Undesirable effects such as thick watershed lines and isolated minima are not uncommon during the process of flooding catchments and locating watershed points. In the proposed algorithm, flooding starts from predetermined regional minima, during which the label of the neighboring pixels is determined based on conditional neighborhood equations and geodesic distance from the outer edge of the nearest plateau. 2.11 (a) After adding Gaussian noise (/^=0, a=0.01) to the original Akiyo image; (b) result of open-close of noise contaminated image; (c) result of open-close through reconstruction.
3.7 (a) Original tennis image; (b) Multiscale gradient morphological operation on the table tennis image with a threshold value of 13; (c) Segmentation produced by Meyer's watershed algorithm; (d) Segmentation produced by the proposed watershed algorithm. 3.8 (a) Original Cermet image; (b) Multiscale gradient morphological operation on the Cennet image with a threshold value of 30; (c) Segmentation produced by Meyer's watershed algorithm; (d) Segmentation produced by the proposed watershed algorithm. 3.9 (a) Original MRI brain image; (b) Multiscale gradient morphological operation on MRI brain image with a threshold value of 12; (c) Segmentation produced by Meyer's watershed algorithm; (d) Segmentation produced by the proposed watershed algorithm.
3.11 (a) Original steel fissure image; (b) Morphological multiscale gradient operation on steel fissure image with a threshold value of 15; (c) Segmentation produced by Meyer's watershed algorithm; (d) Segmentation produced by the proposed watershed algorithm. Blood cell image with a threshold value of 26; (c) Segmentation produced by Meyer's watershed algorithm; (d) Segmentation produced by the proposed watershed algorithm.
5.13 (a) Original Lena image; (b) Morphological multi-scale gradient operation on Lena image with a threshold value of 12; (c) Segmentation produced by Vincent's watershed algorithm [VS91]; (d) Segmentation produced by the proposed watershed algorithm. Tennis image with a threshold value of 12; (c) Segmentation produced by Vincent's watershed algorithm [VS91]; (d) Segmentation produced by the proposed watershed. 5.15 (a) Original road traffic image; (b) Morphological multi-scale gradient operation on road traffic image with a threshold value of 20; (c) Segmentation produced by Vincent's watershed algorithm [VS91]; (d) Segmentation produced by the proposed water.
5.16 (a) Original blood cell image; (b) Morphological multilevel gradient operation on a blood cell image with a threshold value of 26; (c) Segmentation performed by Vincent's watershed algorithm [VS91]; (d) Segmentation due to proposed watershed. The success of the automated image analysis process largely depends on the accuracy of the segmentation technique used.
It is possible to segment an image into regions of common properties by locating each region for which there is a significant change in the properties across the boundary. If an image is noisy or if the region features differ by a small amount between regions, a detected boundary can often be broken. The edge-based segmentation was not very successful because of small gaps that allow merging of dissimilar regions.
At each iteration, the histogram of the image is drawn and the largest peak is separated from the rest of the image. Weska et al [WNR74] have suggested the use of a Laplacian operator in choosing the luminance threshold. A gray-value histogram formed from only those pixels of the original image that lie at coordinates corresponding to very high or very low values of the Laplacian tends to be bi-modal with a distinct valley between adjacent peaks.
Banu and Faugeras [BF82] proposed a gradient relaxation method for solving the unimodal histogram problem. The relaxation process is iterative where pixel values are changed so that the histogram is no longer unimodal and the threshold can be easily detected.
ALGORITHMS FOR IMAGE SEGMENTATION 4
Markov Random Field-based Algorithms
ALGORITHMS FOR IMAGE SEGMENTATION 5
- Watershed Transform Approach
- WATERSHED TRANSFORM APPROACH
- Applications of Watershed Transform Techniques
- Review of the Different Classes of Watershed Algorithms
The set of clustering feature vectors is updated for each image in the sequence which is also used for segmentation of the next image. Among the existing segmentation algorithms, the watershed transform has proven to be a very useful and powerful tool for morphological image segmentation due to its straightforward formation and implementation, as well as its ability to identify important contours of closed of a certain image. As depicted in Figure 1.1, a gradient image obtained by applying an appropriate gradient operator to a grayscale image is considered to be a topographical surface, where the grayscale value of a pixel indicates the height of that pixel.
The entire set of points of the surface, the steepest path of which reaches a minimum, forms the basin [BL79, Beu90a] associated with this minimum. All the points that drain into a common catchment are part of the same catchment. Each pixel in this digital image is given a label during the transformation of the catchment area of a regional minimum.
A watershed is then defined as the set of pixels whose respective downstream paths all end in the same labeled local minimum. The watersheds will fill with water, starting at these local minimums, and dams will be built at points where water coming from different watersheds meets. When the water level reaches the highest peak in the landscape, the process is stopped.
Thus, the watershed transformation refers to labeling an image so that all points of a given watershed have the same unique label, and a spatial label, different from all watershed labels, is assigned to all points of the watershed. Various definitions for watersheds in both continuous and digital space have been proposed in the literature [Beu90a,VS91,BM93,RM01,NC03]. It is known that the transformation of the watershed gradient usually causes over-segmentation (many basim of small watersheds).
The performance of the watershed transformation largely depends on the algorithm used to calculate the gradient. A multi-scale gradient image [Wan97, CKP98] is used as input to the catchment transfonn algorithms in the current work. Although the first application of watershed algorithms was in the field of topography, they have become more popular in several industries: biomedical signal processing [ADR95], medical image processing [WHOF96, RDKJ02], and computer vision [Cee95].
Additionally, various results have been reported on the use of the watershed algorithm for color and multi-spectral image segmentation, object-based video coding [Wan98, Cro97, SYL03] and remote sensing systems fcr satellite and radar images. Many other examples of transphone-based watershed segmentation can be found in the literature [Beu90a,MB90,Beu90b,BM93]. The first set of algo rates aims to detect the watershed by simulating the flood process [BL79, MB90, BM93.
Mey94, VS91, RMol] while the second group directly follows watershed lines using arrow techniques [Beu90a, BM93].
