Image segmentation based on merging of sub-optimal segmentations
Introduction
The problem of the segmentation of images generally implies the partitioning of an image into a number of homogeneous regions, so that joining two of these neighboring regions gives rise to another heterogeneous region. Many ways of defining the homogeneity of a region exist, depending on the context of the application. Moreover, a wide variety of methods and algorithms are available to deal with the problem of the segmentation of images (Fu and Mui, 1981, Haralick and Shapiro, 1985, Pal and Pal, 1993). These methods can be broadly classified in four categories (Zhu and Yuille, 1996):
- •
Edge-based techniques.
- •
Region-based techniques.
- •
Deformable models.
- •
Global optimization approaches.
The edge-based techniques are based on information about the boundaries of the image. Therefore, they try to locate the points in which abrupt changes occur in the levels of some property of the image, typically brightness (Canny, 1986, Rosenfeld and Kak, 1982). On the other hand, those methods that use spatial information of the image (e.g. color or texture) to produce the segmented image fit into the region-based techniques (Chen et al., 1992, Sahoo et al., 1988). These methods depend on the consistency of some relevant property in the different regions of the image. The deformable models are based on curves or surfaces defined within an image that moves due to the influence of certain forces. They can be classified in various groups, principally snakes, deformable templates and active contours (Blake and Isard, 1998, Kass et al., 1988). All of these techniques avoid the use of a global criterion when segmenting the image, which is contrary to the global optimization approaches (Geman and Geman, 1984, Kanungo et al., 1994).
As a consequence, several solutions can be obtained when applying different segmentation techniques to the same image due to the wide variety of segmentation algorithms available for any context or application. The results of applying different techniques are usually dissimilar and normally the aforementioned solutions do not indicate evidently which of these methods performs a better segmentation of the image, even when the application context is well defined. Moreover, in some circumstances, it could be useful to combine algorithms with different goals if they are complementary. First, if the objective of the segmentation is not defined in a precise way. And second, when the objective can be expressed as a combination of several simple goals. For example, in order to detect blue and red objects it is useful to consider some algorithms to detect red objects and others to detect blue ones. So the information extracted from segmentation algorithms with different goals should not be always considered as contradictory.
Based on this, we propose the idea of using various segmentations of the same image to create a final segmented image. This will adequately combine the characteristics of these segmentations and the negative particularities of each of the segmentations can be masked by the others, so that the result is not dependent on the selection of a particular algorithm or on some initial conditions. In this way, the results obtained will be unified in one unique solution and this will have an effect on the quality of the final segmentation. Each of these segmentations, which from now on will be called partial segmentations, can be performed using different techniques or even, the same technique with a different parametrization or with different initial conditions.
In Fig. 1 the functional scheme of the proposed algorithm is shown. To obtain a final solution, an oversegmented image is created from partial segmentations. Afterwards, a region-merging algorithm is applied in such a way that those regions that are similar, according to certain criterion, are merged (Zhu and Yuille, 1996, Beaulieu and Goldberg, 1989, Shafarenko and Petrou, 1997). The proposed criterion used by our algorithm is based exclusively on the information collected from the partial segmentations in the oversegmented image creation process. In addition, our approach is based on the introduction of a repulsing force between neighboring regions that measures the tendency of these to remain separated. The structure of this work is as follows. In Section 2 the concept of shadowed zones is introduced and how we can obtain the oversegmented image from the partial segmentations is explained. In Section 3 the concept of the force of repulsion between regions that indicates the tendency of two regions to merge or remain separated in the final segmentation is introduced. Section 4 describes the merging process, while Section 5 shows the results using, as case of study, a segmentation evaluation function as stopping criterion of the merging process. Finally, Section 6 presents the main conclusions of this work.
Section snippets
Generation of the oversegmented image
The segmentation of an image can be considered as a labeling problem, in the sense that those pixels that belong to the same region have been assigned the same label, which can be specified in terms of a set of states or a set of labels. Let H be a discrete set with m states, H = {1, 2, … , m} and let L be the set of labels. In our case, a state represents a region of the Euclidian space of the image (set of pixels). Labeling is not more than a function that assigns one of the labels of L to each of
Force of repulsion between regions
One important contribution of our proposal is that all the relevant information to obtain the final segmented image is obtained exclusively from the different partial segmentations, both for creating the oversegmented image and for applying the subsequent region-merging algorithm. It should be noted that this strategy does not take into account local characteristics such as size, shade of average grey intensities, etc. Based on the results obtained for all the pixels of the image in each of the
Region-merging algorithm
The structure of the data that we have used for representing the partitions of the oversegmented image is a region adjacency graph (RAG) (Haris et al., 1998). The RAG of a segmentation of K regions is defined as a weighted undirected graph, G = (V, E), where V = {1, 2, … , K} is the set of nodes and E ⊂ V × V is the set of edges. Each region is represented by a node, and between two nodes R, S ∈ V there is an edge (R, S) if the regions are neighbors. Fig. 4 shows the RAG of the oversegmented image of Fig. 3. A
A case of study: evaluation function Q
Most of the results obtained by segmenting an image are evaluated visually and qualitatively. The principal problem is searching for an uniform criterion to evaluate the results. Different methods for evaluating the segmentation of images have been published (Levine and Nazif, 1985, Zhang, 1996). For instance, the evaluation function Q proposed by Borsotti et al. (1998) for color images. This function is a variant of that proposed by Liu and Yang (1994). One of its main advantages is that do
Conclusions
In this paper, a new heuristic segmentation algorithm is presented, based on the use of various segmentations of the same image that are combined to create an oversegmented image. These segmentations can be performed using different techniques or even the same technique with different initial conditions. A region-merging algorithm is then applied to this oversegmented image. The main idea of our proposal is to obtain high quality segmentations from a set of low quality partial segmentations. A
Acknowledgments
This study has been financed by the MCYT (Spain) through the research project TIC2001-3694-C02.
References (23)
- et al.
Quantitative evaluation of color image segmentation results
Pattern Recognition Lett.
(1998) - et al.
A survey on image segmentation
Pattern Recognition
(1981) - et al.
Survey, image segmentation techniques
Comput. Vision Graphics Image Process.
(1985) - et al.
An improved seeded region growing algorithm
Pattern Recognition Lett.
(1997) - et al.
A survey of thresholding techniques
Comput. Vision Graphics Image Process.
(1988) A survey of evaluation methods for image segmentation
Pattern Recognition
(1996)- et al.
Hierarchy in picture segmentation: A stepwise optimization approach
IEEE Trans. Pattern Anal. Machine Intell.
(1989) - et al.
Active Contours
(1998) A computational approach to edge-detection
IEEE Trans. Pattern Anal. Machine Intell.
(1986)- et al.
Split-and-merge image segmentation based on localized feature analysis and statistical tests
CVGIP: Graph. Models Image Process.
(1992)
Decomposition of 3D medical images into visual patterns
IEEE Trans. Biomed. Eng.
Cited by (44)
FractalRG: Advanced fractal region growing using Gaussian mixture models for left atrium segmentation
2024, Digital Signal Processing: A Review JournalMSMCNet: Differential context drives accurate localization and edge smoothing of lesions for medical image segmentation
2023, Computers in Biology and MedicineRFPNet: Reorganizing feature pyramid networks for medical image segmentation
2023, Computers in Biology and MedicineAutomatic identification of the area covered by acorn trees in the dehesa (pastureland) Extremadura of Spain
2020, Computers and Electronics in AgricultureCitation Excerpt :Many methods are available for image segmentation. In general, methods are classified into the following categories: clustering and classification (Ojeda-Magaña et al., 2018; Rajaby et al., 2016; Tan et al., 2013), edge or contour detection (Alonso et al., 2010; Fan et al., 2001) growth of regions (Pichel et al., 2006) and morphological (Boutalis et al., 2002). Within the first category, partitional clustering algorithms such as the c-means algorithms (Bezdek, 2017) divide a colour image (e.g., a cartesian RGB feature space) into a given number of clusters or groups, keeping the colour within each region as homogeneous as possible according to the optimisation criteria used.
A novel image segmentation method based on fast density clustering algorithm
2018, Engineering Applications of Artificial IntelligenceCitation Excerpt :The ability to work without reference images allows unsupervised evaluation to operate over a wide range of conditions and with many different types of images. Recently, Pichel et al. proposed a similar system (Pichel et al., 2006), using unsupervised evaluation to evaluate algorithm performance in real time and to adjust algorithm parameters, which obtains ideal processing result and improved efficiency. The test images are from the Berkeley Segmentation Dataset.
Fast processing of foreign fiber images by image blocking
2014, Information Processing in AgricultureCitation Excerpt :Fast and precise segmentation has always been of great concern to people. Various image segmentation methods are reported in the literature [2,11,15] some of which are used in the Automated Visual Inspection system in agriculture [3,7]. In recent years, researchers have developed more efficient, but also more complicated methods for segmentation.