Labeled regions module

#include "diplib/regions.h"

dip::uint dip::Label(dip::Image const& binary, dip::Image& out, dip::uint connectivity = 0, dip::uint minSize = 0, dip::uint maxSize = 0, dip::StringArray boundaryCondition = {}, dip::String const& mode = S::ALL)

Labels the connected components in a binary image

The output is an unsigned integer image. Each object (respecting the connectivity, see Connectivity) in the input image receives a unique number. This number ranges from 1 to the number of objects in the image. The pixels in the output image corresponding to a given object are set to this number (label). The remaining pixels in the output image are set to 0.

The minSize and maxSize set limits on the size of the objects: Objects smaller than minSize or larger than maxSize do not receive a label and the corresponding pixels in the output image are set to zero. Setting either to zero disables the corresponding check. Setting both to zero causes all objects to be labeled, irrespective of size.

The boundaryCondition array contains a boundary condition string per image dimension, or one string to be used for all dimensions. Valid strings are:

• "" and "mirror": the default behavior, causing the labeling to simply stop at the edges.
• "periodic": imposing a periodic boundary condition, such that objects touching opposite edges of the image are considered the same object.
• "remove": causing objects that touch the image edge to be removed.

boundaryCondition can also be an empty array, using the default behavior for all dimensions.

mode can be "all" (default) or "largest". If set to "largest", only the largest object is retained, and will have a label of 1.

Returns the number of connected components found. The returned value is thus the maximum value in the output image.

dip::UnsignedArray dip::GetObjectLabels(dip::Image const& label, dip::Image const& mask = {}, dip::String const& background = S::EXCLUDE)

Gets a list of object labels in the labeled image. A labeled image must be of an unsigned type.

If background is "include", the label ID 0 will be included in the result if present in the image. Otherwise, background is "exclude", and the label ID 0 will be ignored.

void dip::Relabel(dip::Image const& label, dip::Image& out)

Re-assigns labels to objects in a labeled image, such that all labels are consecutive.

Note that disjoint objects will remain disjoint, as this function only replaces each label ID with a new value. The output image will have consecutive label IDs, in the range [1, N], with N the number of unique labels in label. Pixels with a value of 0 will remain 0 (background).

dip::GetObjectLabels returns a list of unique labels in label, and can be used to determine N.

void dip::Relabel(dip::Image const& label, dip::Image& out, dip::Graph const& graph)

Re-assigns labels to objects in a labeled image, such that regions joined by an edge in graph obtain the same label.

graph should be obtained through dip::RegionAdjacencyGraph and modified to obtain a useful segmentation. For example:

dip::Image input = ...
dip::Image label = dip::Watershed( dip::GradientMagnitude( input, { 2 } ), {}, 2, 1, 0, { "labels" } );
dip::MeasurementTool measurementTool;
auto msr = measurementTool.Measure( label, input, { "Mean" } );
dip::Graph graph = RegionAdjacencyGraph( label, msr[ "Mean" ], "watershed" );
graph = graph.MinimumSpanningForest( { 1 } ); // make sure we don't use the unconnected vertex 0 as root.
graph.RemoveLargestEdges( 100 );
dip::Relabel( label, label, graph );

void dip::SmallObjectsRemove(dip::Image const& in, dip::Image& out, dip::uint threshold, dip::uint connectivity = 0)

Removes small objects from a labeled or binary image.

If in is an unsigned integer image, it is assumed to be a labeled image. The size of the objects are measured using dip::MeasurementTool, and the labels for the objects with fewer than threshold pixels are removed. The connectivity parameter is ignored. Note that if this image contains disjoint objects (i.e. multiple connected components with the same label), it is the size of the object as a whole that counts, not the size of individual connected components.

If in is a binary image, dip::Label is called with minSize set to threshold, and the result is binarized again. connectivity is passed to the labeling function.

The operation on a binary image is equivalent to an area opening with parameter threshold (see dip::AreaOpening). The same is not true for the labeled image case, if labeled regions are touching or if objects are disjoint.

void dip::GrowRegions(dip::Image const& label, dip::Image const& mask, dip::Image& out, dip::sint connectivity = -1, dip::uint iterations = 0)

Grow (dilate) labeled regions uniformly.

The regions in the labeled image label are dilated iterations steps, according to connectivity, and optionally constrained by mask. If iterations is 0, the objects are dilated until no further change is possible.

If a mask is given, this is the labeled equivalent to dip::BinaryPropagation, otherwise it works as dip::BinaryDilation on each label. The difference between dip::GrowRegions and dip::Dilation (which can also be applied to a labeled image) is that here growing stops when different labels meet, whereas in a normal dilation, the label with the larger value would grow over the one with the smaller value.

The connectivity parameter defines the metric, that is, the shape of the structuring element (see Connectivity). Alternating connectivity is only implemented for 2D and 3D images.

If isotropy in the dilation is very important, compute the distance transform of the background component, then apply dip::SeededWatershed.

void dip::GrowRegionsWeighted(dip::Image const& label, dip::Image const& grey, dip::Image const& mask, dip::Image& out, dip::Metric const& metric = {S::CHAMFER,2})

Grow labeled regions with a speed function given by a grey-value image.

The regions in the input image label are grown according to a grey-weighted distance metric; the weights are given by grey. The optional mask image mask limits the growing. All three images must be scalar. label must be of an unsigned integer type, and grey must be real-valued.

out is of the type dip::DT_LABEL, and contains the grown regions.

Non-isotropic sampling is supported through metric, which assumes isotropic sampling by default. See dip::GreyWeightedDistanceTransform for more information on how the grey-weighted distance is computed.

void dip::SplitRegions(dip::Image const& label, dip::Image& out, dip::uint connectivity = 0)

Ensures a gap between regions with unequal labels.

In the output image, no two regions will be connected according to connectivity. out is of the same type as label, which must be an unsigned integer type, and scalar.

To create a one-pixel gap between regions, regions must shrink, and they must shrink unequally. If all regions were to shrink equally, we would create a two-pixel gap. This function chooses to be biased towards larger-valued labels: where two objects touch, the lower-valued region shrinks.

This function works by finding pixels that have a neighbor with a larger value, and setting these pixels to zero (the background label).

void dip::MakeRegionsConvex2D(dip::Image const& label, dip::Image& out, dip::String const& mode = S::FILLED)

Make each object a single, convex shape.

For each label, it finds the convex hull and draws it into the image. Where the convex hulls of objects overlaps, the larger label prevails, meaning that the other object might not be convex any more where the larger label overlaps it.

mode can be "filled" (the default) or "hollow". With "hollow", the convex hull outlines are drawn into an empty image, such that overlapping convex hulls can be visually identified. With "filled", the operation is increasing (like a morphological closing), meaning that out >= label is guaranteed.

Note that defining a convex hull in a discrete space is ambiguous, and therefore this function is not idempotent (i.e. running it a second time on its own output can produce a different result).

If the input image is binary, a single convex hull is computed and drawn. The output image will be binary also.

The image must have two dimensions, and be scalar.

dip::RangeArray dip::GetLabelBoundingBox(dip::Image const& label, dip::uint objectID)

Returns the bounding box for all pixels with label objectID in the labeled or binary image label.

label must be a labeled image (of type unsigned integer) or binary, and must be scalar and have at least one dimension.

In the case of a binary image, objectID should be 1 or 0, no other values are possible in a binary image.

When no pixels with the value objectID exist in the image, the output dip::RangeArray will be an empty array. Otherwise, it will have as many dip::Range elements as dimensions are in the image.

To obtain the bounding box of all labels at once, use dip::MeasurementTool::Measure with the features Minimum and Maximum.

dip::Graph dip::RegionAdjacencyGraph(dip::Image const& labels, dip::String const& mode = "touching")

Construct a graph for the given labeled image.

Each region (object/label) in the image labels is a node. Edges represent neighborhood relations between regions. Because the dip::Graph class uses the vertex ID as an index into an array, it is recommended that this function be called with a labeled image where labels are more or less consecutive. If dip::Maximum( labels ).As< dip::uint >() (i.e. the largest label ID) is much larger than dip::GetObjectLabels( labels ).size() (i.e. the number of labels), then the dip::Graph object will have many unused vertices, and hence waste space. Use dip::Relabel to modify the labels image to have consecutive labels.

mode indicates how to construct the graph. It can be one of the following strings:

• "touching": two regions are neighbors if they have at least one pixel that is 1-connected to the other region. That is, the two regions directly touch.
• "watershed": two regions are neighbors if there is a background pixel that is 1-connected to the two regions. That is, the two regions are separated by a 1-pixel watershed line. Note that, in this case, two regions that directly touch will not be recognized as neighbors!

In both modes, region with ID 0 (the background) is not included in the graph. But note that there is a vertex with index 0, which will not be connected to any other vertex. To include the background, simply increment the label image by 1.

Edge weights are computed as follows: The fraction of boundary pixels for region 1 that connect to region 2 is determined. The fraction of boundary pixels for region 2 that connect to region 1 is determined. One minus the larger of these two fractions is the edge weight. Thus, edge weights have a value in the half-open interval [0,1). If one of the two regions has a very large fraction of its perimeter connected to another region, then the edge weight is very small to indicate a strong connection.

Vertex values are not assigned.

dip::Graph dip::RegionAdjacencyGraph(dip::Image const& labels, dip::Measurement::IteratorFeature const& featureValues, dip::String const& mode = "touching")

Construct a graph for the given labeled image.

This function is similar to the one above, but edge weights are derived from the absolute difference between featureValues for the two regions joined by the edge. Vertex values are set to the feature value for the region.

The input featureValues is a view over a specific feature in a dip::Measurement object. Only the first value of the feature is used. For features with multiple values, select a value using the dip::Measurement::IteratorFeature::Subset method, or pick a column in the dip::Measurement object directly using dip::Measurement::FeatureValuesView.

For the labels that do not appear in featureValues, their vertex value will be set to 0.