基于多尺度变换和稀疏表示的图像融合算法研究

发布时间：2018-07-16 19:41

【摘要】：图像融合就是把来自传感器关于同一场景获取的多幅源图像有效地合成一幅高质量的图像,得到的这幅图像比单一传感器捕获的源图像信息更加丰富、精准和可靠,有利于人类感知、计算机识别、检测等后续工作。本文主要研究基于多尺度变换和稀疏表示的图像融合算法。首先在回顾当前流行的几种多尺度变换融合算法的基础之上,重点研究了基于平移不变剪切波变换的医学图像融合算法;其次,针对传统小波的缺点,分别构造了平移不变双树复剪切波变换和非下采样四元数剪切波变换,研究了其性质;最后,通过分析当前融合规则中的不足,提出了一些改进的融合规则,再结合新的多尺度变换,从而分别提出了基于平移不变双树复剪切波变换和非下采样四元数剪切波变换域的图像融合算法。本文的主要研究内容如下:1.综述了图像融合的国内外发展现状及图像融合中衡量融合图像性能的评价标准,介绍了平移不变剪切波、双树复小波、四元数小波、压缩感知、稀疏表示和脉冲耦合神经网络等概念和相关理论。2.为了提高医学图像融合的质量,提出一种基于平移不变剪切波变换和压缩感知的医学图像融合算法。首先,利用平移不变剪切波变换对源医学图像进行分解;其次,对低频子带,提取子带的空间频率和区域能量,再结合各个子带的相似程度进行融合;对高频子带,考虑到系数数据量比较多,在脉冲耦合神经网络中引入压缩感知进行融合;最后对融合子带进行逆变换得到融合图像。实验结果表明:该算法不仅提高了融合图像的质量,而且还提高了算法的运行效率。3.针对传统小波变换的缺点,通过级联双树复小波变换和剪切波滤波器组,构造了平移不变双树复剪切波变换。提出了一种基于平移不变双树复剪切波变换和稀疏表示的红外线和可见光图像融合算法。首先,对源图像进行形态学变换,再利用平移不变双树复剪切波变换对处理后的图像进行分解;其次,对低频子带,先进行稀疏处理,再结合稀疏系数的拉普拉斯能量和以及S函数进行自适应的融合;对高频子带,给出了自适应双通道脉冲耦合神经网络的融合规则;最后,利用平移不变双树复剪切波变换逆变换重构得到融合图像。实验结果表明:该算法可以有效地提高融合图像的清晰度和纹理特征,优于传统的红外线和可见光图像融合算法4.针对传统小波变换的缺点,通过级联四元数小波变换和剪切波滤波器组,构造了非下采样四元数剪切波变换。提出一种基于非下采样四元数剪切波变换域的多聚焦图像融合算法。首先,利用非下采样四元数剪切波变换对多聚焦图像进行分解;其次,对低频子带,先利用训练的字典进行稀疏处理,再结合稀疏系数的1l范数、区域能量和S函数进行融合;对高频子带,采用子带的空间频率、边缘能量和相似匹配度进行融合处理;最后,对融合系数进行逆变换得到融合图像。实验结果表明:该算法改善了融合图像的边缘和细节信息,优于经典的融合算法。
[Abstract]:Image fusion is an effective synthesis of a high quality image from the multiple source images obtained from the same scene on the same scene. This image is more abundant, accurate and reliable than the source image information captured by a single sensor. It is beneficial to human perception, computer recognition, detection and other follow-up work. The image fusion algorithm of scale transformation and sparse representation. First, on the basis of reviewing several popular multi-scale transform fusion algorithms, the medical image fusion algorithm based on translation invariant shear wave transform is studied. Secondly, the translation and non lower transformation of the translational invariant double tree complex shear wave are constructed for the shortcomings of the traditional wavelet. The properties of the four element number shear wave transform are sampled. Finally, by analyzing the shortcomings of the current fusion rules, some improved fusion rules are proposed, and the new multiscale transformation is combined, and the image fusion algorithm based on the translation invariant double tree complex shear wave transform and the non lower sample four element number shear wave transform domain is proposed respectively. The main contents of this paper are as follows: 1. the development status of image fusion at home and abroad and the evaluation criteria for the performance of image fusion in image fusion are reviewed. The concepts and related theories, such as translation invariant shear wave, double tree complex wavelet, four element wavelet, compression perception, sparse representation and pulse coupling God channel network, are introduced in order to improve the medical map.2. Like the quality of fusion, a medical image fusion algorithm based on translation invariant shear wave transformation and compression perception is proposed. First, the translational invariant shear wave transform is used to decompose the source medical images. Secondly, the frequency and regional energy of the subbands are extracted from the low frequency band, and the similarity degree of each subband is combined. In the frequency subband, considering the large amount of coefficient data, the fusion of compressed sensing is introduced in the pulse coupled neural network. Finally, the fusion subband is converted to the fused image by inverse transformation. The experimental results show that the algorithm not only improves the quality of the fused image, but also improves the operation efficiency of the algorithm.3. for the deficiency of the traditional wavelet transform. A translation invariant double tree complex shear wave transform is constructed by cascade double tree complex wavelet transform and shear wave filter banks. An infrared and visible light image fusion algorithm based on translation invariant double tree complex shear wave transformation and sparse representation is proposed. First, the morphological transformation of the source image is carried out, and then the translation invariant double tree complex shear is used. The shear wave transform is used to decompose the processed images; secondly, for the low frequency subband, the sparse processing is done first, then the Laplasse energy of the sparse coefficient and the S function are combined adaptively. The fusion rules of the adaptive double channel pulse coupling neural network are given for the high-frequency subband. Finally, the translation invariant double tree complex shear is used. The experimental results show that the proposed algorithm can effectively improve the clarity and texture features of the fused image. It is superior to the traditional infrared and visible image fusion algorithm 4. for the shortcomings of the traditional wavelet transform. Through the cascade four element wavelet transform and the shear wave filter bank, the non down recovery is constructed. A multi focus image fusion algorithm based on the non subsampled four element number shear wave transform domain is proposed. First, the multi focus image is decomposed by the non subsampled four element number shear wave transform. Secondly, the low frequency subband is used in the training dictionary for the sparse processing, and the 1L norm of the sparse coefficient is combined with the 1L norm of the sparse coefficient. The domain energy and the S function are fused; the frequency of the subband, the edge energy and the similarity matching degree are fused for the high-frequency subband. Finally, the fusion coefficients are changed to the fusion image by inverse transformation. The experimental results show that the algorithm improves the edge and detail information of the fused image, which is better than the classical fusion algorithm.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】