search

HITS-AIN / PINK

Parallelized rotation and flipping INvariant Kohonen maps
GNU General Public License v3.0
21 stars 11 forks source link

Use circular rotation region for the euclidean distance #55

Closed BerndDoser closed 4 years ago

BerndDoser commented 4 years ago

Close #46

tjgalvin commented 4 years ago

I've manage to give this a go on some images of mine and the circular euclidean distance region appears to be working correctly now for images with a single channel. That last commit that fixed the offsets and deltas did the trick. I have not tried this yet for multi-channel images but I think it would work. I will let you know if I run into any problems.

For reference attached is a quick figure of a single neuron. On the left is the circular ED region. On the right is the entire neuron that is 2**0.5 larger than the input images and circular ED region. You can see for the image on the left we have gotten rid of the super-imposed square type pattern that is there when we use a Cartesian ED region shape.

Thank you @BerndDoser for the great new feature!

Screenshot 2020-04-11 20 49 48

BerndDoser commented 4 years ago

Thank you @tjgalvin for the great help. I am also very happy that we have found such a clean solution for the rotational euclidean space. @tjgalvin has prepared the precalculation of the number of pixels in a circle row, which allows us the usage of a memory mapping procedure and shows an excellent performance.