https://bugs.kde.org/show_bug.cgi?id=452878
Bug ID: 452878
Summary: Feature to search for images by color
Product: digikam
Version: 7.5.0
Platform: Microsoft Windows
OS: Microsoft Windows
Status: REPORTED
Severity: wishlist
Priority: NOR
Component: Searches-Similarity
Assignee: digikam-bugs-n...@kde.org
Reporter: nicolas.bris...@airbus.com
Target Milestone: ---
STEPS TO REPRODUCE
1. I was asked to find some images by color in my 100k+ image collection, in
order to organize some decoration based on the dominant color tone
2. I looked around all of digikam and I could not find a "search by color" tool
:-(
3. I tried to approach the result by using the similarity search and drawing a
sketch consisting of a plain color patch
OBSERVED RESULT
I get a very limited list of images (default: 10 if I remember correctly, which
I can only increase to 50), which are mostly in the right tone - but not even
quite uniform. I guess that's due to the algo used, as discussed in bug
#207188: it isn't actually meant for what I'm trying to do...
EXPECTED RESULT
I would expect to find all (or at least a much higher number than 50) images
from my collection that are "close enough" to the selected color
SOFTWARE/OS VERSIONS
Windows: 10
(since it's a feature request it actually applies to all OSes)
ADDITIONAL INFORMATION
I looked at some sites providing something like this feature, and found some
hopefully relevant ideas:
- A blog at https://medium.com/@joshuawcomeau/search-by-colour-cb1ba49aa9aa
describing some experiments in order to determine the "relevant" color, ie the
one that actually sticks out and not necessarily the dominant background color.
Not quite what I'm after, but it can also be an interesting alternative and
probably some interesting ideas in terms of preprocessing
- An article at https://reasonabledeviations.com/2019/10/02/k-means-in-cpp/
describing k-means, an algorithm which I happen to have used recently (from
python / scikit though) and which sounds very appropriate for the need at hand
- A website that provides more or less exactly what I'm looking for (albeit
neither open-source nor applicable to my personal collection...) at
https://labs.tineye.com/multicolr/#colors=f6d727;weights=100; and which has
some hints on how it works on this page:
https://services.tineye.com/MulticolorEngine - see in particular the "What are
the major functions supported by the API?" section
What I imagine would be along the lines:
1. Add a tool to compute "color fingerprints", similar to the existing
fingerprinting for similarity search. This tool would compute the n (fixed or
varying, to be defined - but probably around 5 to 10 at most) dominant ucolors
using k-means and store their relative frequency if possible. Some
preprocessing may be applied (color paletization, downsizing, ...) if it helps
for performance reasons
2. A "Color" tab could be added to similarity search along Duplicate / Image /
Sketch, in which the user could select a few colors with their percentage (see
https://labs.tineye.com/multicolr/#colors=f6d727;weights=100; for inspiration)
3. Some options may be added, like maximum tolerance or maximum number of hits
to return. Why not also an option to use the first algo linked above, to focus
more on the colors that stick out than the most frequent one(s)
4. Once the colors and options are set, a "simple" (euclidian or better:
perceptual) distance search on the fingerprints is performed and the found
images are listed as for the other similarity search tools
--
You are receiving this mail because:
You are watching all bug changes.
