Example Superpixels

• Example of image segmentation on a standard test image using Felzenszwalb-Huttenlocher
• 

  Input image.

• 

  Region borders highlighted in red.

• 

  Regions are assigned their average color.

Image segmentation is an important (and very much unsolved) problem in computer vision. In this example, different techniques are used to break the image up into regions (or superpixels). The goal of this segmentation is to simplify the image's description, which can then be used for object detection/recognition.

Example Code:

• ExampleSegmentSuperpixels.java

Concepts:

• Image Segmentation
• Super Pixels

Related Examples:

• Color Segmentation
• Watershed with Seeds
• Thresholding
• Tutorial Image Segmentation

Relevant Applets:

• Image Segmentation

Example Code

/**
 * Example demonstrating high level image segmentation interface.  An image segmented using this
 * interface will have each pixel assigned a unique label from 0 to N-1, where N is the number of regions.
 * All pixels which belong to the same region are connected.  These regions are also known as superpixels.
 *
 * @author Peter Abeles
 */
public class ExampleSegmentSuperpixels {

	/**
	 * Segments and visualizes the image
	 */
	public static <T extends ImageBase>
	void performSegmentation( ImageSuperpixels<T> alg , T color )
	{
		// Segmentation often works better after blurring the image.  Reduces high frequency image components which
		// can cause over segmentation
		GBlurImageOps.gaussian(color, color, 0.5, -1, null);

		// Storage for segmented image.  Each pixel will be assigned a label from 0 to N-1, where N is the number
		// of segments in the image
		GrayS32 pixelToSegment = new GrayS32(color.width,color.height);

		// Segmentation magic happens here
		alg.segment(color,pixelToSegment);

		// Displays the results
		visualize(pixelToSegment,color,alg.getTotalSuperpixels());
	}

	/**
	 * Visualizes results three ways.  1) Colorized segmented image where each region is given a random color.
	 * 2) Each pixel is assigned the mean color through out the region. 3) Black pixels represent the border
	 * between regions.
	 */
	public static <T extends ImageBase>
	void visualize(GrayS32 pixelToRegion , T color , int numSegments  )
	{
		// Computes the mean color inside each region
		ImageType<T> type = color.getImageType();
		ComputeRegionMeanColor<T> colorize = FactorySegmentationAlg.regionMeanColor(type);

		FastQueue<float[]> segmentColor = new ColorQueue_F32(type.getNumBands());
		segmentColor.resize(numSegments);

		GrowQueue_I32 regionMemberCount = new GrowQueue_I32();
		regionMemberCount.resize(numSegments);

		ImageSegmentationOps.countRegionPixels(pixelToRegion, numSegments, regionMemberCount.data);
		colorize.process(color,pixelToRegion,regionMemberCount,segmentColor);

		// Draw each region using their average color
		BufferedImage outColor = VisualizeRegions.regionsColor(pixelToRegion,segmentColor,null);
		// Draw each region by assigning it a random color
		BufferedImage outSegments = VisualizeRegions.regions(pixelToRegion, numSegments, null);

		// Make region edges appear red
		BufferedImage outBorder = new BufferedImage(color.width,color.height,BufferedImage.TYPE_INT_RGB);
		ConvertBufferedImage.convertTo(color, outBorder, true);
		VisualizeRegions.regionBorders(pixelToRegion,0xFF0000,outBorder);

		// Show the visualization results
		ListDisplayPanel gui = new ListDisplayPanel();
		gui.addImage(outColor,"Color of Segments");
		gui.addImage(outBorder, "Region Borders");
		gui.addImage(outSegments, "Regions");
		ShowImages.showWindow(gui,"Superpixels", true);
	}

	public static void main(String[] args) {
		BufferedImage image = UtilImageIO.loadImage(UtilIO.pathExample("segment/berkeley_horses.jpg"));
//		BufferedImage image = UtilImageIO.loadImage(UtilIO.pathExample("segment/berkeley_kangaroo.jpg"));
//		BufferedImage image = UtilImageIO.loadImage(UtilIO.pathExample("segment/berkeley_man.jpg"));
//		BufferedImage image = UtilImageIO.loadImage(UtilIO.pathExample("segment/mountain_pines_people.jpg"));
//		BufferedImage image = UtilImageIO.loadImage(UtilIO.pathExample("particles01.jpg"));

		// you probably don't want to segment along the image's alpha channel and the code below assumes 3 channels
		image = ConvertBufferedImage.stripAlphaChannel(image);

		// Select input image type.  Some algorithms behave different depending on image type
		ImageType<Planar<GrayF32>> imageType = ImageType.pl(3, GrayF32.class);
//		ImageType<Planar<GrayU8>> imageType = ImageType.pl(3,GrayU8.class);
//		ImageType<GrayF32> imageType = ImageType.single(GrayF32.class);
//		ImageType<GrayU8> imageType = ImageType.single(GrayU8.class);

//		ImageSuperpixels alg = FactoryImageSegmentation.meanShift(null, imageType);
//		ImageSuperpixels alg = FactoryImageSegmentation.slic(new ConfigSlic(400), imageType);
		ImageSuperpixels alg = FactoryImageSegmentation.fh04(new ConfigFh04(100,30), imageType);
//		ImageSuperpixels alg = FactoryImageSegmentation.watershed(null,imageType);

		// Convert image into BoofCV format
		ImageBase color = imageType.createImage(image.getWidth(),image.getHeight());
		ConvertBufferedImage.convertFrom(image, color, true);

		// Segment and display results
		performSegmentation(alg,color);
	}
}