【作者】 焦竹青；

【导师】 徐保国；

【作者基本信息】 江南大学 ， 控制理论与控制工程， 2011， 博士

【摘要】 多源图像融合是将源自同一场景或目标的多幅图像合成一幅新的图像,以获取对该场景或目标更为准确和全面的描述。图像融合过程充分利用了不同源图像之间的冗余信息和互补信息,使融合后的图像具有更高的可信度、较少的模糊和更好的可理解性,更适合于人类视觉感知或计算机检测、分类和识别。多源图像融合主要分为空间域图像融合和变换域图像融合。变换域融合方法利用了多尺度、多分辨率分析方法在表征信号局部特征上的优点,弥补了空间域图像融合方法在细节表现力上的不足。但是这类融合方法改变了变换系数,若融合规则选取不当,容易造成源图像信息的损失。本论文的研究工作主要围绕变换域中的多源图像融合展开。从有利于场景理解和目标识别的角度出发,在分析变换域图像融合理论和先进方法的基础上,寻找能够保留源图像更多有用信息,有效提高融合图像质量的新方法。论文的主要研究内容如下:(1)从图像融合的层次架构出发,探讨了多源图像融合的基本流程和基本方法,归纳了图像融合效果的主要评价指标及其选取原则;分析了基于多分辨率分析的变换域图像融合技术,并对几种典型变换域融合方法的融合效果进行了比较。(2)在研究多源图像融合及非下采样Contourlet变换理论的基础上,提出了一种基于区域特征的多聚焦图像融合方法。为了有效获取源图像的边缘和细节信息,利用非下采样Contourlet变换对图像进行多尺度分解,再根据所得到的子带系数的区域特征及其接近度分别采用相应的融合规则。该方法的融合效果优于传统的空间域融合方法和变换域像素级融合方法。(3)结合脉冲耦合神经网络,对变换域多源图像融合方法进行了研究。在多聚焦图像融合中,利用非下采样Contourlet变换有效捕捉源图像的特征信息,并根据脉冲耦合神经网络点火映射图的邻域接近度,为非下采样Contourlet变换获得的子带系数提供融合规则。根据红外图像与可见光融合的要求,利用自约束限定函数定义神经元链接强度,并引入脉冲耦合神经网络的迭代过程,在此基础上对非下采样Contourlet变换的子带系数进行融合处理。提出的相应融合方法针对多聚焦图像及红外与可见光图像,均能够有效提高融合图像的质量。(4)研究了变换域多源图像融合中的多目标优化问题。在分析多目标优化理论与算法的基础上,提出了一种自适应差分进化算法。该算法采用自适应变异因子、动态交叉概率函数以及精英最优的排序策略,具有良好的搜索能力和收敛性,将其应用到基于非下采样Contourlet变换的多源图像融合多目标优化中,有效解决了图像融合过程的综合评价问题。更多还原

【Abstract】 Multi-source image fusion will integrate multiple images derived from the same scene or target collected into a new image to obtain more accurate and more complete description about the scene or target. Image fusion process can take advantage of the complementary information and redundant information in different source images, so as to obtain a fusion result with higher reliability, less blur and better intelligibility. As a result, the fusion image is more suitable for human vision perception and computer processing, such as detection, classification and identification.Multi-source image fusion can be divided into space domain image fusion and transform domain image fusion. The transform domain fusion methods utilize a multi-scale, multi-resolution analysis on the expression of the advantages on local signal characteristics, so that it makes up the lack of expression in detail for space domain fusion methods. However, such fusion methods change transform coefficients, which is likely to cause loss of source image information in inappropriate fusion rules.In this thesis, most of the research works focus on multi-source image fusion in transform domain. From the benefit point of scene understand and target identification, this research is to analyze the transform domain image fusion of basic theory and advanced algorithms, and look for the new ways to retain more useful information of source images and improve fusion image quality effectively. The main contents are as follows:(1) From the level structure of image fusion, the basic process and basic methods of multi-source image fusion are discussed, while the evaluation criteria and their selection principles of image fusion effect are summarized Then the transform domain image fusion technology based on Mmulti-resolution analysis is analysized, and the fusion effect of several typical fusion methods in transform domain are compared by experiments.(2) On the basis of researching multi-source image fusion and non-subsampled Contourlet transform theory, a multi-focus image fusion method is proposed based on regional characteristics. In order to effectively obtain the edge and detail information, the source images are decomposed under multi-scale by use of nonsubsampled Contourlet transform, while the corresponding fusion rules are employed to according to the regional characteristics and approach degree of subband coefficients. The fusion effects of this method are better than those of traditional space domain fusion methods and transform domain fusion methods on pixel level.(3) Combined with pulse coupled neural network, the multi-source image fusion methods in transform domain are researched. In multi-focus image fusion, the non-subsampled Contourlet transform is used to the capture feature information of source images, and according to the neighborhood approach degree in ignition mapping images of pulse coupled neural network, the fusion rules are provided to sub-band coefficients from non-subsampled Contourlet transform. According to the requirements of infrared and visible light image fusion, the self-constraint restrictive function is used to define neuron link strength, which is introduced into iteration of pulse coupled neural network. On this basis, fusion process is performed on the subband coefficients generated under non-subsampled Contourlet transform. The proposed corresponding fusion methods for multi-focus image and the infrared and visible images are able to effectively improve the quality of fused images.(4) The multi-objective optimization problem of multi-source image fusion is researched in transform domain. Based on the analysis of multi-objective optimization theory and algorithms, an adaptive differential evolution algorithm is proposed. With adaptive variance factor, dynamical crossover probability function and optimal elite ordering strategy, the algorithm reflects not only good search capability but also good convergence. When applied to multi-objective optimization of multi-source image fusion of transform domain, it will be an effective solution to the comprehensive evaluation in the image fusion process.更多还原