【作者】 郑有志；

【导师】 覃征；

【作者基本信息】 清华大学 ， 计算机科学与技术， 2009， 博士

【摘要】 随着多种图像传感器在军事、民用领域的广泛应用,将多幅图像综合成一幅图像的融合技术具有越来越重要的研究意义。由于传感器种类多、图像数据量大、图像特征复杂等因素给图像融合技术带来众多困难和挑战。本文对图像融合问题中多尺度分解算法、分解表示的合成算法和融合图像的质量评估方面进行了深入分析,做了以下几个方面的工作:(1)提出自适应可协调经验模态分解算法(AC-EMD)。AC-EMD算法是完全数据驱动的分解算法,解决了目前图像融合中的多尺度分解算法存在自适应能力差的问题。算法分解过程中将待融合的多幅源图像根据其图像内容自适应并相互协调地分解成一系列具有物理意义的内蕴模函数图像组和一个趋势图像组,AC-EMD多尺度表示具有比金字塔和小波分解更好的图像表示特性。(2)提出了两种塔型的经验模态分解算法。第一种是金字塔结构的分解算法PEMD,PEMD把AC-EMD分解算法和拉普拉斯金字塔分解算法有机结合,有效地降低了图像分解的冗余度。第二种是AC-EMD分解算法与Contourlet分解算法相结合的EMD-CT分解算法,EMD-CT把AC-EMD算法的自适应性、塔型分解的数据结构、Contourlet算法的多方向特性三者结合在一起,算法不仅降低了分解表示的冗余度,而且该分解算法具有多方向的图像表示特性。(3)提出了主成分PCA和一致性检验相结合的融合规则,克服了传感器图像信息量不均等,融合图像局部不一致的情况;提出了基于区域分割的融合规则,利用图像目标区域的特性,更好解决了融合图像信息不一致、不连贯的问题。前者运行效率高,后者融合图像质量更好。(4)提出了两种不需要参考图像的融合图像质量评估指标:基于Renyi信息熵的质量评估指标和基于结构相似性的质量评估指标。前者在计算融合图像与输入图像互信息时利用Renyi信息熵的优点,并同时考虑了互信息交叉重迭问题对评估指标的影响;后者不仅考虑了融合图像和输入图像的结构相似度,而且考虑了输入图像之间的结构相似度对评估指标的影响。更多还原

【Abstract】 With the increasing applications of multi-image sensors in a wide range of areas, such as military domain and civil domain, image fusion has become a more and more important issue of combing multiple source images into a single one. The factors of multi sensor modalities, plenty of image data, and complex features of image, have made vital challenges in image fusion techniques. This paper gives an intensive study on the multiscale decomposition, the synthesis algorithm of multiscale representations, and the quality evaluation metrics of fused images. The work is summarized as follows:Firstly, adaptive coordinate empirical mode decomposition (AC-EMD) is proposed to solve the poor performance adaptivity of multiscale decompositions (MSDs) in image fusion algorithms. AC-EMD is a fully data-driven multiscale decomposition which self-adaptively and coordinately decomposes the source images into a number of“well-behaved”intrinsic mode functions (IMFs) as well as a residual image. The representation of AC-EMD has better physical features of images than that of pyramid and wavelet decompositions.Secondly, two types of pyramidal empirical mode decomposition are proposed. One is pyramid empirical mode decomposition (PEMD). PEMD transform is less redundant, and combines the merits of the Laplacian pyramid and the properties of AC-EMD. The other is a hybrid representation of empirical mode decomposition (EMD) and contourlet transform (CT), named the EMD-CT. EMD-CT transform shares high adaptivity of AC-EMD, data structure of pyramidal transform, while owning multidirection analysis of contourlet transform. The proposed EMD-CT has not only reduced the redundancy of AC-EMD, but also achieved directional representation of the source images. Thirdly, a fusion rule based on the combination of PCA and consistency checking is proposed to overcome the uneven of the source image information of multisensors and incontinuity of fused image in synthesizing multiscale representations. To overcome the inconsistency and inconsequence problem of the fused image completely, we also propose a region-based fusion rule for synthesizing multiscale representations, using the regional properties of target image. The former fusion rule has lower computational complexity, while the latter achieves the better quality of the fused image.At last, two objective image quality metrics without referenced image are proposed: the image quality metric based on the Renyi entropy and the image quality metric using the structural similarity. The former metric measures the total amount of information that fused image contains about source images based on the merit of Renyi entropy, which avoids the overlapping problem of mutual information. The latter metric considers not only the similarity between the source images and the fused image, but also the similarity among the source images.更多还原