【作者】 蒋年德；

【导师】 王耀南；

【作者基本信息】 湖南大学 ， 控制科学与工程， 2010， 博士

【摘要】 随着传感器技术的发展,多传感器图像数据类型和数量急剧丰富,多传感器图像融合技术在军事目标识别、智能机器人、遥感、医学图像处理和制造业等诸多领域内得到了广泛的应用,更成为了图像理解、计算机视觉和遥感等领域的研究热点。论文在多尺度分析理论指导下对像素级多传感器图像融合方法开展相关的研究工作,融合图像包括了遥感图像、多聚焦图像、红外与可见光图像和医学图像等多种类型。论文首先对研究背景、图像多尺度分析概念、图像融合技术及其相关概念、图像融合技术发展现状和应用等进行了总体介绍。总结了现有的多传感器图像融合方法和国内外学者的研究成果,然后将已有的图像融合方法进行了分类并分析了它们的优缺点。对基于多尺度变换的融合方法原理和已有的相关研究工作进行了较为详细的阐述,给出了本论文所用到的融合图像质量客观评价指标计算公式。借鉴前人的研究工作,论文重点研究了基于多尺度变换的图像融合方法,包括基于金字塔变换、基于小波变换、基于Curvelet变换和基于非下采样Contourlet变换的图像融合方法,然后提出了几种新的融合方案,给出了它们的融合规则函数表达式,分别使用遥感图像、多聚焦图像、红外与可见光图像以及医学图像进行了实验验证。论文对图像融合方法和融合规则都做了一些研究,主要成果和创新点概括如下：(1)根据视觉对比度掩蔽特性,考虑分解系数中的局部均值和窗口内每个系数,基于金字塔多尺度变换提出了改进的图像对比度融合方法。(2)通过研究源图像小波分解子带系数邻域内的相关性,基于小波变换提出了邻域内相关系数与平均梯度的图像融合方法,该方法具有较好的融合效果。(3)通过研究主分量变换融合方法和小波变换融合方法各自的特点,提出了一种基于主分量变换与小波变换结合的自适应遥感图像融合方法。该方法有效地融合了高分辨率图像的低频分量信息,消除了小波变换融合方法图像中的分块效应,其融合图像边界清楚,在保留光谱信息的同时空间细节信息也得到提高。(4)针对小波变换不能够有效地获取图像中的几何特征(如曲线)等奇异性问题,研究了Curvelet变换理论,然后根据Curvelet变换系数特点,考虑图像中的弱边缘,基于第一代Curvelet变换提出了一种自适应的遥感图像融合方法,该方法的融合图像较好地保留了源图像的光谱信息,其空间细节信息也得到了增强。(5)Curvelet变换很好地刻画了图像中曲线等奇异性,图像的纹理活跃性程度可通过不同方向上的Curvelet系数能量变化来反映。综合考虑局部窗口内每个Curvelet系数能量及其均值,基于第二代Curvelet变换提出了改进的系数能量对比度融合方法。实验证明,对于多聚焦图像,其融合图像与源图像的信息相关程度较高并且差异性小,很好地保留了源图像的边缘特征。(6)针对传统IHS (Intensity Hue Saturation,IHS)变换融合方法的光谱信息丢失问题,结合第二代Curvelet变换提出了一种改进的融合方法,两组不同的遥感图像实验及分析表明,该融合方法充分综合了IHS变换和Curvelet变换优势,有效地提取了原始遥感图像特征。(7)通过研究多尺度变换域系数注入融合遥感图像方法,考虑到注入融合方法使多光谱图像中某些特征被全色图像的特征掩蔽,提出了改进的非下样Contourlet变换系数注入融合方法,实验证明,该方法尽可能保留了多光谱图像中的光谱信息和高分辨图像中的细节信息。对论文提出的每种融合方法,都给出了相应的实验,从主观和客观两个方面对提出的融合方法进行了分析评价,得到了一些有价值的结论。实验结果及分析表明,所提出的融合方法可以有效地实现多聚焦图像、医学图像、红外与可见光图像以及遥感图像的融合,论文提出的融合方法对解决图像融合问题具有重要的指导意义。更多还原

【Abstract】 With the development of sensor technology that brings the abundance of image sources, multisensor image fusion technique has been attracting a large amount in a wide aritey of applications such as military target recognition, intelligent robot, remote sensing, medicine image processing, manufacturing etc. In the mean time, multisensor image fusion technique has been a research hotpot in such area as image understanding, computer vision and remote sensing and so on. The mainly goal of this dissertation is to research fusion methods for multisensor image in pixel-level under the guidance of multiscale analysis theory. The fused objects include remote sensing images, multifocus images, infrared and visible light images and medicine images.Firstly, this dissertation introduces the research background, the concept of the image multi-scale analysis, and the related concepts, development statue and applications of image fusion technology. Secondly, the dissertation reviews the existing multisensor image fusion methods, and summarizes the research works of domestic and foreign scholar for multisensor image fusion. It also classifies the image fusion methods in various ways and briefly analyses their advantages and disadvantages. Thirdly, the dissertation gives the image fusion principle based on multiscale transform and the existing related research works in detail. Furthermore, it gives the objective assessing indexes formula of fused image quality used in this dissertation. Finally, the dissertation mainly focuses on the research of multisensor image fusion methods based on multiscale transform, which include pyramid transform, wavelet transform, Curvelet transform and nonsubsampled Contourlet transform. By means of previous research works, it proposes several new fusion schemes and gives the function expression of fusion rules, respectively. To verify the fusion schemes, the experiments have done with remote sensing images, multifocus images, infrared and visible light images and medicine images.This dissertation studies the image fusion methods and some fused rules. The main contributions of this dissertation are summarized as follows:(1) According to the feature of visual contrast masking, a new fusion method using improved image contrast based on multiscale pyramid transform is presented, which has considered each decomposition coefficient and the mean of all in a local window. (2) By studying the correlation between neighborhood subband coefficients of the source images that is decomposed by wavelet transform, a new image fusion method is proposed by using neighborhood correlation coefficient and average gradient based on wavelet transform, which has better fusion result in some problems.(3) After studying the characteristics of fusion method based on principal component transform and wavelet transform respectively, an adaptive remote sensing image fusion method based on principal component transform combined with wavelet transform is proposed. The low frequency information of highresolution image is effectively injected into the fused image. The proposed method eliminates the blocking effect of fusion image based on wavelet transform. In addtion, the border of fused image is clear. When preserving spectral information; spatial detail information is also improved.(4) The wavelet transform could not obtain effectively the geometric features and the singularity of image, such as curves. It maybe affects the fused image and the fusion result. After studing Curvelet transform theory and according to the characteristics of Curvelet transform coefficient, an adaptive fusion method is presented to fuse remote sensing image based on first-generation Curvelet transform, which takes the weak edge of images into consideration. The fused image not only preserves spectral information of the original multispectral image well, but also enhances spatial detail information largely.(5) Curvelet transform depictures the curve singularity of image well, so the activity level of image texture could be reflected by the changes of Curvelet coefficient energy in different directions. Considering the each energy of Curvelet coefficient and the mean of all in a local window, a new image fusion method using improved energy contrast based on second-generation Curvelet transform is presented. For the fusion of multifocus images, the experimental results show highly correlated informations and small difference between the fused image and source images. Besides, the fused image prserves the features of source image well.(6) The spectrum information may be lossed using traditional IHS transform to combine remote sensing image, so a new image fusion method is proposed by using the second-generation Curvelet transform to improve IHS transform. Two different remote sensing images experiments and analysis show that the propsed fusion method takes advantage of the features of IHS transform and Curvelet transform, and is able to extract the feature of original remote sensing images effectively.(7) By studying the fusion method using coefficient injection for remote sensing images based on wavelet transform will be bring the problem that the panchromatic image features could mask some features contained initially in multispectral images, a a novel and effective fusion method with improved coefficients injection based on nonsubsampled Contourlet transform is proposed. The experiments show that the method is as much as possible to retain the spectral information of multispectral image and the detail information of highresolution image.Each image fusion method is correspondily experimented on in this dissertation. It analyse the performance of proposed fusion methods from the subjective and objective aspects of the fused image and obtained some valuable conclusions. Experimental results and analysis show that the proposed fusion method can realize the integration of multifocus image, infrared and visible light images, medicine images and remote sensing images. The proposed image fusion methods have important guiding significance to solve the image fusion problemt.更多还原