【作者】 宋乐；

【导师】 林玉池；

【作者基本信息】 天津大学 ， 测试计量技术及仪器， 2008， 博士

【摘要】 异源图像融合是协同使用不同类型的图像传感器,并将各种图像信息有效地结合起来,形成高性能感知系统来获取对同一目标的一致性描述的过程。该技术从多信息的视角进行处理,可综合各图像之间的互补信息和冗余信息,扩大传感器的工作范围,增加置信度、改善系统的可靠性和可维护性,降低对单一传感器的性能要求,可更为准确、可靠、全面地获取对目标或场景的信息描述,不仅可以使处理后的图像更适合人的视觉观察,而且可以为进一步的图像处理提供更有效的信息。将异源图像中具有代表性的红外图像和可见光图像作为研究对象,结合红外技术、图像处理与人工智能理论,分析、吸收国内外关于图像融合的成功经验,设计了基于空间域和变换域的两类新型异源图像融合算法,并建立了一种新型的融合图像质量综合评价标准,在此基础上,建立了闭环式图像自适应融合体系,可由系统自动选择符合用户需求的融合方案,其分层式处理结构保证了融合算法的快速性和准确性,增强了图像融合的自动化程度。首先研究了基于空间域的异源图像融合方法。对现有的空间域融合方法进行了分析,如加权平均融合法、灰度极值法等,通过融合实验,比较了各自的优缺点。在此基础上,提出了一种基于K-L变换(Karhunen-Loeve Transform)的自动加权融合算法。它保持了现有的加权平均融合法的快速实用性,通过主成分分析,将源图像中的有用信息集中到相互独立的新主成分矩阵中,由此确定加权平均融合方法中源图像的加权系数,以实现更加理想的融合效果。实验结果表明,基于K-L变换的自动加权融合算法的效果优于现有的常用空间域融合算法。重点研究了基于变换域的异源图像融合方法。对变换域中典型的多尺度分析理论进行了论述,对其中具有代表性的小波变换技术及其在图像融合中的应用做出了详细说明,结合具体的实验,讨论了不同小波基、不同分解层数和不同融合规则对融合效果的影响,分析了小波变换融合法的优势和局限性。在此之上,重点讨论了各向异性多尺度分析融合法,以NSCT(Non-sampled Contourlet Transform,非下采样Contourlet变换)为主要工具,给出了该方法在图像融合中应用的特点,并以此为基础,设计了全新的融合规则,提出了一种基于NSCT和PCNN(Pulse Couple Neural Network,脉冲耦合神经网络)的异源图像融合方法,充分利用了NSCT的各向异性、多方向性和平移不变性,消除了以往多尺度分解带来的频率混叠现象,又结合了PCNN的全局耦合特性,利用其特有的生物学背景,提高融合图像的整体视觉效果。实验结果表明,使用该方法得到的融合图像清晰自然,适合人眼观察,其融合效果优于现有的小波变换法等变换域融合算法。研究和分析了异源图像融合质量的综合评价方法。在介绍了现有主、客观图像评价方法的基础上,对已有的多种评价指标进行了分类,讨论了不同融合目的下的指标选取规则以及单个评价指标的局限性;而后,重点研究了现有的多评价指标综合化方法,比较了几种典型方法的优缺点。并在此基础上,提出了一种基于FNN(Fuzzy Neural Network,模糊神经网络)的图像融合质量综合评价方法,结合了模糊逻辑推理的结构性知识表达能力和神经网络的自学习能力,将多种典型的图像融合客观评价指标进行模糊化,以主观评价结论作为先验知识,通过网络学习自动生成评价指标权重等相关参数,并通过动量因子提高了网络的学习效率。实验结果表明,该方法可对图像融合质量进行全面、准确的评价,在很大程度上克服了单纯由人眼判决产生的主观性和单因素客观评价指标的片面性。研究了异源图像融合自适应体系。基于上一章中给出的图像融合综合评价方法,建立了一种闭环式的自适应图像融合模型,以融合算法的复杂度为标准进行分层,将多种特性不同的图像融合算法进行串行排列,由系统自动选取符合用户需求的融合方案,克服了开环融合系统灵活性差的缺点。此外,模型内的算法可灵活添加,具有很好的可扩展性。为了验证模型的有效性,选择加权平均融合法、小波变换融合法、Contourlet融合法和NSCT-PCNN融合法为算法集进行融合实验。实验结果表明,该体系建立了融合结果与用户需要之间的有机联系,融合过程中无需人为参与,图像融合的自动化程度得到了加强。更多还原

【Abstract】 Different-source image fusion is the process of forming a high-performance sensing system by collaborating different types of image sensors and gathering all kinds of image information effectively in order to obtain a coherent description of the same object. This technology processes the information from multiple perspectives and combines complementary information and redundant information of several different images. The working available range of sensors can be expanded. The reliability and maintainability of the sensing system can be improved. The requirements of the single sensor can be reduced. The description of information of the target or the scene can be more accurate, reliable and comprehensive, which not only benefit the visual observation process by human, but also provide more effective information for further image processing.The infrared and visible images are selected as the study objects, which are representative images from different source images. Based on infrared technology, image processing and artificial intelligence, two kinds of image fusion algorithms for different source images, including spatial domain and transform domain, are designed with the help of analyzing and absorbing the successful experience from home and abroad. A novel method for image fusion quality evaluation is established. According to this newly-designed criterion, a closed-loop adaptive image fusion scheme is proposed. The scheme can select the most optimized fusion program automatically. The hierarchical structure ensures the speed and accuracy of the fusion algorithm. The degree of automation is thereby enhanced.An image fusion algorithm based on principle component analysis is presented. Current image fusion algorithms of spatial domain are analyzed and compared, such as weighted average method and gray-level extreme value method. After that, an automatic weighted average method based on K-L (Karhunen-Loeve) transform is proposed. This algorithm preserves the practicability of the current weighted average method. Through the principle component analysis, useful information is concentrated into the independent new matrixes. The weighted coefficients of original images in weighted average method are calculated correspondingly, so that a better fusion result is achieved. Experimental results show that on the use of this image fusion algorithm based on K-L transform, the fused image is superior to the images of the existing fusion method of spatial domain.Image fusion methods of transform domain is researched. The typical multi-scale analysis theory in transform domain is introduced, especially the representative wavelet transform technology and its application in image fusion. Experiments are carried through in order to deduce the advantages and disadvantages of wavelet transform. Based on the analysis above, the image fusion methods of anisotropic multi-scale analysis is emphasized. The NSCT (Non-sampled Contourlet Transform) is used as the main tool. The application of NSCT in image fusion is proposed and a novel fusion strategy is designed correspondingly. A new image fusion method based on NSCT and PCNN (Pulse Coupling Neural Network) is developed. This method takes advantage of the properties of anisotropic, multi-directional and shift invariance of NSCT. The frequency-aliasing problem of the previous multi-scale decomposition is eliminated. The general coupling characteristic of PCNN is also adopted. The view effect of the fused image can be improved with the help of its biological background. Experimental results show that the fused images based on the new method are clear and natural, which is suitable for the human observation. Compared with the current wavelet method, the better fusion effect is achieved.Comprehensive evaluation method for different source image fusion quality is studied. Based on the analysis of the existing subjective and objective image evaluation methods, multiple evaluation factors are classified. The factor selection rule and the limitation of single factor are also discussed. Thereafter, current integration methods for multiple factors are researched and compared. Based on the information above, a mathematical model for image fusion evaluation is designed on the foundation of FNN (fuzzy neural network), which combines the structural knowledge expression ability of fuzzy logic and the self-learning ability of neural network. Multiple typical image fusion objective evaluation factors are fuzzificated with the model. The subjective evaluation results are used as the prior knowledge. The weight factors of the evaluation parameters are generated by the network study. The efficiency of the network study is improved through a momentum factor. Experimental results show that the newly-developed evaluation method can conduct a comprehensive and accurate evaluation, which overcomes the one-sidedness from the human subjective judgments and the single objective evaluation factors.Adaptive image fusion scheme for different source images is investigated. Based on the above new comprehensive evaluation method, a close-loop adaptive image fusion model is established. The layers in the model are assorted serially according to the calculation complexity of fusion algorithms, including multiple fusion methods of various features. The scheme can select the optimized fusion scenario according to the user requirement. The shortcoming of the shortage of flexibility of the open-loop system is overcome. Furthermore, the algorithms in the model can be supplemented easily, which has a good expandability. In order to verify this model, weighted average fusion method, wavelet-based fusion method, Contourlet-based fusion method and NSCT-PCNN-based fusion method are selected as the algorithm group of the experiment. Experimental result shows that the new scheme set up the relationship between fusing result and the user requirement. The total fusion process is unmanned and the automation degree is improved.更多还原