【作者】 陈浩；

【导师】 王延杰；

【作者基本信息】 中国科学院研究生院（长春光学精密机械与物理研究所） ， 机械电子工程， 2010， 博士

【摘要】 多源图像融合是指综合两个或者多个源图像信息,获得对同一场景更为准确、更为全面和更为可靠描述的图像。目前,由于多尺度变换具有良好的时频域局部特性,因此它被广泛的应用于图像融合领域,当源图像采用多尺度变换进行分解后,所得到的分解系数会处于不同的尺度上,因此可以更有针对性的选择融合准则,实现系数最优化的融合,从而最终改善融合图像的质量。在基于多尺度变换的图像融合算法中,比较成熟和应用较为广泛的当属基于拉普拉斯金字塔的图像融合算法和基于小波变换的图像融合算法。但这两种方法都有其局限性,在基于拉普拉金字塔的图像融合算法中,源图像经拉普拉斯金字塔分解后不仅会产生大量的冗余信息,致使融合过程中数据量增大,而且分解后产生的信息不具备方向性,在基于小波变换的图像融合算法中,虽然小波分解后不会造成数据量增大,且有一定的方向性,从而在一定程度上弥补了拉普拉斯金字塔分解的不足,但小波分解只能对低频信号进行,不能对高频信号进行,同时分解后如何选择一个具有优良特性的融合准则也是一个问题,最重要的是,由于小波基不具备各向异性,因此往往不能实现对图像最为稀疏的表达,这些都会对最终的融合图像质量产生不利影响。因此,针对这些问题,本论文开展了以下几方面工作:(1)针对小波变换只能对低频信号进行分解,不能对高频信号进行分解这一局限性,选用既能对低频信号进行分解,又能对高频信号进行分解的小波包变换来对源图像进行分解和重构,并对融合准则进行了改进以实现红外图像与可见光图像融合。(2)针对融合准则的问题,特别介绍了脉冲耦合神经网络(Pulse Coupled Neural Network, PCNN),并将PCNN进行了有效的改进使之作为融合准则使用;同时为了解决小波变换过程中存在大量的卷积运算,会造成运算复杂,计算量增大,储存空间需求增多等问题,改用提升格式小波变换来对图像进行多尺度变换;最后将提升格式小波变换与改进后的PCNN结合起来应用于医学图像融合。(3)针对小波基不具备各向异性,不能够对图像实现稀疏的表达这一局限性,选用了具有多尺度、多方向性的非下采样Contourlet变换来对源图像进行分解和重构,并将其与形态学处理结合起来应用于多聚焦图像融合。本论文所采用的一系列图像融合算法都是以多尺度变换为基础的,实验结果表明,它们都能取得比较好的融合效果。更多还原

【Abstract】 Multi-source image fusion means that integrating information of two or more source images to get a new image which can represent the scene exactly, entirely and reliably. Recently, due to such good properties as localization, multi-scale transform has been widely used in image fusion. After the source images are decomposed by using multi-scale transform, the coefficients to be got will belong to different scales then, the corresponding fusion rules will be chosen to fuse the coefficients perfectly to improve the quality of the fused image.In the filed of image fusion algorithm based on multi-scale, the image fusion algorithm based on either Laplacian pyramid or wavelet is mature and used widely. But both of them have some limitations, when the Laplacian pyramid is used to fuse images, some redundant information will be got to make the data size increase in the fusion processing, and the Laplacian pyramid can not represent the directional information of the image accurately. Compared with the Laplacian pyramid, although the wavelet transform can not result in increasing the data size and have some directional information, it can only decompose low frequency signal, not high frequency signal. At the same time, when source images are decomposed by the wavelet, how to get the perfect fusion rule is a problem then, the most important is the wavelet base has no property such as anisotropy so that the wavelet can not represent image sparely. All about above will influence the quality of the fused image. Focusing on these problems, the main contributions of this dissertation are summarized as follows:(1) Aiming at the wavelet can only decompose the low frequency signal, not the high frequency signal, the wavelet packet transform which not only can decompose the low frequency signal but also can decompose high frequency signal is used to decompose and construct the source images then, the infrared and visible images are fused by combining the wavelet packet transform and the improved fusion rules.(2) Aiming at the fusion rule, the pulse coupled neural network (PCNN) is introduced especially and improved effectively then, the improved PCNN is used as fusion rule. At the same time, the wavelet transform based on lifting scheme is proposed to simplify the computations and save the memory spaces then, the medical images are fused by combining the wavelet transform based on lifting scheme and the improved PCNN.(3) Aiming the wavelet base has no property such as anisotropy so that the wavelet can not represent image sparely, the nonsubsampled contourlet transform which has property such as multi-scale and multi-direction is chosen to decompose and construct the source images and then, the multi-focus images are fused by combining the nonsubsampled contourlet transform and morphology.A series of image fusion algorithms used in this dissertation are based on multi-scale transform, when these image fusion algorithms are used to fuse the images, the experimental results show that the good fused images can be got.更多还原