GISdevelopment.net ---> AARS ---> ACRS 1990 ---> Digital Image Processing

Simultaneous segmentation of multiband images

F. Cheevasuvit, W. Surakampontorn
Faculty of Engineering
King Mongkut's Institute of Technology Ladkrabang
Bangkok 10520, Thailand


Abstract
A method for simultaneous segmentations of a multi band images is described in this article .The purposed method is based on the application of graph theory to one of the multi band images .Then the segmented image from multi band is formed by agglomerating the vertices of the graph and simultaneously merging the corresponding pixels in the other bands where each superimposed segmented region of multi band must satisfy a given homogeneity criterion. The final multi spectral classification is derived from the segmented image.

Introduction
It is well known that an image segmentation technique is a powerful tool for automated interpretation of the satellite images recently many researches concern about region segmentation which attempts to group pixels in to surfaces regions based on the homogeneity or similarity of image properties or features have been proposed among all the techniques of region based segmentation we found that the graph theory technique is very useful one since it gives an almost optimum result of segmented image. Particularly it has been shown in the reference 4 that only one band of the multi band satellite spectral images is required by such a segmentation technique in this paper in order to increase the accuracy and precision if the segmentation process the technique of region segmentation based on graph theory is employed to classify a multi spectral satellite image where by the multi spectral bands images are used to obtain a segmented image. The example that used to demonstrate the algorithm is implemented on a micro-computer IBM PC/AT compatible.

Segmentation procedures
For the analysis of the cover type remote sensing data such as the satellite images of the type shown in figure 1 the multi spectral classification is a very useful process therefore in this section the proposed image segmentation technique will apply to the satellite images in addition for our method a graph theory is applied to the band that gives the highest contrast . The procedure of simultaneous segmentation of multi band images can be described as the following steps.
  1. Define the homogeneity threshold value(Î) which will be applied to all of the image band.

  2. Map one of the bands on to a graph each pixel in the band is mapped on to a vertex in the graph where the vertex weights are the intensities of corresponding pixels. Each vertex links to the adjacent vertex in the form of 4-way connected direction.

  3. Calculate the link weights between the adjacent vertices which are obtained by the absolute difference value of vertex weights as shown in figure 2.


    Fig.1 Example of Landsat multispectral images.


    Fig.2 Shows the mapping of the band 7 image onto a graph.


  4. Search the least weighted link in the graph and satellite from two vertices which connected value by this least weighted link the difference between maximum and minimum of intensity value.

    In the same time the difference between maximum and minimum intensity value from the two corresponding pixels (or regions) in the other bands will also be calculated the least weighted link will be removed if all the differences values which are calculated from this step is less than the given homogeneity threshold value then the higher segmented region will be formed by agglomerating these two vertices (or regions).

    On the contrary if the difference value from any bands is not less than the given homogeneity property, the two vertices (or regions) still separate.

  5. Repeat the step (4) until the last least weighted link the final result will give a segmented image which can image supper imposed to all of the image band and each super imposed segmented region of each image band is satisfied the given homogeneity property.
Implementation example
In this section, the proposed algorithm for simultaneous segmentation of multi band images is applied to the land sat imageries shown in fig 3 the algorithms is implemented on a micro computer IBM PC /AT compatible we apply the graph theory to the band since it gives the highest contrast after processing according to the procedure in the section 2 we obtain the super imposed segmented image as shown in fig 4 where the given homogeneity threshold value (Î) is chosen to be equal to 18 while in the figures 5 shows a random coloured classification from the mean value of segmented region from our algorithm.


Fig.3 Landsat multispectral images over the Bangkok area,
(a) band 4, (b) band 5, (c) band 6 and (d) band 7.


Fig.4 The segmented image obtained from the proposed
algorithm where the given threshold (Î) is 18.


Fig.5 Random coloured classification by the
mean value of segmented region.


Conclusion
The method of simultaneous segmentation of multi band images has been developed in this article the main feature of the proposed method is that segmented boundaries can be found very accurately since the pixel information of all the multi spectral bands are simultaneously considered therefore a high precision multi spectral classification can be obtained from the segmented image.

Reference
  • T.V. Robertson, K.S. Fu and P.H Swain " multi spectral image partitioning " LARS Inf. Note 07137 Purdue Univ. 1973.
  • S.L. Horowitz and T Pavlidis, " Picture segmentation by directed split and merge procedure Proc. 2nd int. joint Conf. on pattern recognition Aug. 13-15, 1974, pp. 424-433
  • Y. pramotepipop and F. Cheevasuvit " modification split and merge algoritm for image segmentation Proc. 9th asian conf. on Remote Sensing Nov. 23-29 1988 pp. q26 1-q26.6
  • F. Cheevasuvit image segmentation based on graph theory 23rd int symospim on remote Sensing of environment Apr. 18-25 1990.
  • D. cheriton and R.E tarjan finding minimum spanning trees SIAM J Comuter 1976 5 pp. 724-742.
  • O.J. Morris, M. de J. Lee and A. G. Constantinides "Graph theory for image analysis an approach based on the shortest spanning tree" proc. IEE vol. 133 pt, F, no. 2 April 1986, pp. 146-152.