IMAGE SEGMENTATION BY ADVANCED MORPHOLOGICAL FILTERING AND CLUSTERING

Marcin IWANOWSKI, Sawomir SKONECZNY, Jarosaw SZOSTAKOWSKI Warsaw University of Technology, ul.Koszykowa 75, 00-662 Warszawa, POLAND ABSTRACT This paper is devoted to a segmentation method consisting of two steps: advanced morphological filtering by reconstruction and clustering by k-means algorithm. First step is based on morphological reconstruction and two filters are applied: opening by reconstruction and closing by reconstruction. This kind of operation has very important advantage from the point of view of segmentation - it preserves the borders of regions. Traditional filters (opening, closing, linear filters) remove noise, but on the other hand they cause some blur effects, which can be the serious obstacle for correct segmentaion. Morphological filtering by reconstruction has very good filtration properties without changing the shapes. Second step of segmentation consists of clustering. Two versions of k-means clustering algorithm will be described: classic and fuzzy one. First, 'crisp' version will be applied to the cases with a knowledge regarding number of clusters given a priori. Fuzzy version should be used when it it is difficult to define cluster number. The algorithm will automatically adapt number of clusters into the structure of the image. A combination of filtering by morphological reconstruction and clustering makes possible to consider two kind of information: spatial (filtering) and spectral (clustering). Key words: clustering, fuzzy logic, k-means, mathematical morphology, reconstruction INTRODUCTION The presented method is divided into two steps: morphological filtering nad clustering. First step is the preprocessing step performed in order to consider spatial dependencies in the image. The main areas of the image are transformed to more homogeneous. It means that all of the pixels with the graytone considerably lighter or darker than their background are removed in order to avoid misclassifications during the second, clustering step. Typical solution for such kind of problem is opening, closing, or linear filtering. But they change the shapes of regions on the image (bluring), which is the main obstacle for using this kind of filters in segmentation process. Segmentation should extract the shapes in the image, so each distortion of shape is dangerous. In order to solve this problem more advanced mathematical morphology is used: geodesic operators. Based on this operators two filters can be defined: opening by reconstruction and closing by reconstruction. They both remove lighter or darker pixels or areas on the image's surface, but they do not change the shapes. So they are very well suited for the segmentation purposes.

The second step of proposed method is called clustering and is destined to a clasification of image's pixels. To do this the k-means algorithm is used. This algorithm is iterative and consists of two steps: clasification and calculation of cluster centers. Each iteration gives better approximation of the cluster centers. The algorithm is itertively performed while the cluster centers are moving. Each pixel is classified to the cluster with the closest center. Clusters are represented by theirs centers, and they correspond with the regions on the image. Two kinds of a k-means algorithms are considered. The first one is traditional and gives as the result the number of clusters which is given by the user. The second is based on fuzzy logic. It can give as the result the number of clusters different from those given at the beginning of the algorithm. In this case the number of clusters is adapted to the real structure of the image. The algorithm starts with given number of clusters and later this number can increase or decrease depending on the spectral distribution of image's pixels. MORPHOLOGICAL FILTERING BY RECONSTRUCTION Mathematical morphology is a very efficient non-linear tool for signal and image processing. An introduction to mathematical morphology is to find in (Serra, 1982; Serra, 1988). In this paper morphological filtering by reconstruction is considered. The following part describes briefly theoretical background for opening and closing by reconstruction. Both operations belong to one group called geodesic morphological transformations (Soille, 1994). Geodesic dilation (erosion) of size 1 of the marker image f with respect to the mask image g where f g ( f g ) is defined as infinimum (supremum) image of mask image and marker image after dilation (erosion) of size one - with elementary structuring element, and is denoted as: g(1) ( f ) ( g(1) ( f ) ): g(1) ( f ) = (1) ( f ) g , g(1) ( f ) = (1) ( f ) g, (1) Geodesic dilation (erosion) of size n consists of n succesive geodesic dilations (erosions) of size 1: g (n) ( f ) = g (1) ( g(1) .... g (1) ( f )...)) , (gn) ( f ) = (g1) ( (g1) .... (g1) ( f )...)) (2)
n times n times

Reconstruction by dilation (reconstruction by erosion) of a mask image f from a marker image g where f g ( f g ) is defined as succesive geodesic dilations (erosions) of f with respect to g performed until idempotence and is denoted by: Rg ( f ) ( Rg ( f ) ):
Rg ( f ) = (gi ) ( f ) , Rg ( f ) = (gi ) ( f ) ,

(3)

where i such that (idempotence): (gi ) ( f ) = (gi +1) ( f ) , (gi ) ( f ) = (gi +1) ( f ). More information about reconstruction is in (Soille, 1994; Vincent, 1992). Opening and closing by reconstruction are defined as follows:

(Rn) ( f ) = Rg ( (n) ( f )),

* (Rn) ( f ) = Rg ( ( n) ( f )).

(4)

Unlike traditional opening and closing, opening by reconstruction and closing by reconstruction preserves shapes in the image. It is very important while using these operations as a first step of segmentation. It makes possible removal of the local peaks of gray-intensity without changing the borders of regions (which happens in traditional opening and closing).

One of the most efficient filters are alternating sequencial filters (ASF), which are consist of two families of filters. Each family contains operation of the same type but of different, increasing size. It can be defined as follows: ASF i = (i ) (i ) ... ( 2) ( 2) (1) (1) (5)

, are the families of filters, i is defined as a size of filter. In most of cases and represent opening and closing (Meyer and Serra, 1989; Serra and Vincent, 1992). In our case we use these operation by reconstruction where means closing by reconstruction and - opening by reconstruction. We will call this filter later ASF by reconstruction (ASFR): ASFR i = R ( ) R ( ) ... R ( 2) R ( 2) R (1) R (1)
i i

(6)

CLUSTERING Up to now a huge amount of segmentation algorithms was developed. In general one can divide them into two groups (Soille, 1992): based on spectral and spatial properties of the image. To the first group belong all the algorithms based on the intensity of image's pixels like e.g. clustering. Second group bases on the relationships between pixels and theirs neighbors. To this group belongs the most popular morphological segmentation technique watershed (Beucher, 1990; Meyer and Beucher, 1990). This paper describes only first type of segmentation - clustering. In our case not only spectral properties are considered because first step - morphological filtering was performed. Those step have removed the undesirable intensity peaks using the spatial information, so the spatial constraints are in such a manner considered. Proposed clustering method bases on the k-means technique (Duda and Hart, 1973; Selim and Ismail 1984; luzek, 1991; Iwanowski, 1996). K-means algorithm aims at classifying each member of given set of input data into given number of clusters. Algorithm deals with distances between samples and cluster centers. This distance is iteratively minimized in two steps performed iteratively. First step distributes samples into the clusters, second updates cluster centers. Each sample is classified to the cluster which center is the nearest in given metric. Updating of cluster centers is performed by computing a new cluster center by placing it in the middle (arithmetic mean value) of all of the samples belonging to the cluster. These two steps are performed iteratively while clusters centers change their position. If in the last performed iteration cluster centers did not move, it means that operation of samples classification is finished. In case of applying k-means algorithm for images notions of sample and cluster center are represented by pixel's graytone. During the calculations the distance between samples and cluster centers is calculated by taking absolute value of difference between graytones of sample and cluster center. Cluster centers obtained finally represents graytones of the output image. These graytones refers to regions on the image surface. Whole algorithm can be written as follows: 1. Set initial cluster centers (output graytones). 2. Classify pixels. Each pixel is labeled by the id number of the closest cluster. 3. Calculate new cluster centers (output graytones). 4. If in cluster centers obtained in step 3 are different to centers obtained in previous iteration then go to step 2, otherwise go to step 5.

5. Using cluster centers (output graytones) classify pixels of the input image into output one. Change value of each pixel to the value of the closest cluster center. Initial cluster centers (point 2 above) can be calculated in any way. We have placed them equally distanced in a whole range of graytones. Next remark refers to the spectral property of this kind of segmentation. Because k-means does not consider neither position of pixel nor its neighbors, there is no need to make all of the calculations on image. Image histrogram is sufficient to do this. This solution is faster because it needs smaller amout of computations. In version of the algorithm described above number of clusters is fixed and cannot be changed. This solution is not always sufficient. Sometimes one would have flexible algorithm which can change the number of cluster according to the real situation on the image. In this case fuzzy version of k-means algoriithm (fuzzy k-means - FKM) can be applied. Some fuzzy versions of k-means was developed. In this paper version close to (Huntsberger et al., 1985) is proposed. The main difference between traditional and FKM algorithm is representation of the result. In traditional version each point is labelled by the cluster number. Fuzzy version produces so called 'fuzzy partition' - a vector which includes membership values. These values tell us 'how much' given point is similiar to each of the clusters. This is not only indication of the closest cluster center but also a description of similiarity with other clusters. The schema of algorithm is similar to that presented above. Only some changes should be made. Calculation of cluster centers is performed using the following equation: ci

=
n

j =1 n

( ij ) m x j

( ij ) m j =1

(7)

(all symbols will be explained later) Classification step computes new vector with membership values and is performed for each pixel (index p) as follows: - compute distance between pixel and cluster centers: 1 i k ; dip = x p ci - if one or more distances d are equal to zero perform for all 1ik:

ip = 0 if

dip = 0

otherwise for all 1ik:

otherwise ip = 1 l (l - number of distances d equal to 0)

(8)

ip =

(
j =1

(9)

dip 2 /( m 1) d jp

where: n - is the number of pixels, ci - is center of i-th cluster, x p - is i-th sample (pixel), ip is the element of fuzzy partition, it describes membership value: 'how much' p-th sample belong to i-th cluster. Parameter m describes degree 'fuzziness' of classification (nature of clustering). m=1 gives 'hard' or 'crisp' clustering, bigger m give fuzzier classification. Presented above descriptions refers to 'pure' FKM algorithm. In order to apply it to segmentation some additional steps should be made. FKM algorithm is performed until convergence. When the fuzzy-partition elements stops changing their elements, next step is

executed. Cluster centers are compared in order to check if there are clusters with centers close to each other. If two or more clusters have their centers close, they represent in fact one cluster. In this case number of clusters k decreases and FKM algorithm is performed once again. When number of clusters is stable next step is performed. Now all the pixels with the membership values high enough are classified to the clusters with the highest membership value. The thresholding operation with threshold which describes how 'high' is 'high enough' is called -cut. All the pixels not classified (their membership values were too small to be classified properly) are classified once again using FKM. Whole algorithm can be summarized as follows: 1. Set whole image as a set of pixels for classification. 2. Perform FKM: 2a. Initialize fuzzy partition. 2b. Generate fuzzy partition (membership values for each pixel). 2c. Calculate new cluster centers. 2d. If new cluster centers are different to previous ones go to step 2b. 3. Check if real number of clusters after FKM is equal to k if yes then go to step 4, otherwise set k=real number of clusters and go to 2. 4. -cut, classify all the pixels with membership values higher than to the closest clusters. 5. If not all pixels were classified in step 4 then set not-classified pixels as a new set for classification and go to step 2, otherwise stop the algorithm. RESULTS In order to apply proposed algorithm some experiments were carried out. We used 256graytone images, resolution 256x256 pixels.

Fig. 1. (a) - input image (mangan), (b) - after ASF of size 2, (c) - after ASFR of size 2, (d) after ASFR of size 2 and k-means (k=2). Fig. 1. shows the results of two kinds of filtering: by ASF and by ASFR. First one consists of four operations: closing of size 1, opening of size 1, closing of size 2, opening of size 2 (ASF of size 2). It is visible that all the disturbances on the background have disappeared, but at the same time shapes of objects have been changed.

Fig. 2. (a) - intensity of pixels from Fig. 1a, (b) - after ASFR of size 1, (c) - after ASFR of size 2, (d) - after ASFR of size 2 and k-means (k=2). Second filter gives correct result. It is ASFR filter of size 2 contained closing by reconstruction and opening by reconstruction (size 1 and 2). Image is filtered without any change of shape of objects. Image filtered by ASFR is much better input for k-means algorithm than image without filtering. Only the important regions exists and theirs borders are unchanged. In this case 'crisp' version of k-means algorithm was applied. Number of clusters is known - we have to extract darker shapes on lighter background - and is equal to 2. Result of k-means of size 2 is presented in Fig. 1d. Fig. 2. includes the 1-D views of the image from Fig. 1a made by horizontal cut of image at y=120. At the first image (Fig. 2a) initial image is shown - image containes a lot of local graytone disturbances which, from the point of view of 'pure' image analysis, represent noise. This noise should be removed. K-means applied directly on this image gives bad result, because apart from the proper areas also a lot of noise will be considered as objects. Next picture (Fig. 2b) shows the result of filtration by ASFR of size 1. Some smaller peaks disappear, but larger remain, so the size of ASFR should be increased. At the next picture (Fig. 2c) result of ASFR of size 2 is presented. All the important peaks are removed so kmeans can be now applied. Result of this operation is shown on Fig. 2d.

Fig. 3. Other examples: (a) granite, (b) granite after ASFR of size 2 and k-means (k=3), (c) cooper, (d) cooper after ASFR of size 1 and k-means (k=2). Fig. 3. contains other examples of proposed segmentation method also with 'crisp' k-means. Fig. 3a presents structure of granite. Three main areas exists on this image. This case needs ASFR of size 2 and k-means with k=3 - result is shown on Fig. 4b. Other example contains stucture of cooper (Fig. 3c.) which requires ASFR of size 1 and kmeans (k=2) - see Fig. 3d. In examples shown above number of clusters was easy to define a priori. But it is sometimes in not possible. Number of clusters can be not available a priori. This can happen in scene analysis. In this case we apply fuzzy version of k-means (FKM).

Fig. 4. (a) input image (scene), (b) ASFR of size 5, (c) FKM k=3, m=3, =150, (d) FKM k=5, m=3, =150. Image shown in Fig. 4a represents a scene. From the point of view of segmentation it contains a lot of noise, especially on the right side. ASFR is applied in order to remove possibly lot of noise. In this case ASFR of size 5 was applied - larger than in examples shown before, because the structure of noise is more complicated. Filtrated image is presented on Fig. 4b. Number of clusters can be defined here only roughly. In proposed FKM algorithm some parameters should be defined. Paramter k represents initial number of clusters. Parmeter m defines how 'fuzzy' behavior of the algorithm is (the bigger m is, the more 'fuzzy' algorithm is). Parameter indicates threshold for classified pixels. During the experiments the following parameters were chosen: m equal to 2 or 3, 130<<190. Segmented images with different parameters k are shown of Fig. 4c and Fig. 4d. The real number of cluster is in both cases bigger than defined by parameter k. Result is adapted to the image's structure. In this case 'crisp' k-means would give not 'natural' result - extracted regions would be not good fited to the real regions on the input image. CONCLUSIONS One of the biggest obstacles for applying k-means-like algorithms is always the spectral nature of this algorithm. It causes problems with bad-classified single points or small areas in the image. Because of that one should find a supplement that can precede the clustering step, and which can take under consideration spatial dependencies of the image. Natural answer to the question of finding such a supplement is filtering. But the most popular linear filters as well as non-linear (e.g. morphological) have one important disadvantage for segmentation: they change the shape of the areas on the image. Segmentation by definition should extract shapes of the areas, so each disturbance of it is a problem. In order to avoid this, an advanced morphological filtering is proposed: alternating sequential filter by reconstruction. As it was shown on the attached examples, this solution seems to be very efficient. It removes local disturbances of graytone without changing the shape of areas. Also two types of k-means algorithm was applied: traditional version anf fuzzy one. First one can be applied to the images where number of clusters can be defined by the user. But not in all cases it can be done. For such a case fuzzy k-means is applied. This variant needs an initial number of clusters, but during the run of the algorithm this number can change according to real image content. Experiments carried out show that proposed combination of advanced morphological filtering and clustering by k-means and fuzzy k-means considers both spectral and spatial nature of image and gives good and promising results. ACKNOWLEDGEMENTS

Authors would like to thank M. Mynarczuk from Instytut Mechaniki Grotworu PAN in Krakw for the set of geological images. REFERENCES S.Beucher. Segmentation d'images et morphologie mathematique. Ph.D. Thesis, School of Mines of Paris, June 1990. R.O.Duda, P.E.Hart, Pattern classification and scene analysis. Willey and Sons, 1973. T.L.Huntsberger, C.L.Jacobs, R.L.Cannon. Iterative fuzzy image segmentation. Pattern Recognition Vol.18., No.2., pp. 131-138, 1985. M.Iwanowski. K-means algorithm for greytone reduction. Technical Report, School of Mines of Paris, N-13/96/MM, 1996. F.Meyer, S.Beucher. Morphological Segmentation. Journal of Visual Comm. and Image Representaion. Vol.1, No.1, September, pp. 21-46, 1990. F.Meyer, J.Serra. Filters: from theory to practice. Acta Stereologica 1989; 8/2; Proceeding of 5-th European Congress for Stereology, Freiburg, September 1989, pp. 503-508. S.Z Selim, M.A.Ismail. K-means-type algorithms: a generalised convergence theorem and characterisation of local otimality. IEEE Trans. Patterm Anal. Machine Intell. 6(1), pp. 81-87, 1984. J.Serra. Image analysis and mathematical morphology. Academic Press, 1982. J.Serra, ed. Image analysis and mathematical morphology. Volume 2: theoretical advances. Academic Press, 1988. J.Serra, L.Vincent. An overview of morphological filtering. Circuits Systems Signal Processing Vol.11, No.1, 1992. P.Soille. Partitioning of multispectral images using mathematical morphology. Technical Report, School of Mines of Paris, N-22/92/MM, 1992. P.Soille. Geodesic Transformations in Mathematical Morphology: an Overview. Technical Report, School of Mines of Paris, N-24/94/MM, 1994. A.luzek. Komputerowa analiza obrazu. Wydawnictwa Politechniki Warszawskiej, Warszawa 1991. L.Vincent. Morphological Grayscale Reconstruction: Definition, Efficient Algorithm and Aplications in Image Analysis, IEEE Comp.Vis. and Patt.Rec., Proceedings, 1992, pp. 622-635