/hydrus/ - Hydrus Network

>>10338

Hydrus is using ImageHashes currently and the Levenshtein distances between those to determine image similarity. This is already in and usable via the special page "duplicates processing".

By saying you did some research, did you just google "image matching" and create a huge file with links to sources? Do you seriously expect the developer to take a look at all of these? I don't know man, if I'm not misunderstanding you, this isn't research at all.

>>10344

I did some basic research on the principles behind SIFT/SURF and BRISK/ORB, which are using a radically different approach compared to image hash. That pad is just a larger reference list that we can explore further.

SIFT is just an image pyramid (image hierarchy with progressively lower resolution) with points of focus (high contrast and/or corners) combined with affine transformation medallions for those points

https://opencv-python-tutroals.readthedocs.io/en/latest/py_tutorials/py_feature2d/py_sift_intro/py_sift_intro.html

SURF is just SIFT, but with more box filters for black-and-white contrasts (basically wavelets are JPEG compression like systems for it) to save more time from traditional SIFT

https://opencv-python-tutroals.readthedocs.io/en/latest/py_tutorials/py_feature2d/py_surf_intro/py_surf_intro.html

BRIEF takes a different approach than SIFT and SURF, where instead of finding edge markers and use them as a base, it uses random straight lines to get the right pattern (which outperforms regular patterns)

https://opencv-python-tutroals.readthedocs.io/en/latest/py_tutorials/py_feature2d/py_brief/py_brief.html

ORB is a blend of FAST and BRIEF rolled into one, which beats the other three by a long shot (but there are more methods on the reading list that are more recent)

https://opencv-python-tutroals.readthedocs.io/en/latest/py_tutorials/py_feature2d/py_orb/py_orb.html

There are others like FAST, BRISK, FREAK, KAZE and AKAZE that are less common, and people are experimenting with other methods.

If you want something that absolutely works try https://github.com/pippy360/transformationInvariantImageSearch which is pretty easy to understand (albeit expected to be slower than modern methods)

>>10345

Also https://vision.fe.uni-lj.si/cvww2016/proceedings/papers/04.pdf (Quantitative Comparison of Feature Matchers Implemented in OpenCV3) and

https://sci-hub.tw/10.1109/m2vip.2016.7827292 (Comparison of OpenCV’s Feature Detectors and Feature Matchers)

All the algorithms listed are in OpenCV, so there is nothing to lose by implementing some of these.

>>10345

>>10346

Alright got it. It seemed at first just like a huge link dump and "hey I did some research". The hint about all this already being available in OpenCV is great. Sorry for reacting like I did then.

This information is useful and interesting, but I unfortunately do not have time to go through it in any clever way. I can't work on big things like new duplicate finding algorithms in normal weekly work–it'll have to wait for the next iteration of the duplicate system. I'd also like to prioritise the duplicate processing workflow, which is the larger problem with the system right now.

I am most interested in face and other feature detection in cv. I expect to use some of this stuff when we push towards machine learning auto-tagging.

>>10348

For the dev: https://medium.com/machine-learning-world/feature-extraction-and-similar-image-search-with-opencv-for-newbies-3c59796bf774

https://github.com/creotiv/computer_vision

>>10338

In case of ghostbin closing down https://pastebin.com/KPaYiXNM