【作者】 孟霆；

【导师】 李一兵；

【作者基本信息】 哈尔滨工程大学 ， 通信与信息系统， 2012， 博士

【摘要】 随着数字化信息的到来，特别是数字图像处理技术的不断发展，使得医学图像成为医生诊断、治疗的有力辅助手段成为可能。同时，利用这些图像预处理技术，还可以进一步的改善医学图像的质量，以此来提高诊断的准确性。传统的空域处理方法和傅里叶分析方法不能做到将信号在高维空间中进行稀疏表示，也不能有效的捕捉图像中丰富的纹理信息。针对这一问题，随着数学调和分析理论的发展，Donoho等人提出多尺度几何分析概念。这一理论的提出，为信号在高维空间寻找一种最优表示成为可能，并通过其各向异性的特点捕捉图像中丰富的方向纹理信息，在图像去噪、增强、分割等方面得到了广泛的应用。因此，本文将多尺度几何分析方法引入细胞图像处理中，其主要研究内容包括以下几个方面：针对采集过程中由于光照不足而导致的对比度低、灰度分布不均等问题，提出一种结合非下采样多尺度分析与Retinex模型增强方法，通过对图像照射分量子图像的估计，实现反射分量图像的提取，使图像的对比度得到提升。针对图像在拍摄时，由于景深效应造成的对焦不准问题，本文提出一种基于多尺度分析的多聚焦图像融合算法。该方法通过建立分解系数的能量直方图，对系数分类，从融合源图像中选取合适的融合系数，得到清晰的融合图像。针对图像采集中受到噪声干扰的问题，本文分别讨论了基于Curvelet、Contourlet、NSCT三种多尺度分析工具的去噪算法。为了充分的利用分解系数间的相关性，提出了一种基于NSCT变换的三变量去噪模型，将统计分析理论引入去噪过程，推导去噪收缩函数表达式。由稀疏表示理论可知，信号的稀疏性是影响重构精度的重要因素。因此，从提高信号稀疏性角度出发，提出了一种基于三维协同滤波和系数分类处理的去噪方法，通过平面向立体的转化以及分类处理思想，进一步改善去噪图像的质量。考虑到实际中细胞的信息常以视频信号的形式存储，还提出了一种结合方向Context模型的Surfacelet视频去噪方法，将视频信号看做三维信号进行处理，对其进行立体多尺度几何分析，实现噪声的去除。针对传统分割方法应用于细胞图像存在的速度慢、边缘分割精度低以及误分现象等问题，本文分别提出了基于粒子群的二维Otsu方法和基于Contourlet域HMT模型与改进上下文结构的图像分割算法，实现准确、快速的轮廓分割。最后，本文针对生物细胞结构和形态变化的研究中细胞形态的识别和分割，讨论了基于脉冲耦合神经网络(PCNN)的细胞图像识别计数方法，利用PCNN交叉熵及自动波特征及正向传播和反向传播特性研究，提出了一种基于简化PCNN模型进行分割的细胞图像识别计数方法。更多还原

【Abstract】 With the coming of digital information, especially the continual development of imageprocessing technology, it enables the medical image being the powerful supplementary meansof diagnosis and treating. Simultaneously, using these image preprocessing techniques canfurther improve the quality of medical images, thus improving the veracity of diagnosis.Traditional spatial processing method and Fourier analysis method neither sparsely representsignal in high dimensional, nor effectively catch abundant texture information of images. Inorder to solve this problem, with the development of Mathematics harmonic analysis theory,Donoho et al propose the definition of multi-resolution analysis. The propose of this theorymakes it possible to seek an optimal presentation in high-dimensional space, and catchabundant image texture information by its anisotropy, and this theory has been widely used inimage denoising, enhancement, segmentation, etc. Thereby, in this thesis multi-resolutionanalysis method is introduced to the cell image pre-processing, the main contents include thefollowing aspects:Aiming at the problem that the lack of light to low contrast and the ununiformities ofgrayscale distribution, this paper present a method for image enhancement by combined withthe NSCT and Retinex model. Based on the estimation of image exposure component subimage to realize image extraction and promote the image contrast.Aiming at the problem that image may be partially blurred due to depth of field effect.This paper proposed a multi-focus image fusion method based on multiscale analysis. Itclassify coefficient and select properly fusion coefficients by establish energy histogram ofdecomposition coefficient to get clearly fusion image.Aiming at the interference caused by noise in image transmission, denoising methodsbased on multi-scale analysis tool——Curvelet、Contourlet、NSCT transform are proposed.In order to take advantage of the correlation of coefficients, a trivariate denoising model basedon NSCT transform is put forward. The denoising shrink function is derived by introducingstatistical analysis theory into denoising processing. According to sparse representation theory,the signal sparse is the main factor which affects the reconstruction precision. Therefore, adenoising method based on3-d collaborative filtering and coefficients classification isproposed at the point of improving the signal sparse. The image quality is further improved through the conversion from the plate to stereo and the classification processing. Consideringthat the cell information is usually stored as video signal, a video denoising method based onSurfacelet transform which combines with Context model is put forward. The video signal isregarded as3-d signal to be processed and stereo multi-scale analysis is applied to it for noiseremoval.The traditional segmentation methods applied to cell images may lead the problems,such as slow speed, low accuracy of edge segmentation and segmentation error. This paperproposes two-dimensional Otsu method based on particle swarm and HMT model onContourlet domain and improved context structure image segmentation algorithm respectively,and it achieves accurate and fast segmentation contour.Finally, this paper discusses cell image recognition counting method based on the pulsecoupled neural network aiming at the identification and segmentation of cell morphology inthe research of biological cell structure and morphological change, it uses PCNN crossentropy automatic features and forward propagation and back propagation characteristics,proposing a cells image recognition counting method of dividing based on simplified PCNNmodel.更多还原