中国科学院沈阳自动化研究所机构知识库
Advanced  
SIA OpenIR  > 光电信息技术研究室  > 学位论文
题名: 基于机器学习的多类目标识别方法研究
其他题名: Multi-Class Object Recognition by Machine Learning
作者: 吴士林
导师: 朱枫
分类号: TP181
关键词: 目标识别 ; 图像分割 ; 多类 ; 条件随机场模型 ; Joint-boost算法
索取号: TP181/W84/2011
学位专业: 模式识别与智能系统
学位类别: 博士
答辩日期: 2011-05-28
授予单位: 中国科学院沈阳自动化研究所
学位授予地点: 中国科学院沈阳自动化研究所
作者部门: 光电信息技术研究室
中文摘要: 图像理解是计算机视觉领域的研究热点之一,通常情况下,它要求将图像中所有点标注为它所对应的景物类别。这一过程也可以看作是对图像中的所有类别景物目标的识别。当景物比较复杂且类别数目比较多时,一般采用基于机器学习的方法实现多类目标识别任务。其基本思路为将一定量的已经完成理解标注的图像作为训练集,用机器学习的方法建立图像特征与标注之间的内在模型(参数),利用这种模型实现对新图像的标注即目标识别。这类方法具有很好的通用性,它也是目前机器学习领域的重要研究方向之一。主要研究内容集中在学习算法的设计、目标特征的选取以及各目标模式以及关系的建立等等。 本文以自然场景的图像理解为背景,对学习模型的建立,特征的选取和融合,目标间关系的表达等方面开展了研究。主要研究内容与成果如下: 1)纹理是图像中区分不同景物的重要特征之一,利用纹理对目标模式进行建模一直是本方向的研究热点。本文首先对图像中的纹理进行编码,然后利用Joint-boost学习方法对不同景物的纹理依存关系进行建模,并在此基础上设计了多种景物特征项,提出了一种条件随机场模型(Conditional Random Field)对这些景物特征项进行融合。实现结果表明:Joint-boost学习方法可以很好地建立目标的纹理模式,条件随机场模型在对多种特征项的融合方面起到了重要作用。 2)图像中不同景物之间在是否同时出现方面具有一定的关联关系,有效利用这种关联关系可以提高识别性能。本文建立了景物目标间的关联关系模型,为使关联发挥更大的作用,本文给定图像主题的前提条件下,建立了主题与关联关系的模型,通过机器学习的方法确定了模型中的参数。实验结果表明,这种方法可以显著提高系统性能,证明了本文方法的正确有效。 3)轮廓等结构信息是景物的另一类重要特征,对于目标的识别非常重要。为了充分利用目标的结构信息,本文改进了一种具有全局特性的过分割算法,利用图像的底层特征突出目标的边界和区域特性以提高系统的识别正确率;为了降低分割算法带来的误分割问题,我们在基于区域分割的基础上利用目标间的关联关系和其它特征项,并最终建立了结合主题和区域限制、融合纹理环境、纹理以及位置特征项的目标识别系统。实验结果标明:此目标识别系统可以实现自然场景图像中的景物目标识别。  本文面向自然场景的图像理解,以机器学习为基本方法,对其中特征选择和融合、模型建立、目标关系表达等方面展开研究,具有较强的理论意义。另外,对自然目标进行识别具有潜在的应用前景,具有一定的应用价值。
英文摘要: Image Understanding is one of the hotspots in the field of computer vision; typically, it requires all points in an image assigned to an object category. This process can also be seen as the target recognition of all objects in the image. When the landscape is complex and consisting of numbers of object categories, methods based on machine learning are often chosen for multi-class object recognition tasks. The basic idea is to have a certain amount of annotated images as the training set; using machine learning methods to establish a model (and its parameters) between image characteristics and its annotation; to use this model to achieve a new image annotation, i.e., the object recognition. Such method is very versatile and is also one of the most important fields of machine learning. Main research is focused on the design of learning algorithm, feature selection, the establishment of the relationship between different object categories, etc. In this paper, with the understanding of the natural scene images as the background, we investigate the modeling of learning algorithm, feature selection, feature fusion, and the relationship between differernt targets. The main contents and achievement are as follows: 1) Texture is an important feature to distinguish different objets in the scene; the usage of texture in the modeling of target mode has been a research focus in this direction. Firstly, we encode the texture of the image, and then using Joint-boost learning, we model the dependencies for texture features of different objects. On this basis, a variety of features items are designed, and a conditional random field model is established for the fusion of the feature items. Implementation results show that: Joint-boost learning method performs well in the establishment of target texture feature, and the conditional random field model plays an important role in the integration of multiple feature items. 2) Whether different objects may simultaneously appear in the same image has a certain relationship, and the effective use of this relationship can improve the recognition performance. We model this kind of association of different targets. Given the theme of images, this paper model the relationship of a theme and this association between categories and use a machine learning method to determine the model parameters. Experimental results show that this method can significantly improve system performance, which demonstrates the effectiveness of this method. 3) Structure information such as contour is another important feature of the scene, and plays an important role in target recognition. In order to take full advantage of structure information of objects, we improve an over-segmentation algorithm with global characteristics, and highlights the object boundary and regional characteristics by using the underlying features of images; in oder to reduce the problems caused by false segmentation, we further incorporate the relationship between objects and other features and eventually establish a  target recognition system that combines the thematic and regional limits and integrates texture environment, texture and position features of objects. The results indicate the effectiveness of the target recognition system for target recognition in images of natural scenes. For the task of image understanding in natural scenes, in this paper we use machine learning as the basic method, study on feature selection and integration, modeling, expression of relationship between different objects, having strong theoretical significance. In addition, identification of natural target has potential applications, indicating a certain applied value.
语种: 中文
产权排序: 1
内容类型: 学位论文
URI标识: http://ir.sia.cn/handle/173321/9233
Appears in Collections:光电信息技术研究室_学位论文

Files in This Item:
File Name/ File Size Content Type Version Access License
基于机器学习的多类目标识别方法研究.pdf(4712KB)----限制开放 联系获取全文

Recommended Citation:
吴士林.基于机器学习的多类目标识别方法研究.[博士学位论文].中国科学院沈阳自动化研究所.2011
Service
Recommend this item
Sava as my favorate item
Show this item's statistics
Export Endnote File
Google Scholar
Similar articles in Google Scholar
[吴士林]'s Articles
CSDL cross search
Similar articles in CSDL Cross Search
[吴士林]‘s Articles
Related Copyright Policies
Null
Social Bookmarking
Add to CiteULike Add to Connotea Add to Del.icio.us Add to Digg Add to Reddit
所有评论 (0)
暂无评论
 
评注功能仅针对注册用户开放,请您登录
您对该条目有什么异议,请填写以下表单,管理员会尽快联系您。
内 容:
Email:  *
单位:
验证码:   刷新
您在IR的使用过程中有什么好的想法或者建议可以反馈给我们。
标 题:
 *
内 容:
Email:  *
验证码:   刷新

Items in IR are protected by copyright, with all rights reserved, unless otherwise indicated.

 

 

Valid XHTML 1.0!
Copyright © 2007-2016  中国科学院沈阳自动化研究所 - Feedback
Powered by CSpace