【作者】 李树涛；

【导师】 王耀南； 吴杰；

【作者基本信息】 湖南大学 ， 控制理论与控制工程， 2001， 博士

【摘要】 近年来，多传感器图象信息融合技术在机器视觉、遥感、军事、交通、医学、生物学等领域得到了广泛的关注。多传感器图象信息融合是指对多个传感器采集的关于同一目标或场景的图象进行适当的处理，充分利用多传感器成像间存在的冗余性和互补性，提高图象信息利用效率的技术。 例如，在医学上CT图象具有很高的分辨力，骨骼成像非常清晰，对病灶的定位提供了良好的参照，但对病灶本身的显示较差。而MRI图象虽然空间分辨力比不上CT图象，但是它对软组织成像清晰，有利于病灶范围的确定。如果把它们之间的互补信息综合在一起，那么就能为医学诊断、人体的功能和结构的研究提供更充分的信息。 随着遥感技术的发展，光学、热红外和微波雷达等大量不同传感器对地观测的应用，获取的同一地区的多种遥感影象数据（多时相、多光谱、多分辨率）越来越多。与单源遥感影象数据相比，多源遥感影象数据所提供的信息具有冗余性和互补性。将多源遥感影象数据合理地进行融合，能够产生比单一信源更精确、更完全、更可靠的估计和判断，能够提高影象的空间分辨率和清晰度，提高平面测图精度、分类的精度与可靠性，增强解译和动态检测能力，有效提高遥感影象数据的利用率。 一般认为图象信息融合分为三个层次，即像素级融合、特征级融合和决策级融合。像素级融合是作用于图象的像素点的最底层的融合，其优点是保留了尽可能多的信息，具有最高的精度。其缺点是处理信息量大、费时、实时性差，并且要求传感器是同质的，即传感器产生的数据具有相同的单位。特征级融合是中等水平的融合，它首先将各参与融合的图象进行特征提取，产生特征矢量，而后将这些特征矢量进行融合。其缺点是比像素级融合的精度差。决策级融合是最高水平的融合，它首先对各传感器图象进行特征提取和识别，而后将各自产生的结果进行融合，得到最终的融合结果。其优点是实时性好，对传感器性质没有严格要求，但融合精度最差。 本文主要研究了像素级图象融合的方法及其应用。论文在第一章介绍了图象信息融合的概念、融合的优点、通用融合结构以及目前常用的像素级图象融合方法，包括算术平均、彩色映射、非线性组合、马尔可夫随机场、神经网络、金字塔变换（梯度金字塔、比率低通金字塔、形态学金字塔等）和小波变换等，在第一章还介绍了像素级图象融合的应用。 本文在第二章提出了三种新颖实用的图象去噪算法。第一种是改进型的中值滤波器，即使图象的噪声污染很严重（80W90％）时，该滤波器依然能够很好地恢复有用的图象信号。第二种是基于模糊神经网络的混合滤波器，主要滤波器模块有十字型中心加权中值滤波器、交叉型中心加权中值滤波器和9点中值滤波器，信号经过三种滤波器处理后送入一个训练好的模糊神经网络进行融合处理，得到最终的滤波结果。第三种滤波器是基于神经网络学习的模糊滤波器，传统摸糊滤波器的阈值根据人的经验确定，本文提出通过神经网络来学习以得到最优阈值。实验结果表明所提出的滤波器优于传统的模糊滤波器。 光学传感器（如数码相机）在某一场景进行成像时，由于场景中不同目标与传感器的距离可能不同，这时想使所有目标都成像清晰是很困难的，而采用图象融合技术，即针对不同的目标，得到多幅成像，经过融合处理，提取各自的清晰信息综合成一幅新的图象，能够使图象中的目标都很清晰。在第三章中，论文提出了基于神经网络的多聚焦图象融合方法。具体实现过程概述如下：首先将两幅（或多幅）配准图象进行分块处理，提取两幅图象中对应块的能反映图象清晰度的三种特征，即空间频率、可见度和边缘，将特征归一化后送入训练好的神经网络进行识别，根据得到的结果依据“谁清晰谁保留”的原则构成融合的图象。实验结果表明，该方法对图象配准没有严格要求，即使图象间有一定的像素位移如果图象中目标没有重叠，也能取得很好的融合结果，该方法 n h具有很高的实时性。 近年来，基于小波变换的多传感器融合得到了广泛的关注与研究。但小波变换由于其变换过程中存在2取1的抽样过程，使得小波变换不具备移不变特性。当输入信号有轻微的移动时，所产生的小波变换系数会有很大的变化，这在图象融合领域是不希望发生的。而小波帧变换没有抽样过程，所以它具有移不变的特性。论文在第四章提出了基于小波帧变换及其变种小波帧包变换和树状小波帧变换的多传感器图象融合方法。具体讨论了变换系数活跃度测量、变换系数融合方法、变换系数验证、帧变换分解深度等问题。实验结果表明，在多源图象没有严格地配准、图象间目标有轻微的移动或成像传感器本勇有一定的移动等情况下，小波帧变换仍然能够取得优于小波变换的融合结果。 在图象处理的实际应用中?更多还原

【Abstract】 Image fusion refers to the techniques that integrate complementary information from multi-image sensor data such that the new images are more suitable for the purpose of human visual perception and computer process such as segmentation, feature extraction, and object recognition. Because different kinds of imaging sensors are optimized for somewhat different operating range and environmental conditions, individual sensor may not receive all the information necessary for detecting an object by human or computer vision. Effective combination of such sensors with different features andlor viewing positions could, therefore, extend the capabilities of the individual ones. Numerous applications that would benefit from the use of multiple sensors include display systems in aviation, remote sensing, surveillance, automated machine vision, and medical imaging.In this thesis, some new methods and applications of pixel-level multisensor image fusion are presented.In chapter 1, the fundamental concept, methods and applications of multisensor image fusion are introduced. Some existed methods for image fusion are averaging and weighted averaging, color mapping, nonlinear method, optimization approach, Markov random fields and simulated annealing, artificial neural networks, pyramids, wavelets. The thesis outline is also given in chapter 1.In chapter 2, three novel filters for the removal of impulse noise from corrupted images are proposed. The first is an improved median filter, which can achieve excellent impulse noise cancellation even when the percentage of impulse noise is high. The second is a fuzzy neuron network based hybrid filter, which comprises four basic components: plus-shaped center weighted median filter, cross-shapedcenter weighted median filter, nine pixels median filter, and a fusion center with fuzzy-neuron network. The proposed filter is able to effectively inherit the merits of the used three filters. The third is a fuzzy reasoning filter based on neural network. The output signal of the filter is obtained as the sum of the input signal and nonlinear transformation of the output of fuzzy reasoning, which is obtained by trained neural network.Chapter 3 is concentrated on multifocus image fusion using artificial neural networks. Two spatially registered images with different focuses are decomposed into several blocks. Then, three features reflecting the clear level of every block, i.e., spatial frequency, visibility, and edge, are calculated. Finally, artificial neural networks, i.e., multilayer-perceptron, radial-basis function, probabilistic neural network, are used to recognize the clear level of the corresponding blocks to decide which blocks should be used to construct the fusion result. Experimental results show that the proposed method can perform better than the wavelet transform based method when the objects in the multifocus images are not intersected. Even when the objects in the original images have some motions or the original images are not stringently registered, the proposed method still perform well. And the proposed method can perform real time.It is well known that discrete wavelet transform yields a shift variant signal representation, which means that a simple integer shift of the input signal will usually result in a nontrivial modification of the decomposed coefficients. Thus an image fusion scheme based on the discrete wavelet transform will also be shift dependent, which is undesirable in practical image fusion process(especially considering misregistration problems). In chapter 4, pyramid-structured discrete wavelet frame transform, discrete wavelet frame packet transform, tree-structureddiscrete wavelet frame transform are used to multisensor image fusion. The concrete fusion process, such as activity level measurement, coefficient combination methods, consistency verification and decomposition depth are discussed in detail.Until recently, multiwavelets, which are extension from scalar wavelets, have received considerab更多还原