基于超顺磁聚类和图割的复杂红外成像目标自动检测方法

项目名称： 基于超顺磁聚类和图割的复杂红外成像目标自动检测方法

项目编号： No.61303192

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 自动化技术、计算机技术

项目作者： 刘松涛

作者单位： 中国人民解放军海军大连舰艇学院

项目金额： 27万元

中文摘要： 目标检测是红外成像自动目标识别的关键环节。本项目的研究目标是建立以无监督学习方式挖掘目标信息的目标检测框架，实现复杂大视场环境下红外成像目标的快速、精确检测。拟开展三项探索：①基于关键点检测与描述、C4聚类、区域筛选和形状匹配方法，从未标记图像集中无监督挖掘目标信息，指导优化图割算法和提供目标的包围盒标记。②集成图像分类和目标定位的快速目标检测方法，包括融合Gist全局特征和显著图局部特征实现快速图像分类，以及改进高效子窗口搜索方法实现快速目标定位。③基于图割的目标精确分割方法，包括正则参数的自动确定，形状、遮挡、灰度不均匀和模糊信息的建模及图割的实时实现。研究方案的特点是将超顺磁聚类和图割引入红外图像处理领域，以无监督学习方式获取目标信息，基于建立的目标检测快速精确实现框架，有效地解决遮挡、模糊、灰度不均匀、训练样本少和精确且高效等检测难题。

中文关键词： 目标检测；图像分割；超顺磁聚类；转移割；高效子窗口搜索

英文摘要： Target detection is a key step for infrared imaging target recognition. The goal for this project is to construct target detection framework through discovering the target information with unsupervised learning, which can realize fast and precise detection for infrared imaging target under the conditions of complicated scene and large field of vision。 Three explorations are planned：① based on key points' detection and description, C4 clustering, region selection and shape matching, the target information will be discovered unsupervisedly from unlabeled image collects for instructing and optimizing graph cut algorithm, and also providing target's labels with Bounding Box.② fast target detection method by integrating image classification and target localization is proposed, in which the fusion of Gist descriptor and saliency map is used for fast image classification, and the efficient subwindow search method is improved for fast target localization. ③ a series of target segmentation methods based on graph cut are presented,including determining regularization parameters automatically, modeling shape, occlusion, gray inhomogeneity and fuzzy information, and using various measures for real-time graph cut implementation.The obvious features for this proposed scheme are summarized as introducing superparamagnetic clus

英文关键词： Target detection；Image segmentation；Superparamagnetic clustering；Transfer cut；Efficient subwindow search