CASIA OpenIR  > 毕业生  > 博士学位论文
融合目标信息的遥感图像显著性检测方法研究
Alternative TitleSaliency Detection Methods via Fusing Target Information
崔晓光
Subtype工学博士
Thesis Advisor杨一平
2013-12-01
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Discipline计算机应用技术
Keyword显著性检测 Top-down 遥感图像 层次时间记忆模型 字典学习 转向回归核 低秩分解 Saliency Detection Top-down Remote Sensing Image Hierarchical Temporal Dictionary Learning Steering Regression Kernel Low-rank Decomposition
Abstract随着遥感技术的迅猛发展,遥感图像的数据规模越来越大,如何利用有限的计算资源对大幅遥感图像进行实时处理逐步成为研究热点。近年来,认知心理学研究发现人类视觉系统在面对复杂场景时,会迅速将注意力集中在少数几个显著的视觉区域,并优先对这些显著的视觉区域进行处理,这个获取显著性视觉区域的过程被称之为显著性检测。显著性检测可以提供感兴趣的视觉区域,将大部分的无关区域筛除,从而大幅提升后期图像处理的效率。 现有的显著性检测方法多是采用bottom-up机制,用数据驱动的方式获取显著性区域,这类方法能有效的将对底层视觉刺激响应强烈的区域突显出来,然而由于没有考虑目标先验信息,仅通过底层数据难以获得与目标相关的、有价值的感兴趣区域。近年来,top-down显著性检测方法逐步成为研究热点,该类方法通过目标先验信息的引导可以获取到与目标更相关的感兴趣区域。Top-down方法可分为感知模型和计算模型两大类。其中感知模型方法通过调节底层视觉特征的权重引入top-down机制,在自然图像中表现出较好的性能,但面对场景复杂的遥感图像时,简单的调节底层特征图的权重,不足以区分背景和目标。计算模型方法通过提取显著性特征对各区域进行显著性度量,这类方法由于采用了判别性更强的图像特征,其对目标的针对性更强,检测精度更高;然而对于大幅遥感图像,提取显著性特征是极其耗时的过程,计算模型方法通常不能满足遥感图像实时处理的要求。另外,top-down方法需要大量训练样本用以获取特征权重或提取显著性特征,难以应对训练样本不足的应用场景。 针对现有top-down显著性检测方法所遇到的问题,本文分别提出以下三种显著性检测方法,主要工作与贡献如下: 1)提出一种新的感知模型-基于层次时间记忆模型反馈的显著性检测方法,该方法通过层次时间记忆模型引入top-down机制,可以有效的利用目标信息对底层显著图做出引导。该方法模拟了人类视觉注意系统的工作原理,将显著性检测分为注意前期、注意期和注意后期三部分。在注意前期采用Itti模型计算底层显著性图;在注意期,提出一种自适应区域生长的方法用于焦点区域的选取和焦点转移;在注意后期,提出一种改进的层次时间记忆模型,用以对注意焦点区域进行目标概率估计,然后将这个概率反馈至底层显著性图,最终生成融合目标信息的高层显著性图。 2)针对显著性特征提取耗时的问题,提出一种基于字典学习和相关滤波器的显著性检测方法。该方法无需提取显著性特征,具备较好的实时性。传统的相关滤波器方法难以应对类内差异较大的目标检测问题,为了增强相关滤波器的检测性能,采用字典学习的方法对训练图像构建一个滤波器字典,通过滤波器字典自适应的针对测试图像生成一个相关滤波器。相比只用一个特定的相关滤波器的常规方法,本章方法能够显著提高检测性能,同时具备较好的实时性。 3)针对训练样本不足的应用场景,提出一种基于局部非对称转向回归核的显著性检测方法。该方法着重解决少量训练样本下的显著性检测问题。提出基于局部非对称转向回归核的特征提取方法,相比传统转向回归核方法,其提...
Other AbstractAlong with the rapid advance in remote sensing technology, the data of remote sensing image is getting bigger and bigger, as a result, the real-time processing for large remote sensing images within limited computing resources is becoming a hot research topic gradually. Recently, cognitive psychology discovered that, when facing complex scenes, human visual system can rapidly pay attention to few salient regions and deal with the corresponding information preferentially. The process which aims at obtaining salient regions is called saliency detection. By means of saliency detection technology, valuably concerned visual scenes are offered while a large proportion of irrelevant part are sieved, which results in the obvious improvement of image processing efficiency. In the exciting methods of saliency detection, the bottom-up mechanism is usually applied, and the corresponding salient region is captured by data driving. These exciting methods can efficiently highlight the areas which have strong response to the low-level vision stimuli, however, since the priori information of the target is not concerned in these methods, it is unavailable to obtain the valuably interested area which is correlative with targets only by means of the underlying data. Recently, top-down saliency detection method is becoming a hot research field, by this method, more interested area correlative with targets can be captured. Top-down method includes two following models, perceptual model and calculation model. In the perceptual model, top-down mechanism is introduced by adjusting the weight factor of low-level vision feature. This model can be employed to deal with the natural image efficiently, however when facing the remote sensing image with complex scene, it is difficult to distinguish the target from its background only by means of adjusting the corresponding weight factor. While in calculation model, the saliency of each area is measured via extracting salient features. Compared with the perceptual model, since the image feature with strong discriminant is employed in this method, higher detection accuracy can be obtained, and the target can be focused more strongly. However, in case of the large remote sensing image, it is a time-consuming process to extract salient feature, as a result, this method can not satisfy the requirement of real-time processing. Additionally, in order to capture the weight of feature maps or salient features, a large number of training sampl...
shelfnumXWLW1961
Other Identifier201018014629081
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/6572
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
崔晓光. 融合目标信息的遥感图像显著性检测方法研究[D]. 中国科学院自动化研究所. 中国科学院大学,2013.
Files in This Item:
File Name/Size DocType Version Access License
201018014629081_崔晓光.(14798KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[崔晓光]'s Articles
Baidu academic
Similar articles in Baidu academic
[崔晓光]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[崔晓光]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.