基于Exemplar-Classifier思想的高分辨率光学遥感影像目标识别研究

项目名称： 基于Exemplar-Classifier思想的高分辨率光学遥感影像目标识别研究

项目编号： No.41301361

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 天文学、地球科学

项目作者： 张绍明

作者单位： 同济大学

项目金额： 24万元

中文摘要： 目标识别是高分遥感影像解译的核心环节，漏检率和虚警率是目标识别方法的两项主要评价指标，前者要求好的推广能力，后者要求高的严格性，因此现有方法很难对两者进行兼顾。最新成果表明，基于Exemplar-Classifier思想对目标识别中的模板和分类方法进行集成，可使目标识别方法兼具模板方法的严格性和分类方法的推广性，同时降低漏检率和虚警率。目前该思想只应用于计算机视觉领域，在遥感领域尚无应用。申请者提出将该思想引入高分遥感影像目标识别中，通过两个关键问题的解决实现其应用。首先基于非线性降维、聚类等手段研究样本约简和优化组织方法，解决高分遥感影像成像状态多样性带来的正样本集膨胀和构建困难问题；然后基于尺度和空间信息，建立多层次局部可变模型，解决影像超高分辨率引起的复杂目标描述困难问题。最终，得到一种能够同时改善虚警率和漏检率的高分遥感影像目标识别新方法，这对高分遥感影像自动解译具有重要意义。

中文关键词： 遥感影像；目标识别；中层特征；深度学习；影像分割

英文摘要： Target recognition plays an important role in automatic interpretation of high-resolution optical remote sensing imagery (HR RS imagery). False negative and false positive rates are two major issues for target recognition method. Low false negative rate requires good generalization ability of method and low false positive rate can only be obtained by high threshold. Therefore, it is difficult to make both the false negative and the false positive rate low. The state-of-the-art in target recognition research demonstrated that Exemplar and Classifier can be integrated to achieve this goal. We propose to introduce the Exemplar-Classifier method to target recognition of HR RS imagery. To accomplish this, two issues need to be addressed. Firstly, the various imaging conditions can greatly increase the size of the positive training set, which can lead to huge computation cost due to the large number of classifiers created for each positive sample. Non-leaner dimension reduction and cluster techniques will be employed to reduce the size of the positive training set and optimize the structure of it as well. The second issue is the difficulty in characterizing the complex target due to the high resolution of imagery. A multi-level deformable parts model will be developed to solve this problem. By addressing the above two

英文关键词： remote sensing imagery；object recognition；mid-level feature；deep learning；image segmentation