image segmentor emulating eCognition's Multiresolution segmentation source code and user guide

Attention: The documentation and code and executables of this project has been moved to [here](https://github.com/JzHuai0108/ImageSegmentor). 

 

Brief Introduction

This software has realized the followingmethods for segmentation:

 

        Graph based segmentation [2](GS);

        Quad tree segmentation [1] (QT)(similar to that implemented in eCognition [6] and that as a step in HSMR [7],but in a bottom-up fashion);

        Hierarchical Clusteringsegmentation (HC) (i.e., Multiresolution segmentation [3], but with multiplehomogeneity criterion, basic data structures detailed in [5]);

        The two-stage methods, GS-HC,QT-HC (described in [1] in detail)

        Pyramid mean shift segmentationimplemented in OpenCV 1.0

        Mean shift segmentation [8] ofDorin Comaniciu (resembles EDISON [9])

        Watershed segmentation [10] ofVincent and Soille

 

And for edge detection, SUSAN [11], Sobel,Canny edge detector are incorporated.

For classification,

        MRF classification initializedby K-means by ICM, Gibbs sampler, Metropolis etc.

        K-means clustering

For building hypotheses generation onpanchromatic images,

        Hierarchical region growing andthen apply threshold to region properties (i.e., decision tree),

        K-means clustering, connectedregion labeling and then decision tree,

For change detection,

        Rosin’s automatic thresholding[12],

        Gradient image correlationthresholding [13],

        window intensity correlationthresholding [14],

        Window histogram matching usingOpenCV.

        Multilevel parcel based change detectionusing K-means (resembles [15] but using k-means to decide threshold)

How to install?

Requirement: OpenCV 1.0 has been installedon your computer and the system variable path has been set. The GDAL library isused in this software to support for multiple image formats and various vectorformats, such as shp. And the needed source code for GDAL is already includedin the source code zip file.

(1) Download the source code. zip file fromthe CSDN (Search keywords image segmentor source code)

(2) Unzip and Open the .dsw file in VisualC++6.0.

(3) Compile and Debug.

If there is linking problems, you may haveto reset the project setting to link libraries for GDAL and OpenCV.

 

How to use?

The menu contains 4 items forpreprocessing, segmentation, classification and change detection, namely,preprocess, seg, classify and texture CD, respectively.

Normally, Open an image, click one commandin the menu, follow the popup dialogs (if available), the outcome will replace theoriginal image and display in the child window. And if want to make anotheroperation on the original image, use View>Origin to get back to the initialimage.

The following is a table of menu itemscorresponding to image processing functions.

 

commands	function description or requirement
preprocess
histogram equalize	take 256 gray scale image as input
edge detection	sobel	use the first channel of the image to detect edges
	susan	as above
	canny	as above
smooth by Gauss	use the first channel of the image
Gabor transform
Entropy
Opening Morph	open operation of morphology
seg
set parameters	set parameters for segmentation methods, if skipped, parameters take values by default
Quad Tree	Quad Tree segmentation
HierClust	multiresolution segmentation
GraphSeg	graph based segmentation
GraphSeg-HC	first Graph based segmentation, followed by hierarchical clustering, i.e.,region merging
QuadTree-HC	Quad Tree segmentation followed by hierarchical clustering
Preview	after any of the preceding 5 segmentation command, use this command to view region properties and histograms
Mean shift(bg-dzj)	mean shift proposed by Comaniciu
Watershed(Vincent)	Vincent-Soille's segmentation approach
Pyr Mean Shift(OpenCV)	pyramid mean shift implemented in OpenCV
classify
MRF classifier	apply to the first band of the input image
Color MRF	use the intensity of the image in the active window as input
Kmeans	Kmeans classifier to classify pixels in an image
Building Kmeans	Kmeans classification, connected region labeling and decision tree to get building hypotheses
Building	region growing (Hierarchical clustering) and decision tree to get building prototypes
Texture CD	change detection based on texture
Rosin CD	Rosin's automatic thresholding for change detection
Correlation CD	intensity correlation between windows centered on one pixel at two-time images and thresholding by 0.5 to detect changes
Grad Corr CD	gradient correlation between windows centered on one pixel at two-time images and thresholding for change detection
Histrogram Stat CD	match histograms of the window centered on one pixel of two times and threshold to detect changes
Multilevel Diff CD	multilevel parcel based feature extraction and k-means clustering to decide threshold for change detection, a little slower and the change detection result is saved in the same directory as the opened image

Note that there are additional commands nottabulated in the above table which are not complete or waiting for update.

Reliability Concerns

These functions have been tested for a fewtimes. But obviously they are premature and need to be rigorously tested, sosafety using this code is not guaranteed. From my empirical observations, thesegmentation methods, GS-HC, QT-HC and HC are most reliable among other commands.If there is any bug concerning the 3 functions, I would like to exchange views.

Reference

1.        YuminTan, Jianzhu Huai, Zhongshi Tang. AnObject-Oriented Remote Sensing Image Segmentation Approach Based on EdgeDetection. Spectroscopy and Spectral Analysis, vol. 30, no. 6, pp.1624-1627,June 2010.

2.        P. F. Felzenszwalb and D. P. Huttenlocher, Efficient Graph-BasedImage Segmentation, International Journalof Computer Vision, vol. 59 issue 2, Sep. 2004.

3.        MartinBaatz, Multi-resolution Segmentation: an optimization approach for high qualitymulti-scale image segmentation. 2002.

4.        KostasHaris, Serafim N. Efstratiadis,Nicos Maglaveras, and Aggelos K. Katsaggelos. HybridImage Segmentation Using Watersheds and Fast Region Merging. IEEE Transactions on Image Processing, Vol.7, No. 12, Dec 1998.

5.        DavidJ. Robinson, Nicholas J. Redding, David J. Crisp. Implementation of a FastAlgorithm for Segmentating SAR Imagery. Jan 2002.

6.        eCognitionuser guide 4.0. http://www.gis.unbc.ca/help/software/ecognition4/ELuserguide.pdf

7.        BenWuest, Yun Zhang. Region based segmentation of QuickBird multispectral imagerythrough band ratios and fuzzy comparison. ISPRS Journal of Photogrammetry andRemote Sensing 64 (2009) 55-64.

8.        DorinComaniciu and Peter Meer. Mean Shift: A Robust Approach Toward Feature SpaceAnalysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.24, no. 5, May 2002.

9.        EDISON user guide. http://coewww.rutgers.edu/riul/research/code/EDISON/doc/segm.html

10.     Lee Vincent and PierreSoille. Watersheds in digitalspaces: An efficient algorithm based on immersion simulations. IEEE PAMI, vol.13, no. 6, pp. 583--598, 1991.

11.     Stephen M. Smith and J. Michael Brady. SUSAN—ANew Approach to Low Level Image Processing. International Journal of ComputerVision 23(1), 45–78 (1997)

12.     Paul L Rosin. Thesholding for ChangeDetection. Proceedings of International Conference on Computer Vision, 1998.

13.     Liyuan Li,Maylor K. H.Leung. Integrating Intensity andTexture Differences for Robust Change Detection. IEEE Transactions on ImageProcessing, vol. 11, no. 2, Feb. 2002.

14.     Csaba Benedek, Xavier Descombes and Josiane Zerubia.Building Extraction and Change Detection in Multitemporal Aerial and SatelliteImages in a Joint Stochastic Approach. INRIA. 2009.

15.     F. Bovolo and L. Bruzzone, A multilevel parcel-basedapproach to change detection in very high resolution multitemporal images,Proc. IGARSS, vol. 3, pp. 2145–2148, 2005.