【作者】 胡新宇；

【导师】 陈作炳；

【作者基本信息】 武汉理工大学 ， 机械电子工程， 2011， 博士

【摘要】 家蚕微粒子病是由微孢子虫寄生蚕体细胞内引起的可传染性原虫病,它是一种古老的、分布甚广且毁灭性强的蚕病,俗称“蚕癌”,是目前各蚕业生产国规定唯一法定检疫的蚕病。按照优选蚕种、制造无病毒蚕种以及消灭本病的原则,一直沿用母蛾镜检法来防止微粒子病毒通过母体传染,如今已成为防止家蚕微粒子病的主要检测手段。由于人工镜检的方法存在劳动强度大、结果不可复现且容易产生错检或漏检的问题,本文开展基于基于机器视觉的家蚕微粒子病识别方法的研究,将机器视觉技术引入蚕病的检测,主要内容如下：(1)针对微粒子显微图像对比度低、图像不清晰的特点,提出了基于模糊信息的图像增强预处理方法,该方法融合了基于全局的Pal模糊增强与局部模糊对比度增强两种算法思想,实现了改善图像整体对比度和增强目标图像局部细节信息的目的,有利于后续的图像分割处理。(2)针对复杂背景条件下微粒子显微图像的分割技术问题,提出了基于HSI模型的微粒子图像分割技术。根据微粒子图像的颜色特征提取准则,实现了彩色目标对象与非目标杂质图像的直接分离,减少了与微粒子形态相似的其它疑似孢子产生误判的可能性；消除了背景中部分杂质图像对于分割处理效果的不利影响,提高了二维Otsu分割方法对于彩色目标H分量图像分割的适应性。(3)根据微粒子图像的形态特点,研究了微粒子图像的特征提取技术,实现了微粒子图像初始形态特征参数集的选取；针对微粒子图像多特征选择的优化组合问题,提出了基于多特征融合的微粒子图像特征选择技术,该技术通过相关分析方法,实现了初始特征集中特征冗余信息的去除,并采用基于分类器学习的特征优化选择方法,确定了微粒子图像的最佳分类特征集。(4)分析了BP神经网络的学习过程,提出了改进的BP优化算法；针对BP神经网络存在“局部极小值”和收敛速度慢的问题,研究了遗传神经网络的混合训练方法,提出了基于遗传神经网络的家蚕微粒子病识别技术方案,确定了识别系统最优的网络结构,验证了遗传神经网络应用于微粒子图像识别问题的有效性与正确性,并对识别系统方案的软硬件实现技术进行了总体设计。更多还原

【Abstract】 Silkworm pebrine disease is one kind of infectious protozoal disease caused by nosema which parasite inside silkworm cells. It is an ancient, widely distributed and destructive strength of silkworm disease, which is commonly known as "silkworm cancer" and it is the sole statutory quarantine silkworm disease in the sericulture producing countries. In accordance with the principle which is preferred silkworm, silkworm eggs made none of the disease and the elimination of the virus, the female moth microscopic method has been used to prevent maternal transmission of virus particles, and now it has become the particulate detection means of preventing the silkworm diseases.As the manual examination has the problem of high labor intensity and the results can not reproduce and error-prone or undetected, the research to identification technology to the silkworm pebrine based on computer vision have been studied in this doctoral dissertation, and brought computer vision technology into the silkworm disease detection.The main contents are as following:(1) According to the characteristics that the microscopic image of particles has low-contrast and it is not clear, fuzzy image enhancement preprocessing method is proposed based on the fuzzy information. This method combines two algorithms which is the theory of Pal fuzzy enhancement based on global and the fuzzy contrast enhancement based on local and improved global contrast image and enhance the local details of the target image, this is conducive to the subsequent image segmentation.(2) In the light of technical problems for the microscopic image segmentation of particles under complex background and conditions, the technology of image segmentation for particles based on the model of HSI was proposed. According to the extraction criteria of color image feature for particles,the target of color image with a direct non-target separation of impurities was achieved, thus the possibility of false positives for the spores which the form is similar to particles was reduced; as well as the adverse effects for the segmentation which was influenced by the some of the background image of impurities was eliminated, the adaptability of image segmenta-tion which is the H component for color target based on two-dimensional Otsu segmentation method was improved.(3) According to the morphological characteristics of the image particles, the multi-feature extraction techniques of image feature for particles was studied and the original parameters set of features was extracted; in the light of images for particles more than the optimal combination of feature selection problem, a multi-feature fusion of particles in the image feature selection technique was proposed. By the method of correlation analysis, the removal of early start features and redundant information was achieved, the best classification sets of image feature for particles were determined based on the selection method of optimization feature by learning-based classifier.(4) The learning process of BP neural network is analyzed, and the improved BP optimal algorithm was proposed. Because of the problems within the BP neural network which are "local minimum" and "slow convergence", the hybrid training method of genetic neural network was studied, and the recognition technology for silkworm pebrine disease based on genetic neural network was proposed, furthermore the optimal network structure of the recognition system was determined. Based on the genetic neural network, the overall design of the recognition system about hardware and software implementation technology was conducted, and the validity and accuracy of the system was verified.更多还原