引入上下文信息的可见光遥感图像目标检测与识别方法研究
发布时间：2018-08-12 11:59
【摘要】：可见光遥感图像的目标检测与识别是遥感技术中的重要内容。由于受多种成像因素的影响,目标特征常常存在显著的变化,这使检测识别的难度大大增加。有效利用上下文信息可以提升目标检测识别的效率和性能。本文针对引入多层次上下文信息的可见光遥感图像目标检测与识别方面开展了如下工作:在紧邻背景相对稳定场景的目标检测研究中,针对存在与真实目标具有相似特性的虚警干扰问题,提出了一种引入邻域上下文信息的检测框架。在该框架中,提出了梯度方向二进制模式描述子来表征目标的邻域上下文信息,由于该描述子不需要由码本学习进行特征量化,因此在特征提取方面显著提升了效率,并将该描述子嵌入空间金字塔匹配模型,提高了目标检测性能。在场景变化的目标检测研究中,提出了一种引入目标上下文信息的检测框架。针对目标上下文关系存在着同一种上下文约束关系发生在不同类别目标与候选目标之间、不同的上下文约束关系发生在同类别目标与候选目标之间的现象,我们设计基于词汇的目标上下文表述,再组合使用概率潜在语义分析模型来解决这一问题。在目标类别识别研究中,针对特征提取环节存在误差的问题,提出了一种引入内部上下文信息的识别框架,它首先提取目标的稀疏显著特征,再结合目标内部空间区域信息进行特征表述,最后进行分类识别。该方法增强了特征表述的稳健性,从而提升了识别性能。在遥感图像的地物分类研究中,针对同类地物目标光谱特性变化大、形状和纹理特征分布复杂的问题,提出了一种引入多种上下文信息的地物分类框架。它首先为了获取良好的分割对象,提出了一种基于图模型的层次化分割方法获取对象并用于初始分类,然后提出了一种具有旋转不变性的邻域上下文表述用于优化初始分类结果,最后利用马尔可夫随机场(Markov Random Field,MRF)模型引入邻域对象的相关性约束得到最终分类结果。该方法采用分层递进的策略逐步引入不同上下文信息,克服了直接使用MRF模型时对初始分类结果依赖性较大的问题。
[Abstract]:Target detection and recognition of visible remote sensing image is an important content in remote sensing technology. Because of the influence of many imaging factors, the target features often change significantly, which makes the detection and recognition more difficult. Using context information effectively can improve the efficiency and performance of target detection and recognition. In this paper, the target detection and recognition of visible light remote sensing images with multi-level context information are studied as follows: in the research of target detection in the relative stable scene adjacent to the background, In order to solve the problem of false alarm interference which is similar to the real target, a detection framework based on neighborhood context information is proposed. In this framework, a gradient direction binary pattern descriptor is proposed to represent the neighborhood context information of the target. Because the descriptor does not need to be quantized by codebook learning, it improves the efficiency of feature extraction significantly. The description is embedded into the space pyramid matching model to improve the performance of target detection. In the research of scene change target detection, a detection framework is proposed to introduce target context information. There is a phenomenon that the same kind of context constraints occur between different categories of targets and candidate targets, and different contextual constraints occur between the same category targets and candidate targets. We design target context representation based on vocabulary and combine probabilistic latent semantic analysis model to solve this problem. In the research of target category recognition, aiming at the problem of error in feature extraction, a recognition framework with internal context information is proposed, which firstly extracts sparse salient features of target. Combined with the spatial region information of the target, the feature is expressed, and the classification and recognition are carried out at last. This method enhances the robustness of feature representation and improves the recognition performance. In the research of ground object classification of remote sensing image, a new classification framework is proposed to solve the problem that the spectral characteristics of similar objects vary greatly and the distribution of shape and texture features is complex. In order to obtain good segmentation object, a hierarchical segmentation method based on graph model is proposed to obtain objects and be used for initial classification. Then, a rotation-invariant neighborhood context representation is proposed to optimize the initial classification results. Finally, the Markov random field (Markov Random field MRF model is introduced into the neighborhood object correlation constraints to obtain the final classification results. In this method, the hierarchical progressive strategy is used to introduce different context information step by step, which overcomes the problem of dependence on the initial classification results when the MRF model is used directly.
【学位授予单位】：国防科学技术大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP751
本文编号：2178991
[Abstract]:Target detection and recognition of visible remote sensing image is an important content in remote sensing technology. Because of the influence of many imaging factors, the target features often change significantly, which makes the detection and recognition more difficult. Using context information effectively can improve the efficiency and performance of target detection and recognition. In this paper, the target detection and recognition of visible light remote sensing images with multi-level context information are studied as follows: in the research of target detection in the relative stable scene adjacent to the background, In order to solve the problem of false alarm interference which is similar to the real target, a detection framework based on neighborhood context information is proposed. In this framework, a gradient direction binary pattern descriptor is proposed to represent the neighborhood context information of the target. Because the descriptor does not need to be quantized by codebook learning, it improves the efficiency of feature extraction significantly. The description is embedded into the space pyramid matching model to improve the performance of target detection. In the research of scene change target detection, a detection framework is proposed to introduce target context information. There is a phenomenon that the same kind of context constraints occur between different categories of targets and candidate targets, and different contextual constraints occur between the same category targets and candidate targets. We design target context representation based on vocabulary and combine probabilistic latent semantic analysis model to solve this problem. In the research of target category recognition, aiming at the problem of error in feature extraction, a recognition framework with internal context information is proposed, which firstly extracts sparse salient features of target. Combined with the spatial region information of the target, the feature is expressed, and the classification and recognition are carried out at last. This method enhances the robustness of feature representation and improves the recognition performance. In the research of ground object classification of remote sensing image, a new classification framework is proposed to solve the problem that the spectral characteristics of similar objects vary greatly and the distribution of shape and texture features is complex. In order to obtain good segmentation object, a hierarchical segmentation method based on graph model is proposed to obtain objects and be used for initial classification. Then, a rotation-invariant neighborhood context representation is proposed to optimize the initial classification results. Finally, the Markov random field (Markov Random field MRF model is introduced into the neighborhood object correlation constraints to obtain the final classification results. In this method, the hierarchical progressive strategy is used to introduce different context information step by step, which overcomes the problem of dependence on the initial classification results when the MRF model is used directly.
【学位授予单位】：国防科学技术大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP751
【相似文献】
相关期刊论文 前8条
1 杨凯;姜宏旭;;基于失真敏感性的可见光遥感图像压缩补偿[J];中国空间科学技术;2012年04期
2 刘晓沐;季艳;万寿红;;单源可见光遥感图像中的水域识别[J];仪器仪表学报;2006年S3期
3 赵敏;张荣;尹东;王奎;;一种新的可见光遥感图像云判别算法[J];遥感技术与应用;2012年01期
4 赵英海;吴秀清;闻凌云;徐守时;;可见光遥感图像中舰船目标检测方法[J];光电工程;2008年08期
5 李智勇;董世永;杨效军;;卫星可见光遥感图像异常原因分析方法初探[J];航天返回与遥感;2009年01期
6 产晓冰;周川杰;吕原;;基于应用的可见光遥感图像质量评价方法研究[J];无线电工程;2013年11期
7 陈克;尹东;马婧;张开华;;基于层次反应型agent的可见光遥感图像阴影提取[J];中国科学技术大学学报;2009年08期
8 ;[J];;年期
相关会议论文 前1条
1 晏磊;童庆禧;;可见光遥感作业新途径——以中国贵州无人机超低空遥感试验为例[A];中国感光学会第七次全国会员代表大会暨学术年会和第七届青年学术交流会论文摘要集[C];2006年
相关博士学位论文 前3条
1 郭军;引入上下文信息的可见光遥感图像目标检测与识别方法研究[D];国防科学技术大学;2014年
2 陈雁;可见光遥感图像分割与提取研究[D];中国科学技术大学;2010年
3 廉蔺;红外与可见光遥感图像自动配准算法研究[D];国防科学技术大学;2013年
相关硕士学位论文 前2条
1 于鹏;高分辨率可见光遥感图像舰船目标识别方法研究[D];吉林大学;2011年
2 陈荣;西洋水域水深可见光遥感反演研究[D];南京师范大学;2008年
，本文编号：2178991
本文链接：https://www.wllwen.com/guanlilunwen/gongchengguanli/2178991.html
