Image Segmentation Based On Adaptive Multi-Agents

In image analysise, the high level operations such as feature extraction,object recognition and understanding are influenced by the efficiency and robustness of the selected image segmentation method. As a feature has different exhibition forms, complex distributions and the fuzzy character between different features, image segmentation is a challenge problem. How to segment images accurately and adaptively with the help of computer is a pivotal problem which needs to be solved by image processing.After the investigation of many existing image segmentation methods and based on servel modeling method of complex system, the paper models image segmentation problem in a distributed framework which connects the problem and multi-agent system. A multi-agent segmentation algorithm is implemented based on the framework. Its main contents and innovation include following:(1) Two theories of modeling complex system are discussed in-depth. One is Holland's complex adaptive system including the microcosmic stimulation-reaction model and macrocosmic ECHO model which is enlighted by biology. The other is based on agent theory and technology in distributed artificial intelligence. The complex system is characterized by agent model, their learning and cooperation. It is derived from the observation on human society.(2) An adaptive multi-agent modeling method which combines the advantages above is proposed. It overcomes the simplexity of model in the former and the lack of adaptivity in the later.(3) This method is applied to image segmentation field and a distributed and adaptive image segmentation framework (ISF) is proposed. This framework builds the connection of image segmentation problem and multi-agent systems.(4) The image segmentation framework is mapping to an innovation multi-agent System as its concrete implementation. It based on the behaviors of individual agent which is local and adaptive and the interaction of agents which leads the segmentation of image. The approach needn't configure many parameters and only need run once in contrast to Jiming Liu's agent-based method.(5) Experiments on medical images are conducted. The results indicate the approach can implement the automatic segmentation and it needn't much interaction between human and computer. The segmentation effect is basally consistent with the expert of image segmentation field.