Clustering Based On Multi-agent Evolutionary Algorithm And Its Application In Image Segmentation

Image segmentation is to divide an image into different targets, so that these marked targets could be utilized in the following processes of compute vision, such as the image target recognition and the image analysis. With the increasing application of digital images, image segmentation, a critical step of digital image processing, has attracted great attention. Now, plenty research work has been done on this domain and massive image segmentation methods has been proposed. Among these methods, it is a hot field to segment an image based on clustering technology.In this paper, our works mainly focus on the application of clustering on image segmentation, and two clustering methods based on multi-agent evolutionary algorithm are proposed. It is an effective and popular way to treat clustering problem as an optimization task, and these two proposed clustering method in this paper are designed to use the global optimal search ability of the multi-agent evolutionary algorithm. Following is our work in this area:1. Most of the clustering methods have the drawback of setting the initial parameter values, especially the number of clusters. This drawback has greatly hindered the application of clustering in image segmentation. To solve this problem, we proposed a new clustering method that combines the multi-agent evolutionary algorithm(MAEA) and the ISODATA clustering algorithm. In the conventional ISODATA algorithm, clusters will be split and merged iteratively, so that the number of clusters and the cluster centers could be adjusted. Inspired by this strategy, we designed the splitting operator and the merging operator. Then these two operators are added into the MAEA algorithm to utilize the global optimal searching ability of the MAEA algorithm. Consequently, the proposed algorithm could automatically adjust the number of clusters and avoid being trapped in local optimal.2. The clustering validity is usually used as the fitness function in the clustering methods that base on optimal technologies, and massive clustering validities have been proposed now. However, each clustering validity could only consider certain propertiesof the clustering result. Hence, it is more suitable to optimal two clustering validities at the same time. Based on this, a clustering method based on the multi-objective MAEA algorithm is proposed in the paper. Since two different clustering validities are utilized, the clustering result could be effectively evaluated. Besides, the multi-agent framework could guarantee the diversity of the population and hence avoid the algorithm trapped in local optimal. Therefore, the proposed method has a well robustness and global searching ability.