【作者】 李勇；

【导师】 王珂；

【作者基本信息】 吉林大学 ， 通信与信息系统， 2010， 博士

【摘要】 本文主要利用具有多尺度分解特性的传统小波变换、Curvelet变换和非下采样Contourlet变换对多源图像像素级融合进行系统的深入的研究。通过大量的仿真实验得出一系列的重要结论,完成了一定的创新工作。本文的主要研究成果如下:(1)针对高斯函数不能根据图像特征自适应选择空间系数σ,影响其融合效果,本文提出一种基于高斯小波函数,利用灰度共生矩阵特征值自适应选择σ,对图像进行融合的算法,并将这种算法应用于医学图像融合中。(2)针对传统小波变换只具有有限的方向,不能最优地表示含“线”或者“面”奇异的高维函数,提出了一种新的基于Curvelet变换的多聚焦图像融合算法。能够获得优于传统图像融合算法的多聚焦图像融合效果。(3)根据红外和可见光图像的成像特点,提出了一种基于NSCT-PCNN变换结合的图像融合算法。结果表明本文算法获得的融合结果具有更优的视觉效果及客观量化指标,能更好地解决下采样过程中融合信息不完全的问题。(4)针对PCA变换融合光谱失真严重的问题,提出基于2DPCA-NSCT变换的多光谱和全色图像的融合算法,该算法在保持PCA变换融合良好的空间分辨率的同时改善了其光谱失真的问题,尤其在抗噪声性能上有优势。更多还原

【Abstract】 Image fusion is an important component of data fusion. Image fusion is a technique which combines information from multiple images of the same scene to obtain more comprehensive, more accurate description of the image. Image fusion can provide more effective information for further image process, such as image segmentation, target detection and identification, feature extraction, etc. At present, Image fusion is widely used in the field of medical science, remote sensing, computer vision, military, etc.According to hierarchical classification, image fusion falls into the following three categories: pixel level fusion, feature level fusion, and decision level fusion. The basis is pixel level, and pixel level fusion is the object of study in this paper. It is widely known that because of the good performance of multi-scale and time-frequency localization characteristic, wavelet Transform is widely used in the field of image fusion. Traditional wavelet Transform can express points singularity of signal effectively, however it cannot express singularity of lines and curves optimally in the image. What’s more, wavelet Transform only possesses finite directional information (horizontal, vertical, diagonal). Aiming at this issue, Curvelet Transform, Contourlet Transform,and nonsubsampled Contourlet Transform theories of geometrical analysis have been proposed in recent years. These multiscale geometrical analyses not only possess the properties of multi-scale and Time-frequency localization, but also possess the multidirectional and anisotropic properties which can provide better sparse express proficiency.In this paper, traditional wavelet transform, Curvelet transform and nonsubsampled Contourlet transform are utilized to deeply investigate the multi-sources image pixel level fusion.The main content of this paper is as follow:(1) Because Gaussian function cannot adaptively select the space factor according to the image feature, the fusion result can be deteriorated. In this paper, Image fusion algorithm based on Gaussian wavelet function makes use of the eigenvalue of Gray-level Co-occurrence Matrix to select the space factor adaptively. And then this algorithm is employed in the field of medical image fusion. The experiment results show that this algorithm possesses the advantage of maintaining the characteristic information of medical image.(2) Because traditional wavelet transform only possesses the finite direction which cannot express High-dimensional Function of line and two dimensions optimistically, a new multistage focusing image fusion algorithm which based on Curvelet transform is presented in this paper. The algorithm adopts Curvelet transform which possess more edge detection capacity to multiscale decompose for image. For decomposition coefficient of each level, It is the fusion strategyies that low frequency coefficients are weighted on average and high frequency coefficients are adopted by adaptive weight method. Finally, according to Consistency Check, the final fusion image can be obtained. The experiment results show that this method can obtain the result of multistage focusing image fusion which is better than traditional image fusion algorithm, and then all the clear fusion images can be obtained.(3) According to the imaging characteristics of infrared and visible light, based on NSCT-PCNN transform, an image fusion algorithm is proposed. This algorithm make uses of nonsubsampled Contourlet transform to multiscale decompose the output image which is after rectification. In the image, two-dimension and multi-dimensions edge texture information is accurately extracted. Then, PCCN model is employed to the fusion of high frequent subband coefficients. For low-pass subband, window variance fusion rule is adopted. In the experiment, we compare this algorithm with Laplacian pyramid transform, Mallat wavelet transform, Contourlet transform. The results show that the visual effects and objective indicator of the images using our algorithm are better than the others. Our algorithm can resolve the issue of incomplete information of fusion.(4) Aiming at the serious issue of spectrum distortion of PCA transform, the advantages of redundancy de-noising, and high resolution of NSCT fusion are employed. Combining PCA with NSCT transform domain fusion, a multispectral and panchromatic image fusion algorithn based on 2DPCA-NSCT transform is proposed. First, each band of multispectral image is decomposed by PCA transform. The main content will be considered as signal information, and the nonessential content is considered as noise. This can improve the robust ability. Then, the panchromatic images and the first main component are decomposed by NSCT. In frequency domain, the approximate component and multi-directional high frequency coefficients are fused by different fusion rules, which the low frequency coefficients are averagely weighted and fusion rule of window average gradient is adopted in high frequency subband. The experiment results show that this algorithm can not only maintain the good space resolution of PCA transform fusion, but also resolve the issue of spectrum distortion. In particular, it can resist noise disturbs to a certain extent.更多还原