Hyperspectral Image Segmentation Using Seed Points and Minimum Path Estimation Method
Fatemeh Hajiani Islamic Azad University
Khormooj Branch Khormooj, Iran [email protected]
Ahmad Keshavarz Electricity Department Persian Gulf University
Bushehr 75169, Iran [email protected]
Hossein Pourghassem Department of Electrical Engineering, Najafabad Branch,
Islamic Azad University, Isfahan, Iran [email protected] Abstract—In hyperspectral images, segmentation as preprocess
has high importance. In this paper, in two stages, the process of segmentation is done by considering spectrum of pixel in all bands. In the first stage, the image is turned to some subzones by use of flatting zone method, and then a more complete segmentation is done by use of estimation method of minimum path on zones. The suggested method produces new segmentation by use of appropriate choice of seeds and by considering minimum path of pixel to the seeds. The suggested method is implemented on AVIRIS image and produces more ideal number of zones and borders in compare with the other method.
Keywords-hyperspectral image, segmentation, flatting zone, estimate of minimum path
I. INTRODUCTION
Segmentation of hyperspectral image has high importance and as preprocess helps us in the next stages.
Methods of segmentation which turn levels of light changes to small zones are not appropriate methods. For solving this problem, several methods are suggested. One way to solve this problem is the use of a repeating algorithm which acts by determining the number of above zones and incorporation of small zones [1]. In the said method, segmentation is sensitive to amount of threshold. In the other study, segmentation is done by use of incorporation of spectral and contextual features of the first two components of PCA image [2]. Incorporation of two different features is hard and some information destroyed, because of not using of all bands. In the other way categorization is done by choice of band and the image is segmented by level set method [3].
II. SEGMENTATION OF IMAGE
With segmentation, the image is divided to formed parts, so that the same pixels are put in the same zone. Level set and watershed are two ways of segmentation among prior methods. Watershed is one of the strong instruments of morphology which does the segmentation of image by identifying continuant borders between zones [4]. This change is acted based on the change of the size of gray pixels and is enforceable by use of gradient of image [5].
For 2D image segmentation, the level set boundary is the zero level set of an implicit function:r2 rthe level set
methodology tracks the motion of the zero level set boundaries according to forces acting normally to the zero level set curves [6].
III. THE SUGGESTED SEGMENTATION
In this method, the image, with choice of a point which is called seed and incorporating of pixels to it, is segmented in two stages. In the first stage the image is segmented by use of flatting zone method and with regard to this fact that this stage is sensitive to threshold amount, another segmentation is done by the estimate of minimum path on the segmented images in the first stage.
A. Segmentation by flatting zone method
The template is used to format your paper and style the hyperspectral images are in form of discrete multivariable functions with tens or hundreds of spectral bands, which any pixel of them can be regarded as a vector. If hyperspectral image is shown with f, fi x showsx pixel of iband.
Based on flatting zone method which is a local segmentation, the two points xand
y
can belong to a same zone, if the distance between these two pointsxandy
are considered in chain form of points (p0,pi,....,pi,...pl), which all the points pi and pi1are contiguous and the criterion of similarity between contiguous pixels are smaller than [7]. Contiguity of any pixel with next pixels is considered in octet form. In this analysis, the similarity criterion of spectrum size is used. Euclidean distance is defined as one of criterion of distance [8] in form ofl
j
i j i j
orig f p f p
Ed
1
1 2
(1) For rational comparison, the distance between zero and one is measured.
) /(
)
(Ed m M m
Ed orig
(2) Whichmand
M
are the highest and the lowest amounts ofEdorig respectively. is a similarity criterion which shows correlation between two vectors and is defined as 2013 International Conference on Communication Systems and Network Technologies
1
1 1
1 1
i i
i i
p p
p i j p i
j p f p
f
l
(3)
Mand are mean and standard deviation in the pixel. The size of spectrum similarity is incorporation of criterion correlation and distance [9].
2
21
Ed
SSV (4) This kind of segmentation is very sensitive to amount of. It is done for the choice of which mean distance of 16 classes of training sample is measured from each other by use of similarity criterion of spectrum size. Then the amount ofis obtained by mean of three minimum amount of them.
[mean-mean/2 mean+ mean/2] is regard for the choice of threshold. The measured amount ofis 0.55.
B. Segmentation by estimate of minimum path
In this method, at first the border of the zone of segmentation in the first stage is determined and for the pixels on the border, minimum path to pixel of seed and existing pixels in its path are determined. For the choice of seed vectorial medianf(k)is defined which its component are measured by use of total distance of any pixel to the other existing pixel in each zone of segmentation in the first stage [10,11]. Then the minimum amount of this vector is chosen as a seed:
)) ( ( min arg
)) ( ), ( ( min
arg
/
p f
x f p f d k
R R p
R x i
i R
p
i
(5) For the measurement of minimum path, the algorithm of dijkestra is used [11]. By use of the amount of pixels, diagram is drawn in the minimum path from the pixel on the border to the seed and this diagram is estimated by a linear relation, then the error of mean of the path is measured by use of regression. If error of mean is lower than threshold limit, all the pixels in this path are put in the same zone. For this the threshold error is 58. Otherwise the path from seed of the place where its mean error of path is less than threshold error is determined and its pixels are put in new zones and Component of pixels is omitted from the vectorial median. In the next stage, pixels on the border are ignored and the new border is determined. The way is continued unless there is not pixel. Then from the vectorial median, the new seed is chosen and until the omission of all the components of the vectorial median, the stages are repeated.
Based on this, in the segmentation of the first stage, one zone may be divided into several subzones. For the other zone, segmentation of the first stage is repeated like this.
IV. PRACTICAL RESULTS
The used image is taken by AVIRIS sensor this image has
zone method, based on similarity criterion of SSV with a threshold amount of 0.55 and In Figure 2 the minimum path of a pixel to seed on the border of the zone of the first stage of segmentation is shown.
Figure 1. Segmentation based on similarity criterion of SSV.
Figure 2. Minimum path of pixel to seed on the border of the zone of segmentation in the first stage.
is shown for it. In Figure 4 path from seed of the place where it mean error of the path is less than threshold is shown and in Figure 2 pixel of end path as fill circle displayed. Figure 5 shows image of segmentation in the minimum path estimation method with a threshold error amount of 58. In Figure 6, the image of segmentation in level set and watershed is shown.
(a)
(b)
Figure 3. (a) Minimum path of pixel on the borer to seed with existing pixels in path (b) linear relation estimated for the minimum path.
Figure 4. Path from seed of the place where it mean error of the path is less than threshold shown.
The number of zones of segmented image is shown in watershed, level set and estimation of minimum path in Figure 7. For the study of the results of segmentation, the image of thematic is shown in Figure 8.
Figure 5. Image of segmentation in the minimum path estimation method.
(a)
(b)
Figure 6. (a) Watershed method segmentation in (b) Level set method Segmentation
Figure 8. Thematic map
With regard to the diagram, watershed method has the highest number of zones and level set method has the lowest number of method. In the watershed method, any congenial zone is divided into several subzones and the produced borders are not the ideal ones. In the level set method, the incorporation heterogeneous zone can be seen clearly. In this method some parts are not segmented. In the suggested method, congenial pixels are put in a zone and produce more ideal number of zones and borders in compare with the said method. But in this method, there is the possibility of incorporation of contiguous zone with high similarity.
V. CONCLUTION
If the segmentation is done correctly, congenial pixels are put in one zone and this does not produce interference of zones. In watershed method, borders of any congenial zone are divided into several subzones and this makes the number of zone increase. But because the zones are small, the possibility of the interference of zone happens less. In level set method, the number of zones in compare with the other methods is lower, but several heterogeneous zones have incorporated with each other. This method is sensitive to parameters and the start points of curves are very important.
Estimate method of minimum path puts the congenial pixels in one zone and produces ideal number of zones and borders. But in this method, there is the possibility of the incorporation of the contiguous zones.
REFERENCES
[1] D. Brunner and P. Soille, “Iterative area seeded region growing for multichannel image simplification,” Springer International symposium on Mathematical Morphology, pp. 397-406, 2005.
[3] J. Ball and L.M. Bruce, “Level Set Hyperspectral Segmentation:
Near-Optimal Speed Functions using Best Band Analysis and Scaled Spectral Angle Mapper,” Proc. IEEE Geoscience and Remote Sensing Symposium (IGARSS), Denver, CO, August 2006.
[4] G. Li, and Y. Wan, “Improved watershed segmentation with optimal scale based on ordered dither halftone and mutual information”.
IEEE. Computer Science and Information Technology, Vol. 17, pp.296 – 300, 2010.
[5] P. Li, X. Xiao, “Multispectral Image Segmentation by a Multichannel Watershed-based Approach,” Internationa Jornal of Remote Sensing, Vol. 28,No. 19, 10, pp. 4429- 4452, October 2007.
[6] Li, C. and et al's., "Minimization of Region-Scalable Fitting Energy for Image Segmentation", IEEE Trans. Image Processing, vol. 17 (10), pp.1940-1949, 2008.
[7] 3G. Nogel, J. Angulo, and P. Jeulin, “on Distance, Path and Connection for Hyperspectral Image Segmetation,” presented at the 8nd Int. International Symposium on Mathmatical Morphology, Rio de Janeiro, Brazil, Vol. 1, pp. 399-410, 2007.
[8] P. Keranen, A. Kaarna, and P. Toivanen, “Spectral Similarity Measures For Classification in Lossy Compression Of Hyprespectarl Images,” proceedings of the 9’th International Symposium on Remote Sensing, SPIE, pp. 285-296, vol. 4885,Crete, Greece, September , 2002.
[9] J.C. Granahan and J.N Sweet, “An Evaluation Of Atmospheric Correction Techniques Using The Spestarl Similarity Scale,” IEEE International Geoscience and Remote Sensing Symposium, Vol. 5, pp. 2022-2024, 2001.
[10] S.S. Bedi, GS Tomar & Shekhar Verma, “NPT Based Video Watermarking with non-overlapping Block Matching”, Springer- Verlag Berlin Heidelberg Transactions on Computational Science, XI, LNCS 6480, pp. 21 –29, 2010.
[11] J. Serra, “A lattice approach to image segmentation,”springer Science, Journal of Mathematic Imaging and Vision, 24, pp. 83-130, 2006.
[12] E. W. Dijkstra,“A Note on Two Problemin Connection with Graph,”
Numerische Mathematik, pp. 269-271, 1959.