Image segmentation by fusion of low level and domain specific information via Markov Random Fields☆
Introduction
In most of the recent studies, researchers address image segmentation and object recognition tasks in the same framework [6], [17], [1], [31], where top-down and bottom-up approaches are employed cooperatively. It is expected that these methods enrich the segmentation process by incorporating additional information extracted from recognition or detection tasks.
Markov Random Field models (MRF) are widely used in the segmentation literature [17], [20], [30], [25], since they are convenient for integration of information in various forms. However, these methods, construct energy functions in a supervised learning paradigm, where the relations between image parts are learned from a training set. Unfortunately, statistically sufficient amount of labeled dataset may not be available in many practical problems. On the other hand, some common information related to the content of image dataset may be available, especially, if the dataset consists of images from a specific domain which share similar properties. In this study, the information about the problem domain is referred as domain specific information (DSI) and it is formally represented by a domain specific map (DSM). The novelty of this study is the incorporation of the high level information into segmentation process in an unsupervised framework. This is possible only if the expert knowledge about a given problem domain is available. In this case, DSI is used to partially label the image regions to generate the Domain Specific Maps, which are then, utilized in an unsupervised MRF segmentation framework. Experiments indicate that the suggested Domain Specific MRF segmentation improves the segmentation quality with respect to the global consistency error and probabilistic rand Index.
Section snippets
Related work
Depending on the form of available information, segmentation can be complemented either by labeling the whole or parts of the image [17], [22]. Recent segmentation studies take up a bottom-up approach and employ a set of segmentations or classifiers whose outputs are later unified to obtain a segmentation with a corresponding labeling [13], [22], [18], [31], [3], [15]. These studies assume that a labeled dataset for a set of classes is available and they adopt a class-based approach by
Motivation
In application domains such as remote sensing, one seeks a group of objects in a highly cluttered background. For instance, if the goal is to detect the airplanes or airports in a remotely sensed image, the unsupervised segmentation algorithms are quite naive to extract the airport regions or airplanes with a complete segmentation method. Similarly, supervised segmentation approaches require large amount of labeled data. Even if sufficient labeled data is available, due to the large within
Experiments
A set of images from Microsoft Research Cambridge Object Recognition Image Database (MSRC) [32] is utilized in the experiments, where 115 outdoor images with vegetation are selected. All images have a pixel-wise ground truth segmentation. Sample images from this dataset are given in Fig. 2. DSI related to this dataset is defined as,
Normalized Difference Vegetation Index (NDVI) is used to detect vegetation regions [21]. This index is a normalized ratio of red
DS-MRF on a remote sensing application
A large number of MRF based methods exists in the Remote Sensing Literature. Poggi et al. [27] propose a supervised segmentation system which segments a given image based on a tree-structured MRF model. On the other hand, Porway et al. [28] propose a hierarchical model which incorporates bottom-up and top-down processing for image parsing to eliminate contextual inconsistencies. Moser and Serpico [23] propose a classification system which employs Markov Random Fields for fusing data from a set
Conclusion and future work
An MRF based image segmentation method which incorporates the domain specific information on unlabeled datasets, is introduced. The proposed Domain Specific MRF (DS-MRF) segmentation is compared with the state of the art methods on MRSC dataset. Experiments show that the suggested DS-MRF segmentation improves the performance depending on the effectiveness of the DSI coming from the experts.
The suggested segmentation method is applied to the remote sensing images, where the final goal is to
References (32)
- et al.
Color image segmentation and parameter estimation in a Markovian framework
Pattern Recognit. Lett.
(March 2001) - et al.
Colour and shape analysis techniques for weed detection in cereal fields
Comput. Electron. Agric.
(2000) - et al.
Semantic image segmentation and object labeling
IEEE Trans. Circuits Syst. Video Technol.
(2007) - et al.
Contour detection and hierarchical image segmentation
IEEE Trans. Pattern Anal. Mach. Intell.
(2011) - P. Arbelaez, B. Hariharan, Gu Chunhui, S. Gupta, L. Bourdev, J. Malik, Semantic segmentation using regions and parts,...
Exactly Solved Models in Statistical Mechanics
(1982)Spatial interaction and the statistical analysis of lattice systems
J. R. Stat. Soc. Ser. B
(1974)- et al.
Class-Specific, Top-Down Segmentation
- et al.
Fast approximate energy minimization via graph cuts
IEEE Trans. Pattern Anal. Mach. Intell.
(2001) - C. Christoudias, B. Georgescu, P. Meer, Synergism in low-level vision, in: 16th International Conference on Pattern...
Mean shift: a robust approach toward feature space analysis
IEEE Trans. Pattern Anal. Mach. Intell.
Efficient graph-based image segmentation
Int. J. Comput. Vision
Cited by (0)
- ☆
This paper has been recommended for acceptance by Y. Liu.
- 1
Instructor in Akdeniz University, Antalya, Turkey.