Universitaet Koblenz-Landau

Universitaet Koblenz-Landau Fachbereich Informatik
Image Recognition Laboratory
Image Recognition Laboratory Download Page
Ansprechpartner (Email, Telefon)
Dokumentenbestand dieses Servers durchsuchen
Informationein und Hilfe zu diesem Server
 

Manual: Recursive Histogram Splitting

The Algorithm

Recursive histogram splitting is also known as recursive region splitting or recursive splitting.
(original algorithm: R. Ohlander, K. Price, and D. R. Reddy. Picture segmentation using a recursive region splitting method. Computer Graphics and Image Processing, 8:313-333, 1978)
The algorithm splits recursively the picture into regions based on the biggest peak which is found in some histograms.
The biggest peak is searched with 1d-graph theoretical clustering, a technique based on 2d-graph theoretical clustering (J. Matas, Colour-based Object Recognition, University of Surrey)
For each region the algorithm analyzes 9 histograms in the rgb, hsv and yiq colorspace. Detailed information can be found in the online lecture (lectures.ps, lectures.pdf)

A detailed description how the algorithm works:
Stabile, echtzeitfaehige Farbbildverarbeitung,
Volker Rehrmann, Foelbach 1994, page 40ff

Author: Dirk Balthasar, 2002, Universitšt Koblenz-Landau

Description

The zip archives segmentation-linux-1-0.zip and segmentation-windows-1-0.zip contain each the binary version of the command line tool segmentation. With the command line tool segmentation you are able to make segmentations of color images with the recursive histogram splitting and split and merge algorithm. The 'segmentation' needs as input a color image in PPM-Format (Portable Pixmap in binary format with magic number 'P6'). It produces two images: a region image that stores for each pixel the average color of its region and a labeled image that stores for each pixel a label number of its region.

Usage

Syntax: segmentation [options] rhs "name_of_input_image" "name_of_region_image" "name_of_labeled_image"

Options Reference:

-gt [0/127]: Windowsize for graph theoretical clustering of histograms. 1 means that local maxima are searched with a windowsize of 3. (default value LeftRightOffset = 3) Higher values result the coarser granularity of the peaks.
-ms MinSplitPeakSize:If the biggest peak that is found in all histograms for one region is smaller than MinSplitPeakSize the region is not splitted any more. (default MinSplitPeakSize = 20) Higher values result bigger regions, 0 disables this option.
-mr MinSplitTotalSize:If a region is smaller than MinSplitTotalSize it will not be split. Must be bigger than MinSplitPeakSize to have any effect. (default MinSplitTotalSize = 30) Higher values result bigger regions, 0 disables this option.


Example:

Suppose you want to make RHS segmentation of an color image with name example.ppm. First be sure that your color image is in PPM graphics format and the PPM file has the magic number P6. The command
segmentation -gt 3 -ms 20 -mr 30 smg example.ppm region.ppm region.pgm
will produce a region image with name region.ppm containing the mean color of each region in the ppm format and a label image with the name region.pgm.




Editor: Dirk Balthasar
Letzte Änderung: 15.07.2002