Image Segmentation Is An Important Research Area Biology Essay

Published: Last Edited:

This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.

Image segmentation is an important research area in computer vision and many segmentation methods have been proposed. This paper attempts to provide a brief overview of elemental segmentation techniques based on boundary or regional approaches. It focuses mainly on the agent based image segmentation techniques.

1. Introduction:

Image Segmentation is a very important part in image processing. It has applications in image visualization, image coding, image synthesis, pattern recognition, rendering displacement estimation, etc (Ho and Lee, 2003). The categorization of images is generally done on the basis of their energy source; the principal source being the electromagnetic spectrum. Image segmentation is a process of dividing or partitioning of an image into certain regions such that each region is homogeneous and none of the union of two adjacent regions is homogeneous. Mathematically, the image segmentation can be defined as follows: Let F be a set of all image

pixels and p(.) be a homogenous predicate defined over groups of connected pixels then the image segmentation is a partition of the set F into a set of connected subsets or regions (s1, s2, ........,sn) such that

The homogeneity predicate p(.) = True for all regions si and p (sisj) =False for any two adjacent regions si and sj (Ho and Lee 2003). The image segmentation can apply in medical imaging such as the location of tumors and other pathological tests, measurement of tissue volumes, computer guided surgery, diagnosis and treatment planning, etc.Other applications include the location of objects in satellite, face recognition and finger print recognition, etc. The quality of the final output depends on the quality of the segmented output. The uniformity of light intensity is measured by uniformity predicate. (Pal and Pal, 1993).

The segmentation techniques are categorized into four classes: Edge based approaches, clustering based approaches; region based approaches and split/Merge approaches (Ho and Lee, 2003).

In the edge based approach an image is partitioned on the basis of abrupt changes in intensity (Cufi et al., 2001) Image edges are detected and thus linked into contours that represent boundaries of image objects (Gonzalez and Woods, 1992). Most of the techniques use a differentiation filter in order to approximate the first order image gradient or the image Laplacian. Then, candidate edges are extracted by holding the gradient or Laplacian magnitude (Canny, 1986). In the clustering based approach the

Image pixels are stored in increasing order as a histogram according to their intensities. Then, a predefined cluster number is used to separate the intensity histogram based on the intensity values. The number of regions is unsupervised because the locations of pixels in the same cluster may not be adjacent. (Ho and Lee, 2003). Several clustering - based approaches have been proposed, such as fuzzy-c-means (FCM) (Tolias and Panas, 1998; Chun and yang, 1996; Bezdek, 1981) and K means (Pappas, 1992; Chen et al, 1998; Tou and Gonzales, 1974).

In the region based approaches, the segmented contours are always continuous and one - pixel wide (Fu and Mei, 1981). It includes region growing (Chan and li, 1994; Adams and Bischof, 1994; Hojjatoleslami and Kittler; 1998), water sheds (Haris et al., 1998; Gauch 1999; Tremeau and colantoni, 2000) and pyramidal segmentation (Rezaee et al, 2000).

In this approach an input image is first tessellated into a set of homogeneous primitive regions. Then, similar neighboring regions are merged according to certain decision rules (Ho and Lee, 2003). Several split approaches are available, such as pyramidal segmentation (Rezaee, 2000) Watersheds (Hari, 1998, Gauch, 1999, Tremeau and Colantoni, 2000), FCM (Chun and Yang, 1996) and K-Means (Pappas, 1992). In Merge process, region adjacent graph (RAG) (Pavlidis, 1982) and nearest neighbor graph (NNG) (Haris, 1998) are both available structures. RAG and NNG are usually applied with a gredy merge process for removing unimportant regions, until a predefined stop condition algorithms without setting any threshold have been proposed for merging process (Chun and Yang, 1996). Agents are new paradigm of Modern artificial Intelligence (AI) research in computer Science. In this paper some of the Agent based image segmentation methods and some of their applications are described.

2. Agents:

Agents are defined as a software or hardware entities that perform some sets of tasks on behalf of users with some degree of autonomy. Agents can be used in the next generation model engineering -complex, distribution system (Laws, 1980). Agent based approaches have some advantages. They can be adapted to the locality. They are reliable in performance; are less sensitive to the noise, and are simple to represent and implement (Mass, 1995; Liu et al., 1997; and Keshtkar et al., 2005). Agent based approach is widely used in such fields as traffic management, network monitoring, robotics control, electronic commerce, Medicine and computer vision (Demazeau, 2003).

3. Agent Based Approaches:

Boucher and Garbay (1996) proposed a multi agent system for the problem of living cell segmentation and recognition. The agent integrates three basic behaviors, namely, segmentation, interaction and reproduction. The segmentation behavior is based on region-growing methods. The static and motion based criteria are used for pixel evaluation. The interactive nature between the two agents allows them to merge or to negotiate parts of regions. The negotiation is visualized as the process of segmentation and refinement carried out by the agents. The reproduction behavior defines an exploration strategy of the image frames. Agents can trigger other neighboring agents to start functioning or else they can duplicate themselves in the next frame .Segmentation of the frame is carried out in the pipeline and the previous information helps in the process of the current frame. The agents have different behaviors to achieve their tasks. Depending on the events that are collected from the basic behavior, normal region growing, other ones can be activated. The negotiation behavior is a way to perform some segmentation correction. The agent creation is dynamically elaborated and adapted from application constraints. Agents compete themselves to segment an image. They can label merge or negotiate zones. Behaviors into one agent also work under a competition scheme. They are driven by incoming events, which modify their priorities to allow one of them to execute at a given time.

The proposed autonomous agents are distributed computational entities that operate directly in the two dimensional lattice of a digital image and exhibit a number of reactive behaviors (Liu et al., 1997). Individual agents are able to sense the local stimuli from their image environment by means of evaluating the gray level intensity of locally connected pixel and activate their behavior accordingly so as to locate the feature pixels effectively. The agent is able to perform i) feature marking at local pixels and self reproduction of offspring agents in the neighboring regions, if the local stimuli are found to satisfy feature conditions; ii) Diffusion to adjacent image regions if the feature conditions are not held, or (iii) death; if the agents exceed their life span. The direction in which the agent will self reproduce and/or diffuse are inherited from the direction selected by high fitness parents. They presented several experimental results to demonstrate the optimal extraction of image features using the evolution of the distributed autonomous agent. The application of the above mentioned approach are identification of pathological foci of early stage cancer and important anatomical features from ultrasound images of a prostate (Muzzalini et al., 1993).Identification of speculated lesions, micro calcifications and circumscribed lesions in scanning mammograms for breast cancer.

This approach is applied for forest fire detection (Movaghati et al., 2008). Movaghati et al (2008) studied the potential of agent in processing of remote sensing imagery.

KAGAWA et al (1999) proposed simple rules; where in each agent has the features such as color and moves onto a pixel which has most of the similar features. The pheromone which is the idea based on the chemical substance has the property to keep agents away. The pheromone is put by the agents on the pixel. The locus of each agent then becomes a segmented region. When the features of two neighboring regions are similar, they are integrated into a larger region. After several such iterations, the region becomes an object.

The activities of agents in the different regions is given by the following equation

Activity of agents = ---------------------- (1)

Where Disti is the distance which agent i goes in the region and M is the number of agents in the in the region.

In the image segmentation method proposed by Liu and Tang (1999), the goal of the autonomous agent in S (where S is a digital image of discretized two dimensional array of size U X V that contains a number of pixels pertaining to a specific homogeneous segment) is to visit and label all the pixels in the homogenous segment. Agents are designed in such a way that they operate directly on the individual pixels of a digital image by continuously sensing their neighboring regions and checking the homogeneity criteria of relative contrast, regional mean and for regional standard deviation. Based on the sensory feedback from their neighboring regions the agent will accordingly select and execute their behavior. As shown by them earlier (Liu et al., 1997), the agent may breed offspring, move to adjacent pixels or vanish in the image. In this respect they regard the behavior of the agents as being reactive, as it is entirely activated and hence determined by the local environment of the agents.

Melkemi et al (2004) developed a new approach for image segmentation based on multi agent system (MAS) and MRF. A MAS is a group of agents co-operating a local common goal, and among which messages are exchanged in order to co-operate for the achievement of the global goal. In MAS, a set of segmentation agents and a co-coordinator agent are organized in star communication network. The segmentation agent is able to segment image by ICM starting from its own initialization. The co-coordinator agent diversifies these initial configurations by genetic operations (crossover and mutation) in order to accede to good configurations. This model is a hybridization of ICM and Genetic algorithm. Hybridization helps in the task of segmentation intensification.

Melkemi et al. (2005 a) proposed a new chaotic multi agent system (CMAS) in place of MAS (Melkemi et al., 2004) for image segmentation, which improves their previous approach. They introduce a chaotic mapping as a new agent behavior in order to improve the efficiency of MAS. A chaos phenomenon is a set of unpredictable behaviors. The special chaotic characteristic features such as ergodic property, stochastic aspect and dependence on initialization permit this approach to escape the local optimum and converge to a global optimum. Both synthetic and real images have been used to assess the validity and performance of the approach. Experimental results are very encouraging which demonstrate the feasibility, the convergence and the robustness of the method.

A new hybrid Island-Multi agent system for image segmentation was proposed by (Melkemi and Batouche, 2005 b) .Island MAS comprises of a set of agents called Island-agent. The Island-MAS is considered as a distributed hybrid GA in which a population of good initial images is divided into smaller subpopulations called demes and GA is executed on each subpopulation separately followed by ICM which started on its judged good off springs.

At each cycle of evolution of the Island MAS, each island agent: Receives individuals (initial images) from the different Island - agents, performs a GA on current - deme, performs a crossover on peer of parents and performs a mutation on one or several individual. It performs ICM starting from the judged good offspring and Updates the best segmented image. Transmits this new-initial image to the different Island-agents for another segmentation process. The main reason of the popularity of this model is that it can be readily implemented in parallel computers like distributed memory MIMD computer.

Sahba et al. (2007) proposed a new method for image segmentation which used an opposition-based reinforcement learning (RL) scheme. RL means learning by iteration based on interaction with the environment (Sutton and Barto 1998; Singh and Marving, 1996). The RL agent is appropriate for a dynamic environment. In this method, the agent acts for a parameter adjustment to change its environment i.e. the quality of a segmented object. The agent takes some actions as and when it receives an image. It received an objective reward or punishment based on its comparison with the manually segmented version (gold image). The agent tries to learn which action can gain the highest reward. Thereafter, with the accumulated reward, the agent has appropriate knowledge for similar images. The ability of RL agent is that it can be trained using a very limited number of samples and also can gain extra knowledge during the segmentation process. Thus this method performs with less information to than approaches which need good amount of a prior or expert knowledge.

Sahba et al. (2008) applied this (RL) scheme later for the segmentation of prostate in transrectal ultra sound image in the field of biomedical imaging.

Mazouzi et al. (2007) presented a multi agent approach for range image segmentation. The approach consists of using autonomous agents for the segmentation of a range image in its difficult planar regions. Agents move on the image and perform local actions on the pixels, allowing robust region extraction and accurate edge detection. In order to improve the segmentation quality, Bayesian edge regularization is applied to the resulting edges. A new Markov Random field (MRF) model is introduced to model the edge smoothness, used as a priority method in edge regularization. Regions are progressively smoothed by aligning noise pixels to the surrounding planar regions. At the end of the process, region borders consist of thin lines of one pixel wide. The regularization was performed typically for roof edges situated between adjacent regions. The proposed approach aims to improve efficiency and to deal with the problem of result accuracy. Bayesian edge regularization using an appropriate MRF model allows improvement in the segmentation results. The experimental results obtained with real images from the ABW database showed good efficiency of the proposed approach for accurate segmentation of range images.


Many different techniques have been suggested in literature to perform segmentation. These categories include Edge based, clustering based, and region based and split/merge based approaches. The advantage of the edge based approach lies in its is short computation time. The edge grouping process however suffers from a series of difficulties in setting appropriate thresholds and producing continuous, one pixel wide contours. In clustering based approach the difficult threshold setting problem could be avoided by using iterative process. However, this may result in over segmentation. The difficulty experienced in the region growth approach is to set a threshold which is sensitive in measuring similarity. The computation time for pyramidal segmentation is short. Both watersheds and pyramidal segmentation may cause the over segmentation problem. In split /merge process most of the evolutionary algorithms suffer from the slow convergence speed. The autonomous agents enable the optimal extraction of image features and in short less computation time. The agent based image segmentation methods are very feasible, and robust.

With respect to real life applications, the proposed agent based approaches could have segmentation impact on difficult image analysis problems i.e. problem in which conventional edge and contrast enhancement have failed to extract important features. The remote sensing medical images (such as prostate of transrectal ultra sound image, dental radio graphs, Cat-scan(CT )images, and so on) can be optimally segmented with the help of agent based approaches.

Based on the comprehensive literature survey we find that there is considerable scope for improving the accuracy of the segmentation techniques especially for image containing textured backgrounds. Further the time complexity of the segmentation algorithms needs to be simplified or possibilities of implementing these algorithms on a parallel or distributed platform needs to be explored. Also

Hybridization of two or more approaches to take advantages of their best properties can be attempted.