
posted by qvtqht at 12:30 PM on December 3, 2007

Yeah, while you can certainly come up with algorithms to analyze the image's histogram for the distribution of flesh tones, if it's happening later in the day it's most likely human intervention. It's easy to quickly scan a large page of thumbnails and spot nudity, so I don't see why it would be out of the question for them to have a staffer doing that. They probably do not want their free service used by porn sites simply because if they allowed it their service would probably be swamped with it and become overloaded.
posted by Rhomboid at 12:43 PM on December 3, 2007

Response by poster: some further info on my nude resizer. I no longer much use it and have others that cause no problem of deletion. My site has many nudes--very tasteful ones I would not call porn--and I get some 5 thousand hits per day, so no one would call a company to report me. I continue to use the resizer (it is on the desktop and handy) but I avoid nudes. Use another instead for the nudes and for a few years now, no problem with deletions.
posted by Postroad at 2:01 PM on December 3, 2007

You may be interested in this: Detection Software & SDK
posted by mcarthey at 4:05 PM on December 3, 2007

Response by poster: the software mentioned just above this seems what might be at work. I had thought short time ago that since the image is taken from a source, the resizing outfit might determine by the html of the site where photo appears that it focuses upon nudes, therefore they delete based on point of origin of photo. I could test by taking a nice piece of artwork, classic stuff, that has breasts showing and see if my guess works, but since I have a fine place for resizing without fear of deletion, I will not fuss over the issue. Thanks to al for answers.
posted by Postroad at 5:26 PM on December 3, 2007

Questions I've always wanted to ask, I hope you don't mind -- do these nakedness-recognising programs successfully recognise the skin tones of all races? And what are some things which would give false positives?
posted by AmbroseChapel at 5:57 PM on December 3, 2007

ambrosechapel: The one I worked on was pretty good at recognizing different skin tones. One of the QA test cases was that we had to run it against a wide variety of porn and non-porn, and let's just say that we made sure that we had a fairly representative sample when we were compiling our test set.

2 weaknesses that we did notice: The engine was significantly weaker with comic-book type porn (hentai), and also almost blind on the first QA pass to black-and-white pictures (usually, pics that we desaturated in Photoshop).

BTW - getting written permission from our company's HR department to surf for porn at the office was definitely one of the highlights of my career.
posted by deadmessenger at 7:17 PM on December 3, 2007 [1 favorite]

Dan @ Dan's Data did a review of one of these pieces of software some time ago and found that recognition was dubious at best. Here is the link: PornSweeper

(Dan is also a MeFi member, BTW)
posted by ranglin at 8:50 PM on December 3, 2007