Semantic-based high resolution remote sensing image retrieval.

摘要

High Resolution Remote Sensing (HRRS) imagery has been experiencing extraordinary development in the past decade. Technology development means increased resolution imagery is available at lower cost, making it a precious resource for planners, environmental scientists, as well as others who can learn from the ground truth. Image retrieval plays an important role in managing and accessing huge image database. Current image retrieval techniques, cannot satisfy users' requests on retrieving remote sensing images based on semantics. In this dissertation, we make two fundamental contributions to the area of content based image retrieval.; First, we propose a novel unsupervised texture-based segmentation approach suitable for accurately segmenting HRRS images. The results of existing segmentation algorithms dramatically deteriorate if simply adopted to HRRS images. This is primarily clue to the multi-texture scales and the high level noise present in these images. Therefore, we propose an effective and efficient segmentation model, which is a two-step process. At high-level, we improved the unsupervised segmentation algorithm by coping with two special features possessed by HRRS images. By preprocessing images with wavelet transform, we not only obtain multi-resolution images but also denoise the original images. By optimizing the splitting results, we solve the problem of textons in HRRS images existing in different scales. At fine level, we employ fuzzy classification segmentation techniques with adjusted parameters for different land cover. We implement our algorithm using real world 1-foot resolution aerial images.; Second, we devise methodologies to automatically annotate HRRS images based on semantics. In this, we address the issue of semantic feature selection, the major challenge faced by semantic-based image retrieval. To discover and make use of hidden semantics of images is application dependent. One type of the semantics in HRRS image is conveyed by composite objects. Composite objects are those consisting of several individual objects that form a new semantic concept. We exploit a hyperclique pattern discovery method to find co-existing individual objects that form a new concept. We convert the identified groups of co-existing objects as new feature sets and feed them into the statistical learning model for better performance in image annotation. Experiments with real-world datasets show that, with new semantic features added, we can improve the performance of composite object discovery.