【作者】 罗晓清；

【导师】 吴小俊；

【作者基本信息】 江南大学 ， 轻工信息技术与工程， 2010， 博士

【摘要】 多传感器图像融合是以图像为研究对象的信息融合,它把对同一目标或场景的用不同传感器获得的不同图像,或用同种传感器以不同成像方式或在不同成像时间获得的不同图像,融合为一幅图像,在这一幅融合图像中能反映多重原始图像的信息,以达到对目标和场景的综合描述,使之更适合视觉感知或像医疗应用、目标检测与分类等计算机处理任务。它是一门综合了传感器、信号处理、图像处理和人工智能等技术的新兴学科分支。近些年来,多传感器图像融合己成为图像理解和计算机视觉领域中一项重要而有用的新技术。多传感器图像融合的处理通常可在以下三个不同层次上进行：像素级、特征级、决策级。本文在分析国内外多传感器图像融合方法的基础上,主要对像素级和特征级融合技术以及多传感器图像融合评价方法等三个方面进行了研究,论文的主要学术贡献和工作如下：(1)在像素级图像融合算法研究方面提出了三种算法。具体地有：提出一种新的基于窗体以及多目标向量求值的量子粒子群优化算法(VEQPSO)的图像融合方法。低频子带采用基于灰度关联分析的VEQPSO算法完成融合过程。高频部分划分窗体,根据不同的窗体类型,运用窗体信息能量和模糊熵指导融合策略；提出了一种改进的形态学小波多聚焦图像融合算法；针对形态学小波融合方法(MMWF)在重构尺度信号时由于发生位置错误而导致灰度值下溢的问题,采用了检测-重融合的方法,该方法保留了MMWF快速、有效、易于实现等优点,同时融合效果也得到了提高；针对统计模型的遥感图像多分辨率融合方法中需要设置相关门限以及约束条件、参数过多,使得算法复杂度增加的问题,提出一种改进的算法。该算法将约束条件变形,然后将其融合在目标函数中,构造新的目标函数,通过最大化拉格朗日函数求偏导数的方法估计参数。该算法可以避免原始算法中的参数设置,并且融合图像的空间分辨率和光谱保持能力均能达到较好的效果,算法鲁棒性增强并且复杂性降低。(2)在多特征图像融合算法研究方面提出了四种算法。论文在研究基于多特征模糊聚类的图像融合方法的基础上,提出了基于卡尔曼滤波的噪声图像的融合方法。该方法结合了滤波和多特征的优点,可提高融合效果,减少图像噪声对融合的干扰；将多通道Gabor滤波与区域方法结合,提出了新的区域相似性度量方法用于图像融合。研究表明,该方法的融合性能对于Gabor滤波器参数(径向中心频率和方向角)的选取不敏感,算法具有一定的稳定性；针对FCM算法容易陷入局部最优的弱点,选用量子粒子群算法与模糊C均值聚类(FCM)相结合的方法(QPSO-FCM)。鉴于QPSO-FCM具有很好的分割效果,将全局优化算法QPSO引入多特征图像融合过程中,这将有利于融合效果的提高；提出一种新的基于二次融合多特征的融合方法,研究表明,这种二次融合方法对多聚焦图像融合有很好的效果。(3)多传感器图像融合评价方面,在对经典的图像融合评价方法研究的基础上,提出了新的图像融合客观评价方法。考虑到人对区域信息更为敏感,因此算法将图像进行区域分割,利用区域特征矩阵表示区域中的空间、纹理和灰度信息等内容,算法更适合于评价。针对是否有参考图像,本文提出了两种用于评价融合图像的新的度量公式。此外,在上述研究的基础上将二维主成分分析(2DPCA)引入上述评价算法,又获得了一种新的算法；本文提出的图像融合评价新方法考虑了图像像素的局部关系以及区域的显著性,更加符合人类的视觉特征。更多还原

【Abstract】 Multisensor image fusion is a kind of information fusion, which refers to the synergistic combination of different sources of sensory image into a single image. The information to be fused may come from different sensors of the same object or scene,or from the same sensor of different imaging manner, or from different images of different time period. The fused image can reflect multiple properties of source images, which makes it more suitable for the purpose of human visual perception and computer processing tasks such as medical applications, detection or classification tasks.Multisensor image fusion is a new branch of research that involves sensors, signal processing, image processing, and artificial intelligence. Recently, image fusion has been an important and useful technique for image analysis and computer vision.Multisensor image fusion can occur at three different levels:pixel level, feature level, and decision level. However, the main concern of this thsis is to present a study on pixel level fusion, feature level fusion and the evaluation method of image fusion respectively. The main contributions are summarized as follows:(1) Three new algorithms are proposed in pixel level fusion. The first new algorithm of pixel level fusion is proposed based on the combination of windows and vector evaluated quantum behaved particle swarm optimization(VEQPSO), in which VEQPSO incorprating with gray relational analysis is utilized for low frequency band, while information energy and fuzzy entropy of windows are adopted for fusion of high frequency band. Since there is position error which leads to the downward overflow of gray value of images when the scale signal is reconstructed for modified morphorlogical wavelet filter (MMWF), a detection-refusion strategy is proposed in the second proposed algorithm. The proposed method preserves the advantages of MMWF including fast speed, effectiveness, and easy implementation. Furthermore, the performance of the proposed fusion method is improved significantly. To reduce the dependence on parameter and strengthen the robustness of the algorithm based on statistical model, a modified statistical fusion model (MSFM) is proposed in which new objective function is presented and the constraints are simplified. Using this new model, the interrelated spatial information is well enhanced, and the spectral information of multi-spectral images is effectively kept. Furthermore, the proposed algorithm can not only avoid determining a threshod of the conventional statistical model, but the robustness is enhanced and the complexity is reduced as well.(2) Four algorithms are proposed in feature level fusion. A noisy image fusion algorithm based on Kalman filter is proposed based on previous image fusion method using multi-feature fuzzy clustering. The proposed algorithm gains an improved performance since it combines the advantages of Kalman filter and multi-feature. A new region similarity is proposed based on multichannel Gabor filtering and region fusion method. The research indicates that the performance of image fusion is insensitive to the selection of different Gabor parameters including the center frequency and orientation. Furthermore, the proposed method for image fusion is stable in fusion performance. To avoid the local minimum problems of fuzzy C-means(FCM), a new clustering algorithm QPSO-FCM is proposed that incorporates the FCM into QPSO algorithm. Since QPSO-FCM can produce a very good segmentation, which leads to an improved performance of fusion. A new double fusion method based on multi-feature is proposed. The research indicates that the proposed method is very effective for the fusion of multi-focus images.(3) After research on the conventional evaluation methods of image fusion, a novel metrics for evaluation of fused images are proposed based on the similarity of corresponding regions in images. The new metrics are computed on a region-by-region basis, which is more suitable for the evaluation because human eyes are more sensitive to regions. The region information is represented by feature matrix of region, which consists of multi-feature vectors including spatial information, texture and gray value, which can adequately reflect the regional content. Two new quality metrics are proposed which consider the reference image is available or not. Furthermore, two dimensional principal component analysis(2DPCA) is introduced into the process of evaluation, which leads to a new version of the proposed evaluation method. Research indicates that the proposed metrics are more consistent with the nature of human perception as it considers the local image variations and the saliency of region.更多还原