Too much false positives
scratcher28 opened this issue · comments
Tested it with a lot of images, in fact, too much false positives.
The most of test images are flagged as nude.
A lot of really nude images are flagged as normal.
Can you provide us the images you used to test it? Besides that, this is a
proof of concept made only to give a speech at a security conference.
On Mon, Oct 7, 2013 at 8:07 AM, scratcher28 notifications@github.comwrote:
Tested it with a lot of images, in fact, too much false positives.
The most of test images are flagged as nude.
A lot of really nude images are flagged as normal.—
Reply to this email directly or view it on GitHubhttps://github.com//issues/1
.
Atte:
Matías Aereal Aeón
As @mattaereal mentioned, this is a proof of concept for a security conference. We just wanted to have something running ASAP to be able to demonstrate that stealing pron automatically in a massive way is possible.
Besides that, nude detection is a very active research field, and there is lots of material out there to improve the accuracy of this script. Some ideas that come to my mind at the moment:
- Improve the skin color detection algorithm. We use a very basic one.
- Combine the algorithm with other alternatives, and tune some parameters.
- Use a feature-detection based algorithm alongside with this.
- Take the file names and metadata into consideration for classifying the image/video as porn or not.
I hope this helps you. We may need to explain this better in the README.