GRASS GIS manual: i.superpixels.slic

NAME

i.superpixels.slic - Perform image segmentation using the SLIC segmentation method.

KEYWORDS

imagery, segmentation, superpixels, SLIC

SYNOPSIS

i.superpixels.slic

i.superpixels.slic --help

i.superpixels.slic [-n] input=name[,name,...] output=name [iterations=integer] [num_pixels=integer] [step=integer] [perturb=integer] [compactness=float] [minsize=integer] [memory=integer] [--overwrite] [--help] [--verbose] [--quiet] [--ui]

Flags:

-n: Normalize spectral distances; Equvivalent to SLIC zero (SLIC0)
--overwrite: Allow output files to overwrite existing files
--help: Print usage summary
--verbose: Verbose module output
--quiet: Quiet module output
--ui: Force launching GUI dialog

Parameters:

input=name[,name,...] [required]: Name of two or more input raster maps or imagery group
output=name [required]: Name for output raster map
iterations=integer: Maximum number of iterations; Default: 10
num_pixels=integer: Approximate number of output super pixels; Default: 200
step=integer: Distance (number of cells) between initial super pixel centers; A step size > 0 overrides the number of super pixels; Default: 0
perturb=integer: Perturb initial super pixel centers; Percent of intitial superpixel radius; Options: 0-100; Default: 0
compactness=float: Compactness; A larger value causes more compact superpixels; Default: 1
minsize=integer: Minimum superpixel size; Default: 1
memory=integer: Memory in MB; Default: 300

DESCRIPTION
NOTES
EXAMPLES
- Segmentation of Landsat images and NDVI
REFERENCES
SEE ALSO
AUTHORS

DESCRIPTION

i.superpixels.slic performs superpixel segmentation using a k means method, based on the work of Achanta et al. The number of superpixels is determined either with the num_pixels option (number of superpixels) or with the step option (distance between initial super pixel centers).

The compactness option controls the compactness of the resulting superpixels. A larger compactness value will cause spatially more compact, but spectrally more heterogeneous superpixels. This is the most important parameter of the module. A reasonable value should be determined for small test regions before applying the module to a large region.

The resultant number of superpixels will most often be larger than the initial number of superpixels because the initial number of superpixels is used to create seeds and SLIC assigns pixels to seeds. Pixels assigned to the same seed are usually not connected. The final number of superpixels is the number of clumps, also known as connected components, objects, regions, or blobs. The final number of superpixels can be reduced with the minsize option.

NOTES

Input bands

Contrary to the original Achanta et al. SLIC algorithm which allows only RGB input images (which are internally transformed into LAB color space, i.superpixels.slic allows the use of any number of input bands. These bands can be either spectral bands of imagery, or any other pseudo-bands relevant to the analysis at hand (NDVI, texture, precipitation, etc). All band values are normalized to a common 0-1 scale to ensure comparability in the spectral distance calculations. Therefore results will not be identical with the original implementation.

Iterations

In each iteration, the assignment of pixels to superpixels and the superpixels themselves (spatial and spectral means) are updated until either the maximum number of iterations is reached or until the superpixels no longer change. More iterations will provide better results but will increase processing time. The module will stop earlier if convergence is reached, but it can take 1000 or more iterations until convergence is reached.

Minimum segment size

If a minimum segment size of 2 or larger is given with the minsize parameter, segments with a smaller pixel count will be merged with their most similar neighbor.

In the original implementation of Achanta et al., a minimum segment size is internally determined, and segments smaller than this minimum segment size are merged with an arbitrarily chosen neigbour. Therefore results will not be identical with the original implementation, even if a minimum segment size identical to the minimum segment size internally determined in the original implementation of Achanta et al. is used.

Creating seeds for i.segment

If the purpose is to create seeds for i.segment, a small number of iterations (at least 10) should be sufficient. Further on, a large number of superpixels or a small step should be used, and small clumps should not be merged.

Image segmentation

If the purpose is to perform image segmentation with i.superpixels.slic, a larger number of iterations (e.g. 100) should be used in order to obtain more stable superpixels. In this case, larger superpixels can be used and small clumps can be removed with the minsize option.

Normalization of spectral distances (SLIC0)

If the -n flag is used, the spectral distance of a pixel to a given superpixel is divided by the maximum previously observed spectral distance to that superpixel. This is an adaptation of the so-called SLIC0 (SLIC zero) method.

After each iteration, the largest spectral distance to a superpixel is determined from all pixels assigned to that superpixel. In the next iteration, pixel assignment to superpixels is updated and spectral distances of pixels to superpixels are divided by the largest spectral distance of the current superpixel when evaluating a potential assignment of a pixel to a superpixel.

Contrary to the Achanta et al. version of SLIC0, i.superpixels.slic takes into account the compactness value chosen by the user even when the -n flag is used.

SLIC0 implies that more heterogeneous superpixels have a larger maximum spectral distance. For a given pixel, the normalized spectral distance will be smaller for a more heterogeneous superpixel than for a more homogeneous superpixel. This favours more heterogeneous superpixels which can steal pixels from more homogeneous superpixels even if the not normalized spectral distance of a pixel to a homogeneous superpixel is smaller than to a heterogeneous pixel. As a consequence, heterogeneous superpixels can become larger and and even more heterogeneous. This effect becomes stronger with larger differences in the spectral homogeneity of neighboring superpixels, and with a lower compactness value, as spectral difference then gets a bigger weight.

Perturbing initial superpixel centers

Initial superpixel centers can be optimized with the perturb option. The objective of this optimization of initial superpixel centers is to create more distinct and more homogeneous superpixels. Superpixel centers are shifted to more uniform areas, the pixel with the smallest gradient. The module guarantees that no two superpixel centers are shifted to the same position. The perturb option is interpreted as percent of the maximum allowable shift distance such that no two superpixel centers can obtain the same position.

Memory vs disk cache

i.superpixels.slic can handle very large amounts of data. Depending on the amount of data and the available RAM memory, as defined by the memory parameter, the module will either work with memory cache, storing everything in memory, or with a disk cache, storing elements on disk and only retrieving the data when necessary. Disk cache is slower, but the data can be much larger than the RAM memory can hold. By default, the memory parameter is set fairly low for modern computer systems (500MB). Users should thus make sure to adjust the value to their system.

EXAMPLES

Segmentation of Landsat images and NDVI

List Landsat imagery in the full NC sample dataset:

g.list type=raster pattern='lsat*' sep=comma mapset=PERMANENT

Set the computation region to one of the rasters (all have the same extent and resolution):

g.region raster=lsat7_2002_10

Use the list to create an imagery group:

i.group group=lsat subgroup=lsat input=`g.list type=raster pattern='lsat*' sep=comma mapset=PERMANENT`

Perform the segmentation:

i.superpixels.slic group=lsat output=segments num_pixels=2000

Convert the segments to vectors for further analysis and visualization:

r.to.vect input=segments output=segments type=area

Show the boundaries between the segments with false color image in the background:

d.rgb red=lsat7_2002_70 green=lsat7_2002_50 blue=lsat7_2002_30
d.vect map=segments fill_color=none

Let's compute the NDVI using r.mapcalc and assign average NDVI to each of the vector areas:

i.vi red=lsat7_2002_30 nir=lsat7_2002_40 viname=ndvi output=ndvi
v.rast.stats map=segments raster=ndvi column_prefix=ndvi method=average

And now visualize the segments using the NDVI value for coloring the areas with the ndvi color table but keeping the area boundaries black:

g.copy vector=segments,segments_color
v.colors map=segments_color use=attr column=ndvi_average color=ndvi
d.vect map=segments_color width=2 icon=basic/point
d.vect map=segments fill_color=none

Average NDVI in each superpixel area (segment)

REFERENCES

SLIC Superpixels. 2010. Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Susstrunk. EPFL Technical Report no. 149300.
SLIC Superpixels Compared to State-of-the-art Superpixel Methods. 2012. Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(11), 2274 - 2282.
SLIC(0) website

AUTHORS

Rashad Kanavath, India
Markus Metz

Last changed: $Date: 2017-07-22 13:34:36 +0200 (Sat, 22 Jul 2017) $

SOURCE CODE

Available at: i.superpixels.slic source code (history)