Similar image search software (e.g. TinEye)

Question

On one of our community sites, we allow users to upload images. These images have been approved or rejected by our moderators.

In order to limit the work required by our administrators, we want to “register” each image that is rejected in some database, and search this database before sending the image for approval. If such an image has already been rejected, the uploaded image will not be sent for approval.

We can, of course, just record things like file name, size and MD5 images for similarity, but we could definitely find cropped or resized images differently.

TinEye.com provides similar functionality.

Do you know any open source software capable of this? Do you have any other ideas?

Thanks!

Answer 1

To detect images with reduced sizes and with loss of images, you can resize the image to some standard size (for example, 40x40 pixels), and then subtract the known image from the new image and compare the distance to the threshold.

Unfortunately, this does not work with rotation or cropping. In this case, you need to extract the large-scale invariant image functions.

Another problem with this approach is that with a naive implementation, the computational cost is linear in size of the list of known images, so it may be too expensive to compare a new image with all old images.

Answer 2

You can create a list with "similar images" if they do not guarantee 100% similarity. The similarity could be calculated taking into account the imprint of the image (as Winner said, you can scale it to a standard size and build a checksum). Can also be used "medium" color, as well as "color variation".

Based on this, you can display the list of "similar images" (with a mouse click) in the administrator, sorted in the order "most likely similar"

You can also see the image :: Compare http://linux.softpedia.com/get/Programming/Widgets/Perl-Modules/Image-Compare-43727.shtml and jpegDiff http://www.marengo-ltd.com/open_source /index.php

Answer 3

I developed a service similar to TinEYE for private image collections. He currently works in alpha testing, but works great for 50,000 images of my clients ... You can even identify printed images in electronic documents :)

Try http://example.askbubo.com (only looking for images of my clients)

At http://askbubo.com/ you can read more. Also feel free to contact me directly at hajo (at) spratpix (dot) com.

While I plan to charge a fee for the service later, I will gladly give it to the people who participated in the alpha testing and gave a good feedback.