【作者】 程军娜；

【导师】 姬光荣；

【作者基本信息】 中国海洋大学 ， 海洋信息探测与处理， 2010， 博士

【摘要】 海洋浮游植物又称浮游海藻,它们既是海洋生态系统中最重要的初级生产者,也是引发赤潮的主要因素。对浮游植物资源的开发利用、对海洋生态系统的监测,都离不开对其种类的分类和鉴定。目前,对海藻的鉴定,最基本的方法是由生物学专家,观察显微镜下的藻类外观形态实现的。这种方法需要研究者拥有很高的专业水平和丰富的经验,且费时又费力。将计算机图像处理与模式识别技术,与这种最基本的方法相结合,建立浮游植物图像自动识别系统,是实现海藻种类准确、快速鉴定的新途径。国内外已有的浮游植物图像识别系统,普遍存在一些局限性：例如,海藻种类繁多,但大多数图像识别系统能够识别的种类偏少等。本文的研究内容是浮游植物图像自动识别系统中的重要组成部分,以海藻细胞显微图像的自动分割以及纹理特征提取作为主要研究目标。图像分割和特征提取的结果是准确分类识别的基础。近年来,形态学图像分析逐步发展成为图像领域一种新兴的研究方法,其基本思想对非线性图像处理分析产生了不可估量的促进意义,尤其适合于实时性要求高的场合。形态学图像分析具有理论严谨完备、节省内存、算法执行效率高、易于用硬件实现等优势。在浮游植物图像识别的国际先进研究工作中,采用形态学图像分析的方法进行图像分割和部分特征的提取,是一个主流的和有效的途径。多尺度分析是一种类似于人类视觉系统的图像分析方法,它的优势在于：图像在不同的尺度下,可表现出与尺度相关的不同特性。因此,在某种单一尺度下无法实现的图像分割或特征检测任务在另外的或者综合的尺度上能实现。而形态学多尺度图像分析方法兼具形态学和多尺度分析的优点。因此,本文系统归纳了形态学多尺度图像分析的理论体系,跟踪其前沿进展,针对海藻细胞显微图像的特点,围绕图像分割和特征提取这两个目标,解决图像分割和特征提取中的难题。本文的主要研究工作及创新之处包括以下几点：1针对含噪、低对比度、弱边界、复杂形状海藻细胞显微图像的分割,利用形态金字塔变换和连通形态算子,解决分割中的主要问题：(1)为了解决图像边缘增强与噪声放大的矛盾,利用附益形态金字塔变换,结合直接灰度映射、梯度边缘检测算子和改进软阈值滤波方法,提出了一种基于Haar形态金字塔的多尺度边缘增强算法。Haar形态金字塔变换的非线性特性,使得图像中的边缘等重要几何信息在各分辨率的残差图像中很好地保留。直接灰度映射增强了整幅图像的灰度对比度,但同时放大了噪声。梯度边缘检测算子和改进软阈值滤波方法,实现了残差图像边缘的跟踪和增强,而对噪声进行了抑制。实验结果表明：针对低对比度和边缘连续性差的海藻细胞显微图像,该算法可以有效地增强图像的边缘,且没有放大噪声。(2)为了解决图像分割过程中噪声抑制和图像细节保留之间的矛盾,在连通形态变换的基础上,提出了一种自适应面积开算法,结合形态重建和属性细化等连通形态算子,在去噪的同时,保留边缘、角毛等精细结构。2提出了一种完备重构的极值保留自适应提升形态小波。在提升小波框架基础上,提出了一种极值保留自适应提升形态小波,而且证明：在不需要分解过程中的提升判据的前提下,它是可完备重构的。极值保留自适应提升形态小波,在对图像的多分辨率分解过程中,只对图像的平坦区域进行中值滤波,保留图像的局部极值点。3以对海藻细胞显微图像的特征描述为目的,将极值保留自适应提升形态小波和基于灰度级共生矩阵(GLCM)的统计法相结合,提出了一种多尺度纹理特征描述方法。为了降低GLCM的计算量,同时有效、可靠地描述海藻细胞纹理特征,本文将极值保留自适应提升形态小波引入基于GLCM的统计法,充分利用极值保留自适应提升形态小波极值保留的多分辩率特性和中值滤波的抗噪性能,使得该多尺度纹理特征描述方法,相对于单纯的基于GLCM的统计法,既具有计算效率上的优势,又降低了对噪声的敏感度。对中国沿海常见浮游植物细胞图像库中的17种圆筛藻种的纹理图像识别实验结果证明：基于极值保留自适应提升形态小波和GLCM的多尺度纹理特征描述方法,比基于原始图像的GLCM的纹理特征描述方法,平均正确识别率要高约7个百分点。4针对甲藻类细胞原始显微图像边缘模糊、对比度低等不利于甲沟提取的困难,构造了一种边缘保留自适应提升形态小波。该小波在根据图像的局部特征构造自适应性更新算子时,将边缘作为感兴趣的图像特征,更新提升根据是否是边缘像素选择保留、锐化滤波或平滑滤波,使得分解后的近似图像中,图像边缘得以保留,而灰度变化缓慢的区域得到了平滑。实验证明：甲藻细胞经边缘保留自适应提升形态小波分解后的近似图像,在一定程度上突出了目标边界,较之原始图像,更有利于甲沟特征的提取。更多还原

【Abstract】 Algae is also called phytoplankton. In order to make use of Algae resources and to monitor marine ecosystem in time, identification and classification of Algae is an essential work. For the sake of identifying Algae appearing frequently in China’s sea areas efficiently and accurately without professional biologist, research on Automatic Identification and Classification of Algae by cell microscopic image (AICA) is carried out, which take the advantages of digital image processing technique and pattern classification method. Image segmentation and feature extraction play an important role in AICA system, which is the main objective of research in this paper.During recent years, Morphological image analysis is considered to be an important and significant technique for lots of image applications, due to its mature and strict mathematical theory, efficient arithmetic and easiness of hardware implementation. It has become a popular tool for real-time image processing. Therefore, it is not surprising that Morphological image analysis is adopted widely to accomplish the task of image segmentation and some feature extraction in most of the advanced research on automatic identification of Algae by microscopic image around the world. But until now, the developed work about automatic identification of Algae is limited in a few species of algae reported in the literature. Much work should be done to advance the research further.Multiscale analysis is an excellent image analysis method similar to human’s vision, with the benefits of representing and processing image in different scale. Thus distinct properties relevant to its scale can be utilized to image segmentation and feature extraction, in particular for those applications where such task is difficult to realize by single scale image analysis. Accordingly, the combination of Multiscale analysis and mathematical morphology provide us with a powerful tool for nonlinear image processing, which is named as Multiscale Morphological image analysis and is applied to solve the problem in process of image segmentation and feature extraction of Algae cell microscopic image.The systemic theory of Multiscale Morphological image analysis is summarized. Based on the systemic theory about Multiscale Morphological image analysis and its up to date development, such as morphological pyramid transformation and morphological wavelet, some innovation is brought out in the field. The main work of this paper is concluded as followed:1 Algae cell microscopic image have the properties of low contrast, weak edge and mixed with lots of noise, which make it very difficult to separate the cell from the background. In addition, according to the variety of Algae, they have multiplicate and abnormal shapes, which add much to the difficulty of image segmentation. In fact, such kind of image segmentation is a challenging work around the world. So it is critical to enhance the edge and reduce noise for the purpose of perfect image segmentation. But contradictions arise in image enhancement and noise magnification, noise reduction and detail reservation. Thus morphological adjunction pyramid transformation and connected morphological transformation are explored to solve the problem:(1) A multiscale edge enhancement approach is proposed based on Haar morphological pyramid transformation, image enhancement by direct grey-level mapping, gradient edge detector and improved soft threshold filtering, aiming at enhance the edge of image while not magnifying noise.The multiscale edge enhancement algorithm mainly take the advantage of non-linearity inherent in morphological adjunction pyramid transformation, which make it possible that the residual images at different resolution decomposed by morphological adjunction pyramid transformation preserves the geometric structure perfectly. Image enhancement by direct grey-level mapping enlarges the dynamic range of the grey-level of the whole image, which enhance the edge and nosie all together. At the moment, gradient edge detector and improved soft threshold filtering are used to trace and enhance the edge only in the residual images. Accordingly, the reconstructed image has the characteristic of enhanced edge without introducing new noise. Experimental results show that this algorithm can enhance the edge of Algae cell while not enlarging noise in the image.(2) An adaptive area open algorithm is built based on connected morphological transformation. The adaptive area open algorithm, banding with morphological reconstruction and attribute thinning, achieve the goal of eliminating noise while preserving the majority of detail information of Algae cell.2 An extremum reservation adaptive morphological wavelet is put forward based on lift wavelet scheme, which performs median filtering to smooth region while leave extremum untouched during decomposition procedure. And it is proved that complete reconstruction can be fulfilled without storage of criterion at each pixel location produced in the procedure of decomposition which saves the memory of computer. Another interest of this adaptive morphological wavelet is that it is non-seperable wavelet transform in spatial domain and has simple architecture for decomposition and reconstruction.3 A multiscale texture feature description method is developed by means of combination of extremum reservation adaptive morphological wavelet with grey-level co-occurrence matrix (GLCM) based statistics approach.This texture feature description method has the virtue of computational efficiency and less sensitivity to noise compared to GLCM based statistics approach itself, so it can describe the texture feature reliably and effectively. Experiments are carried out on identification of 17 species of coscinodiscaceae collected in image database of phytoplankton appearing frequently in China’s sea by texture feature. Experimental results indicate that this multiscale texture feature description method precede about 7 percentage in average correct recognition rate compared with the GLCM method.4 An edge reservation adaptive morphological wavelet is established in pursuit of feature extraction. Adaptivity is successfully introduced into the update lifting by taking into account local characteristic such as edge, which enable the edge remain untouched or sharpened at different resolution image decomposed by this adaptive morphological wavelet. At the same time, smoothing effect is gained by median filtering at other region.The edge reservation adaptive morphological wavelet is used to deal with the tough task for extraction of faint girdle or sulcus of dinoflagellates. Experiments demonstrate that contrast of girdle or sulcus of dinoflagellates to the whole image are improved in multiscale decomposition image by this adaptive morphological wavelet, compared to its original grey-level image. Thus it is more favorable for the extraction of girdle or sulcus.更多还原