项目名称: 基于多核学习的高分辨率光学遥感图像固定结构人造目标检测方法研究
项目编号: No.41301480
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: 李湘眷
作者单位: 西安石油大学
项目金额: 19万元
中文摘要: 随着遥感技术的发展,高分辨率遥感图像中人造目标的结构、纹理和细节等信息会表现得更加清楚,可辨识目标的种类也大大增加。本项目以高分辨率光学遥感图像为数据源,在深入分析具有固定几何结构的人造目标的特征基础上,以核方法为框架,构建多类特征对应的基础核,采用层次化的混合方法归纳利用这几类信息,增强多类目标之间的可区分性,建立相应的模型学习理论;针对实际目标检测的效率问题,采用产生式的方法建立新的分类模型,提高解的稀疏性,克服海量数据自动判读这一瓶颈;研究实现多核学习的多类模型构建及学习方法,最终实现复杂场景的光学遥感图像中多类目标同时检测。本项研究可以推动高分辨率光学遥感图像在战场监视、城市测绘方面的应用,促进我国航空遥感的应用水平。
中文关键词: 遥感图像;特征融合;多核学习;图像分类;目标检测
英文摘要: With the development of remote sensing technologies, the high-resolution optical remote sensing images contain more structural and textural information of man-made objects. The number of visible objects categories increase greatly as well. Some high-resolution optical remote sensing images are used in this research. Based on kernel theories and features of man-made objects with fixed structure, several basis kernels are constructed and combined by hierarchical methods, illustrating the effect of each feature type on each object category. The multiple kernel learning (MKL)classification model is bulit as well. Considering the efficiency problems in practical detection tasks, production methods are used to build new classification models, leading to sparser solutions. The MKL multi-class classifier is also constructed. The proposed method is capable of dealing with multi-class man made objects detection problems within complex background. The research can improve the application level of high-resolution optical remote sensing images in the area of battlefield surveillance and urban surveying.
英文关键词: Remote sensing image;Feature fusion;Multiple kernel learning;Image classification;Object detection