【作者】 谢振平；

【导师】 王士同；

【作者基本信息】 江南大学 ， 轻工信息技术与工程， 2008， 博士

【摘要】 农产品的无损检测是农业现代化的一个重要组成部分,随着计算机视觉技术的发展,许多新的无损检测技术应运而生,并取得了一些初步成果。在基于机器视觉的农产品无损检测技术中,首要任务是精准地获取仅含有农产品对象本身的数字图像,即从原始图像中分离出感兴趣目标区域的图像分割,其处理性能将直接影响到机器视觉模型的训练效果和最终的定量分析结果。然而图像分割是计算机视觉处理的基础难题之一,尤其对于一般的自然图像,目前还面临着许多困难,需要继续深入研究。本课题中,以农产品图像为重点研究对象,对图像分割问题进行了深入探讨,其中主要采用了两类具有代表性的农产品图像:水蜜桃阴影图像和自然环境下的水果图像。聚类和曲线进化(水平集)方法是目前图像分割中使用较多的两类处理工具,本文中首先在第二与第三章中重点讨论了基于聚类的图像分割方法,并在现有研究基础上提出了一种新的分裂式层次聚类算法(SHPDHC)和一种增强的可能性聚类算法(EPCM),两者均以本文提出的软边界球分策略为基础。上述策略的引入一方面提高了聚类算法的鲁棒性,另一方面增加了算法标识例外点的能力,这两个优势使得新算法在水蜜桃阴影图像分割上取得了较现有聚类方法更好的效果,对于大部分此类图像均能精准地提取出感兴趣的目标区域。比较上述两种新聚类方法,前者的优势在于能自适应地确定分割区域数目,后者则在聚类数目已知时具有更好的强壮性。第四章中,为克服聚类方法在图像分割处理上的固有缺陷,在前面研究模糊聚类的基础上,结合考虑了水平集图像分割模型,提出了融合模糊C均值的Mumford-Shah图像分割模型FCMMS(本质上也是一种水平集图像分割方法);进一步地第五章中,提出了一种融合高斯混合模型的水平集图像分割模型(GMMLS)。前者通过扩展原有一类水平集图像分割模型的目标泛函至模糊形式而得到;后者中,其模型表达及求解过程在假设图像特征数据分布可由高斯混合模型逼近的前提下,以贝叶斯原理和水平集技术为工具推导求得。上述两种新水平集图像分割模型是对现有相关模型的有力补充,就处理性能而言,它们等价于同时继承了聚类与水平集方法的长处,而弥补了各自方法原有的不足。由于秉承了上述优点,两种新水平集图像分割模型不仅在水蜜桃阴影图像上取得了很好的效果,对于一般的自然水果图像(甚至一般自然图像)同样也取得了较令人满意的效果。就这两种新方法本身而言,GMMLS的最大特点是数学理论基础较为完备,在假设成立,且逼近原型数设定合理时,所得结果在理论上有较好的保障,因而在许多情况下,GMMLS获得的分割结果要优于FCMMS。然而后者具有另一个优势,文中理论分析已指出,FCMMS的模型框架同样适用于其它一些模糊聚类方法,又由文献资料可知,许多学者已就图像分割问题提出了大量具有针对性的模糊聚类算法,通过把这些算法融入到水平集图像分割模型中,可方便地提出相应的变形方法,显然,这些图像分割方法可以在实际应用中发挥更大的作用。第六章中,作为第四与第五章的补充研究,定性地讨论了两类新水平集方法的一些扩展技术,这些扩展技术在许多实际应用中同样起着重要的作用。更多还原

【Abstract】 Nondestructive detection of agricultural product is a basic technique for quality assurance, so it is an important component of agricultural modernization. Along with the development of computer vision techniques, now many corresponding nondestructive detection methods are introduced and some good effect are obtained. In these methods, a primary task is to gain the exact object images, i.e., image segmentation, with which the quantitative analyses on product quality might be done. However, image segmentation is one of the difficult problems in computer vision; to improve the power of corresponding nondestructive detection, the image segmentation problems on agricultural products deserve to further exploration. In this paper, two types of agricultural products are mainly examined including peach shade images and natural fruit images.Clustering and curve evolution (level set) techniques are two prominent image segmentation tools, which are also used to image segmentations of agricultural products in this paper. In chapter 2 and chapter 3, we focus on clustering based image segmentation methods, where two new clustering algorithms are presented including: a new divisive hierarchical clustering algorithm SHPDHC and an enhanced possibilistic clustering algorithm (EPCM). These two new image segmentation methods are all based on the soft hyperspheric partition strategies proposed in the chapter 2 of this paper. Compared with several popular clustering algorithms, SHPDHC and EPCM have the better (or equal) clustering robustness and the ability to label the outliers of the dataset. The peach shade images are mainly examined in these two chapters, and the experimental results demonstrate that the better segmentation results can be obtained by above two new algorithms than several classical clustering based image segmentation methods for the most of such type of images. However, for the complex natural fruit images, the satisfied results can not be obtained by above two algorithms. These cases are further researched in the next two chapters.In chapter 4, a new image segmentation model integrating the fuzzy c-means clustering into Mumford-Shah model (FCMMS) is presented. Next, a coupled level set method integrating with Gaussian mixture model (GMMLS) for image segmentation is proposed in chapter 5. Above two new models can be viewed as a combination of clustering and level set image segmentation methods from the viewpoint of model object functional. They can effectively inherit the merits of two original methods, while the respective shortcomings are weakened. The application performance of these two new methods on peach shade images and natural fruit images are very satisfied with our expectations. Furthermore, in many cases, GMMLS can get the better segmentation performance than FCMMS due to its theoretical maturity. Nevertheless, according to the same framework of FCMMS, many extended versions integrating with current fuzzy clustering algorithms for image segmentation might be put forward, which will improve its practical application worthiness.In chapter 6, some advanced techniques of two new image segmentation methods are discussed. In many cases, these advanced techniques might pay the crucial roles and improve the segmentation performance in great part.更多还原