| Remote sensing image change detection is to observe and analysis the same area at different times of two or more remote sensing images, and get the differences between the images by comparing. Then according to the difference image, the change information of the region over time is detect. Currently the general procedure of image change detection algorithms is first to generate the difference image of registered images, and then to classify the difference image. The difference image is divided into two categories, changed and unchanged regions. Clustering is one of method of classifying the difference image. Aiming at the deficiency of existing clustering method in image segmentation, this paper presents an unsupervised change detection method for synthetic aperture radar(SAR) images based on a multi-objective fuzzy clustering algorithm and selective ensemble strategy.1. The first work puts forward a multi-objective clustering algorithm to analysis the difference image. The algorithm establishes two complementary clustering objective functions to evaluate the performance of the clustering algorithm. In the second objective function we introduce gray level difference between neighborhood pixels and the center pixel and weighted Euclidian distance as similarity measure for clustering algorithm, combining spatial neighborhood information in the process of clustering. Compared with the traditional single objective clustering algorithm, the proposed method can reduce the effect of speckle noise on clustering results effectively, without the loss of detail. Due to solving two objective functions at the same time, the proposed method avoids the selection of parameters. The purpose of reducing the effect of speckle noise and enhancing the cluster performance at the same time is achieved in the process of image segmentation or classification. Use the randomly generated initial antibody population instead of the initial clustering center, the propose method reduces the sensitivity of the traditional clustering method in initializing clustering centers. Single objective clustering algorithm run multiple times to get a different solution, multi-objective method only needs run one generation.2. In this work, the selective ensemble strategy is introduced to integrated intermediate image segmentation results, the results of multi-objective clustering can be regarded asa homomorphism classifier of different weights. The final integration result is better than single clustering results. The selective ensemble strategy introduced in this paper is to select a subset of the ordered intermediate image classification results for aggregation. First, the proposed method combines all the intermediate clustering results using a simple majority vote, then reorder the intermediate component individual image classification results using the set of class labels generated in first step as criterion and then select the top 15% to 30% to combine. The output of a multi-objective optimization clustering is a set of mutually non-dominated clustering solutions which correspond to different trade-offs between the two objectives, the result is a set of clustering centers. Using the set of clustering center, we can get different segmentation results. From the perspective of multi-objective optimization, the set of mutually non-dominated solution. Most clustering algorithms introduce third-party solution selection strategy to choose one of them, which is equivalent to add a clustering objective function. Some strategy need to assume that 10% of the experimental data is known, which is not practical for image clustering segmentation. The method can obtain final segmentation result directly and do not need to use third-party solution selection strategy or use the prior knowledge. The final change detection results shows lower error, the method adopted in this paper obtains better segmentation result than other existing methods on change detection in synthetic aperture radar(SAR) images. |
PDF Full Text Request