基于稀疏表示的医学图像融合算法研究

发布时间：2018-04-12 23:03

本文选题：稀疏表示 + 医学图像融合　；参考：《浙江理工大学》2017年硕士论文

【摘要】：随着医学成像设备的快速发展和普及,医学影像技术已成为临床医学中检查和诊断疾病必不可少的手段,然而单一模态的医学图像所提供信息存在一定局限性,为此,学者们提出了医学图像融合。医学图像融合是将不同模态医学图像之间各自优势进行相互融合,弥补单一模态的医学图像的局限性,从而在单幅医学图像中更加直观地提供人体解剖结构、生理状况及病理特性等信息。由于稀疏表示能够提取少数特征用于表示图像全部信息,因此本文将稀疏表示理论与图像融合技术相结合,并进行深入研究,主要内容如下:1)针对医学图像复杂多样性特点,提出一种基于在线字典学习的自适应医学图像融合算法。首先利用在线字典学习算法训练源图像的过完备字典,提高图像特征提取的自适应能力;然后利用OMP算法对源图像进行稀疏表示得到稀疏编码,降低了融合数据维度;再根据源图像之间稀疏编码的能量差异程度和梯度差异程度自适应调整融合规则,若能量差异程度大于梯度差异程度,则根据能量取大规则融合稀疏编码,反之,根据梯度取大规则融合稀疏编码;最后将融合后的稀疏编码与过完备字典进行重构得到融合图像。实验结果表明:与多尺度几何分析、K奇异值分解等图像融合算法比较,本文算法融合的图像客观评价指标信息熵、边缘评价因子均有所提高,主观上纹理清晰、对比度高,能够较好的保留源图像边缘信息。2)针对ROMP算法在压缩感知重构中需预估稀疏度导致重构精度不稳定的问题,提出一种改进的ROMP算法。由于观测信号能够继承原始信号特征,在选择候选集原子过程中引入自适应弱选择标准,依据观测信号的信息量设定弱选择标准,实现稀疏度自适应调整。将该算法应用于压缩感知框架下的医学图像融合,并提出一种结合观测信号结构相似度的融合规则,当待融合的观测信号之间结构相似度较高时,说明待融合的原始信号之间同样具有相似性,以两者信息量的加权作为融合规则。同理,当待融合的观测信号结构相似度较低时,选择信息量较大的观测信号作为融合后的观测信号。实验结果表明:改进ROMP算法的重构图像质量优于OMP、ROMP、SAMP等算法,其峰值信噪比提高了6%左右。应用于医学图像融合时,得到融合图像具有较好的人类视觉特性,轮廓清晰,保留了源图像中大部分特征信息,可在较短时间内得到优质的融合结果。
[Abstract]:With the rapid development and popularization of medical imaging equipment, medical imaging technology has become an indispensable means for the examination and diagnosis of diseases in clinical medicine. However, the information provided by single mode medical images has some limitations.Scholars have proposed medical image fusion.Medical image fusion is to fuse the advantages of different medical images to make up for the limitation of single mode medical image, so as to provide the anatomical structure of human body more intuitively in a single medical image.Physiological and pathological information.Because sparse representation can extract a few features to represent all the information of an image, this paper combines sparse representation theory with image fusion technology, and makes a thorough study. The main contents are as follows: 1) aiming at the complex diversity of medical image,An adaptive medical image fusion algorithm based on online dictionary learning is proposed.Firstly, the online dictionary learning algorithm is used to train the over-complete dictionary of the source image to improve the adaptive ability of image feature extraction, and then the sparse representation of the source image is obtained by using OMP algorithm, which reduces the dimension of fusion data.Then adaptively adjusts the fusion rules according to the degree of energy difference and gradient difference between source images. If the degree of energy difference is greater than the gradient difference degree, then according to the large rule of energy fusion sparse coding, conversely,Finally, the fused sparse coding is reconstructed from the over-complete dictionary to obtain the fused image.The experimental results show that compared with the image fusion algorithms such as multi-scale geometric analysis and singular value decomposition, the objective evaluation index information entropy and edge evaluation factors of this algorithm are improved, the subjective texture is clear, and the contrast is high.A modified ROMP algorithm is proposed to solve the problem that the ROMP algorithm needs to predict the sparse degree in the compression perception reconstruction, which leads to the instability of the reconstruction accuracy. 2) A modified ROMP algorithm is proposed to solve the problem that the reconstruction accuracy is unstable due to the need to predict the sparsity of the ROMP algorithm.Because the observed signal can inherit the characteristics of the original signal, adaptive weak selection criterion is introduced in the process of selecting candidate set atoms, and the weak selection criterion is set according to the amount of information of the observed signal to realize the adaptive adjustment of the sparsity.The algorithm is applied to the medical image fusion under the frame of compressed perception, and a fusion rule combining the structural similarity of the observed signals is proposed. When the structural similarity of the observed signals to be fused is high,It is shown that the original signals to be fused have the same similarity, and the weighted information between them is taken as the fusion rule.Similarly, when the structural similarity of the observed signals to be fused is low, the observation signals with large amount of information are selected as the observed signals after fusion.The experimental results show that the reconstructed image quality of the improved ROMP algorithm is better than that of the ROMP algorithm, and the peak signal-to-noise ratio (PSNR) is improved by about 6%.When applied to medical image fusion, the fused image has better human visual characteristics, clear contour, and retains most of the feature information in the source image, which can obtain high quality fusion results in a relatively short time.
【学位授予单位】：浙江理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】