Multi-Feature Ensemble Based SAR Image Segmentation

With the rapid development of synthetic aperture radar (SAR) system, a great number of SAR image data is acquired. On the other side, the research on theory and technique of SAR image understanding and interpretation has fallen behind. SAR image segmentation is one of the key techniques for SAR image automatic interpretation, and is a current research hotspot attracting lots of scholars’attention. Comparing with other kinds of images (e.g., optical images, thermal infrared images, and so on), SAR images own their unique properties, which bring great challenge to the problem of SAR image segmentation. For example, SAR images are in some degree contaminated by speckle noise; SAR images usually consist of multi-scale objects even in the same scene; the exhibition of the same terrain target in SAR images is often nonstationary and has complex variation. Therefore, the research on SAR image segmentation is of great significance for the development of SAR system.Considering the characteristics of SAR images, this dissertation investigates SAR image segmentation from the feature extraction to the algorithm design, and proposes a few SAR images segmentation algorithms based on multi-feature ensemble, which can be summarized as follows:1. A context based hierarchical unequally merging for SAR image segmentation is proposed, which combines the top-down fashion and bottom-up fashion. Based on the Gestalt laws, three rules are proposed to represent superpixel context that realize a new and natural way to manage different kinds of features extracted from SAR images. The rules are prior knowledge from cognitive science and serve as top-down constraints to globally guide the superpixel merging. The features, including brightness, texture, edges, and spatial information, locally describe the superpixels of SAR images and are bottom-up forces. While merging superpixels, a hierarchical unequally merging algorithm is designed, which includes two stages:a) coarse merging stage; and b) fine merging stage. The merging algorithm unequally allocates computation resources so as to spend less running time in the superpixels without ambiguity and more running time in the superpixels with ambiguity. Experiments on synthetic and real SAR images indicate that this algorithm can make a balance between computation speed and segmentation accuracy.2. A fuzzy clustering based on subspace iterative optimization is proposed, which is characterized by following aspects:1) multiple features have been extracted to accurately describe the objects in SAR images.2) A novel similarity measure using hierarchical ensemble is presented, which integrate these features of different properties respectively in the feature level and the similarity level to avoid the mutual influences between features and maximize discrimination ability of the similarity measure between objects.3) A novel spatial neighbourhood term with preferences has been introduced to the objective function, which can compute the contextual relationship between superpixels more accurately and reasonable than the traditional average methods.4) According to the objective function, a subspace iterative optimization algorithm is designed,which can optimize the objective function iteratively in each feature subspace to avoid the mutual influences between different kinds of features and preserve the distribution structures of the data points in every subspace.3. Immune clonal selection optimization is carefully studied, and we propose two novel immune clonal selection optimization algorithms, which are:immunologic regulation clonal selection algorithm and clonal selection optimization based on orthogonal experiment design. Then, considering the advantages of the alternating minimization and the clonal selection optimization, we combine the two kinds of methods into a framework. The former one can iteratively search along the direction of the gradient steepest descent so as to accelerate the convergence speed of the algorithm; while the latter one can avoid premature convergence and find the global optimal solution with high probability so as to improve the robustness and accuracy of the algorithm.Experiments indicate that the proposed hybrid optimization method can achieve the best performance and is stable and effectiveness for various kinds of SAR images.4. A context based unsupervised hierarchical iterative algorithm for SAR segmentation is proposed, which combines the advantages of the cluster based segmentation algorithms and region growing based segmentation algorithms.While merging superpixels,this algorithm chooses a hierarchical iterative strategy:a modified fuzzy c-means algorithm is first designed to analyze the appearance-based features of superpixels, and then region iterative growing is used to merge the similar superpixels based on contextual analysis in space domain.After that, a new loop of the clustering algorithm and the region growing begins. These two iterative sub-algorithms perform hierarchically and realize a natural and effective way to use different kinds of information to segment SAR images. Experiments on synthetic and real SAR images indicate that the proposed algorithm can obtain excellent segmentation and make a good balance between region consistency and preserving image details. 5. A two-stage algorithm for SAR image segmentation is presented by combing the advantages of the image-space based method and the feature-space based method. The algorithm consists of two stages:1) coarse merging stage, and2) fine classification stage. A context-based region iterative merging (CRIM) algorithm is proposed in the coarse merging stage, which can use many different kinds of information to guide the superpixel merging. The advantage of CRIM is that it can merge most of the superpixels inside the true segments at a very fast speed, so as to improve the accuracy of features and reduce the computation burden for the fine classification stage. Comparing with the traditional region merging algorithms, CRIM can take use of much more information, and a global halt index (HI) is carefully designed to decide the output of CRIM, which has global consistency in homogeneity. In the fine classification stage, the designed fuzzy clustering using hybrid optimization is adopted to produce the final segmentation result, which realizes the combination of the gradient-descent-direction search and the heuristic search. The former one can accelerate the algorithm’s convergence speed; while the latter one can avoid premature and find the global optimal solution with high probability.