ncolor 1.5.3

label matrix coloring algorithm

ncolor

Fast remapping of instance labels 1,2,3,...,M to a smaller set of repeating, disjoint labels, 1,2,...,N. The four color theorem guarantees that at most four colors are required for any 2D segmentation/map, but the stochastic algorithms of ncolor will opt for 5 or 6 to give an acceptable result if it fails to find a 4-color mapping quickly. Also works for 3D labels (<8 colors typically required) and perhaps higher dimensions as well.

Usage

If you have an integer array called masks, you may transform it into an N-color representation as follows:

import ncolor 
ncolor_masks = ncolor.label(masks)

If you need the number of unique labels returned:

ncolor_masks, num_labels = ncolor.label(masks,return_n=True)

If you need to convert back to 0,...,N object labels:

labels = ncolor.format_labels(ncolor_masks,clean=True)

Note that format_labels with clean=True will also remove small labels (<9px) by default. This behavior can be changed with the min_area parameter.

The integer array ncolor_masks can be visualized using any color map you prefer. The example in this README uses the viridis colormap. See example.ipynb for more details.

Thanks to Ryan Peters (@ryanirl) for suggesting the expand_labels function. This is applied by default to 2D images (optionally for 3D images with expand=True, but this can give bad results since objects in 3D have a lot more wiggle room to make contact when expanded). This preprocessing step eliminates cases where close (but not touching) or dispersed objects previously received the same label. I dug a layer back to use ndimage.distance_transform_edt for a speed boost. If undesired for 2D images, use expand=False.