【作者】 杨强；

【导师】 王华军；

【作者基本信息】 成都理工大学 ， 应用地球物理， 2015， 博士

【摘要】 传统图像融合及超分辨率重建算法是在源图像全部像素信息基础上进行融合与重建,存在时间和空间复杂度较高的问题。压缩感知理论利用测量矩阵对信号进行降维测量,在获取图像的同时对图像数据进行压缩,降低了采样频率,减少了图像数据采样量,极大的降低了数据传输、处理和存储的要求,为当前大数据量的图像融合和超分辨率重建提供了新的研究思路。基于压缩感知理论的图像融合及超分辨率重建有利于改善图像退化引起的空间分辨率下降问题,能够获得更多的图像特征和细节参数,在图像特征提取、分类识别等方面有重要意义。本文研究了基于压缩感知理论的图像融合及超分辨率重建算法,并将实现的算法应用于遥感图像融合及超分辨率重建之中,取得了较好应用效果。论文首先对压缩感知的主要理论成果进行了深入分析研究,提出了基于改进K-SVD算法的图像稀疏表示方法,提出了基于改进Hadamard矩阵的图像测量与重构算法;其次,研究了基于压缩感知理论的图像融合方法,阐述了基于小波变换和傅里叶随机测量矩阵的图像融合算法,提出了基于改进K-SVD和Hadamard测量矩阵的图像融合算法;最后,提出了基于压缩感知和学习字典的单幅图像超分辨率重建算法。论文的主要贡献及创新点:(1)提出基于改进K-SVD的图像稀疏表示算法研究了图像稀疏表示问题,提出了基于改进K-SVD的图像稀疏表示方法。改进算法在K-奇异值分解算法的基础上,根据图像的几何特征,构造了平滑、边缘轮廓和纹理结构等多成分字典,再对多成分字典进行分类逐一更新,实现图像稀疏表示。基于改进K-SVD的图像稀疏表示算法突破了对正交基的限制条件,对图像的不同特征区域采用不同的正交基,用不同区域的正交基组合形成一个框架,利用超完备基来表示图像。和传统的K-SVD算法相比,改进算法构造的字典更加紧湊,获得的稀疏因子较小,对于多特征的图像稀疏表示有更好的效果。(2)提出基于改进Hadamard矩阵的图像测量与重构算法研究了自适应测量矩阵的构造及优化问题,对确定性的测量矩阵进行了分析研究,提出了改进的Hadamard测量矩阵快速构造算法,并对Hadamard测量矩阵进行优化。改进的Hadamard测量矩阵独立随机变元更少,测量矩阵的构造可以用快速循环移位变换实现,同时构造的测量矩阵有更好的稀疏性,避免了测量矩阵中存在四方阵的情况。在重构算法OMP中,采用改进优化后的Hadamard测量矩阵取得了较好的重构效果。重构算法在PSNR值、重构迭代次数、重构算法执行时间等方面取得了较满意的实验效果。(3)提出基于改进K-SVD和Hadamard测量矩阵的图像融合算法深入研究了基于压缩感知理论的图像融合方法,阐述了基于小波变换和傅里叶随机测量矩阵的图像融合算法,提出了基于改进K-SVD和Hadamard测量矩阵的图像融合算法。基于改进K-SVD和Hadamard测量矩阵的图像融合算法首先将待融合图像用改进的K-SVD算法进行稀疏表示,再采用改进的Hadamard观测矩阵进行降维测量,并在压缩感知域内采用加权融合规则方法对测量值进行系数融合,最后用OMP重构算法实现了对融合图像的重构。实验结果表明,融合算法仅作用于压缩感知采样后的少量测量数据,大大减少了参与融合的像素数量,提高了融合算法的时间效率,图像融合效果较好。(4)提出基于压缩感知和学习字典的单幅图像超分辨率重建算法提出了基于压缩感知和学习字典的单幅图像超分辨率重建算法,阐述了图像的自相似性和学习字典的训练方法,在超分辨率重构中引入低通滤波器,选用改进的局部Hadamard矩阵做为测量矩阵,确保图像重建符合压缩感知理论的RIP准则,最后通过迭代算法实现了单幅图像的超分辨率重建。实验表明,超分辨率重建算法获得的重建图像有较好的视觉效果,PSNR值较双线性插值和SRSR算法有所提高。(5)遥感图像融合及超分辨率重建的实现将本文研究的两种融合算法应用于遥感图像融合中,实现了可见光与红外遥感图像的融合,实现了多光谱图像与可见光遥感图像的融合,取得了较好的融合效果。同时,利用本文提出的超分辨率重建算法实现了单幅遥感图像的超分辨重建,取得了较好的重建效果。更多还原

【Abstract】 The traditional algorithm of image fusion and super-resolution reconstruction use the all pixel information of source image to enhance the image characteristics,the algorithm has a higher complexity degree of time and space.The compressed sensing theory makes use of measurement matrix to reduce the dimension of the signal.It can reduce the sampling frequency and reduce the requirement of data transmission,processing and storage.The algorithm of image fusion and super-resolution reconstruction based on compressive sensing theory are beneficial to improve the spatial resolution of image degradation.It is beneficial to make up for the lack of spatial resolution of the original image.The algorithm can obtain more image features and detail parameters,which is important in image feature extraction,classification and recognition.The dissertation analyzed the algorithm of image fusion and super-resolution reconstruction based on compressed sensing,and applied the algorithm to realize the remote sensing image fusion and super-resolution reconstruction.The algorithm has achieved good application effect.Firstly,the dissertation analyzed the compressed sensing theory,and put forward the improved algorithm of image sparse based on the K-SVD,and proposed the improved algorithm of image measurement and reconstruction algorithm based on the improved Hadamard matrix.Secondly,the dissertation analysised the algorithm of image fusion based on compressive sensing theory,and proposed the image fusion algorithm based on the wavelet transform and the Fourier random measurement matrix,and put forward the image fusion algorithm based on the improved K-SVD and Hadamard measurement matrix.Finally,the dissertation realized the super-resolution reconstruction algorithm of the single image based on the learning-dictionary and compressive sensing.The main contributions and innovation points as follows:(1)The dissertation put forward the improved K-SVD algorithm of image sparse representation.The dissertation studied the algorithm of image sparse representation,and put forward the improved algorithm based on the K-SVD.The improved algorithm obtained the multi-component dictionary based on the image geometric features,such as smooth,edge and texture structure,etc.The improved K-SVD algorithm breaks the restriction condition of the orthogonal basis,and uses different orthogonal bases to form a frame,and represented the image by a super-complete basis.The Dictionaries are more compact based on the improved algorithm,and the sparse factor is smaller compared with the traditional K-SVD algorithm,which has better effect on the sparse representation of multiple features.(2)The dissertation put forward the improved algorithm of Hadamard matrix,and realized the image measurement and reconstruction based on the improved Hadamard.The dissertation studied the construction and optimization of adaptive measurement matrix,and put forward the improved Hadamard measurement matrix.The dissertation analyzed the optimization algorithm of matrix.The improved Hadamard matrix has fewer elements of independent random variations,and the structure of the measurement matrix can be achieved by using the fast cyclic shift transform.At the same time,the structure of the measurement matrix has better sparsity,which avoids the existence of the matrix of the measurement matrix.In the reconstruction algorithm OMP,the improved Hadamard measurement matrix is used to obtain a better reconstruction effect.The reconstruction algorithm obtains satisfactory results in the aspects of PSNR value and the execution time of the algorithm.(3)The dissertation put forward the fusion algorithm based on improved K-SVD and Hadamard measurement matrix.The dissertation analyzed the image fusion method based on the theory of compressed sensing.The dissertation analyzed the image fusion algorithm based on wavelet transform and FuLiye random measurement matrix,and put forward the image fusion algorithm based on the improved K-SVD and Hadamard measurement matrix.In the image fusion algorithm based on the improved K-SVD and Hadamard measurement matrix,the algorithm firstly used the improve K-SVD algorithm to represent the sparsed image,and then used the improved Hadamard measurement matrix to reduce the dimension of the measurement,at last,the fusion image is reconstructed by using the weighted fusion rule method in the compressed sensing domain.The experimental results show that the fusion algorithm is effective in reducing the number of pixels in the fusion algorithm and improving the time and space complexity of the fusion algorithm.(4)The dissertation put forward the algorithm of super-resolution reconstruction based on the compressed sensing and learning dictionary for single image.The dissertation analyzed the super-resolution reconstruction algorithm based on the compression sensing and learning dictionary,this dissertation analyzed the image self similarity,and analyzed the training algorithm of learning dictionary.In the algorithm of super resolution reconstruction,the algorithm introduced the low-pass filter,and using the improved local-Hadamard matrix as the measurement matrix,which ensure that the algorithm meets the RIP criteria.Finally,the dissertation realized the single image super resolution reconstruction with the iterative algorithm.Experiments show that the reconstructed images have good visual effect,and the PSNR value is improved compared with the bilinear interpolation and SRSR algorithm.(5)The dissertation realized the fusion and super resolution reconstruction of remote sensing images.The dissertation realized the fusion of visible image and infrared remote sensing images using the proposed algorithm,and realized the fusion of multi-spectral remote sensing image and visible remote sensing image.At the same time,the super resolution reconstruction of remote sensing image is achieved by using the proposed algorithm.The algorithm has achieved good results of the remote sensing image fusion and super resolution reconstruction.更多还原