载入中....
设为首页 收藏本站 联系我们 网站地图
论文网
您现在的位置: 免费毕业论文网 >> 理学论文 >> 生命环境 >> 正文
搜索: 论文

基于遗传规划物体探测的研究

更新时间 2009-9-11 2:46:31 点击数:

基于遗传规划物体探测的研究
The Research on Genetic Programming Techniques for Object Detection
【中文摘要】 物体目标探测的过程是在图像里自动寻找目标物体。鉴于目前大量的信息都是以电子形式储存,物体目标探测将会迅速成为一个有用的和富有挑战性的机器学习和计算机视觉的任务。在实际运用中,如在卫星遥感图像中探测船只,在医学X光片上找出肿瘤的位置并且分辨出肿瘤的种类。因此,如何解决物体探测问题,是近几年研究的重点之一。本文将利用遗传规划(Genetic programming)来解决不同难度的多类物体的探测问题。遗传规划是一种新型的搜索寻优方法。它仿效生物界中进化和遗传的过程,遵从“优胜劣汰,适者生存”原则,从一组随机生成的初始可行解开始,通过复制、交叉和变异等遗传操作,逐步迭代而逼近问题的最优解。首先,对物体特征值的提取是采用像素统计法来提取,并对特征提取窗口做了不同的区域划分。实验证明,科学的特征提取窗口的划分在对不同的物体探测时能够更有效的提取出物体的特征值,提高物体的探测效率。其次,分析了基于遗传规划的三种不同的物体探测的方法:直接探测法,物体分类探测法,混合法。研究的思路是找到一个既能够保持探测效果,又能够减少训练/进化时间的方法。其中混合法就是通过在训练/进化阶段采用两次训练的方法,第一次...更多训练采用分类的方法,初步训练出能够正确探测出物体的初始种群,缩小在第二次训练/进化过程中的搜索空间。第二次训练是在第一次训练得到的种群的基础上进一步训练/进化,进化出能够正确探测出目标物体的最优遗传程序。通过实验证明,混合法能够在保持探测率的基础上缩短训练/进化的时间。最后,我们通过对两个不同的适应度函数的比较:一个是把反应遗传程序规模大小的参数引入到适应度函数,一个是适应度函数中没有反应遗传程序规模大小的参数。通过实验证明,在适应度函数中加入与程序规模有关的内容,在保证物体探测准确率的前提下,能够达到了缩小程序规模之目的。程序规模的缩小,能减少进化过程中可能的探索空间,能提高物体探测效率。
【英文摘要】 Object detection is the process of automatically finding objects of interest within images. In view of the current capture and a wealth of information stored in electronic form, which will quickly become a useful and challenging machine learning and computer vision tasks. In practical application, such as satellite remote sensing images in the detection of vessels, in the medical X-rays to identify the location of the tumor and identify the type of tumor. In this paper, a new genetic learning algorithm for each of Genetic Programming (Genetic Programming) for object detection and through experiments carried out to try.This article will use the algorithm has nothing to do with the area - Genetic Programming (Genetic programming) to solve the difficulty of the many types of different object detection problem. Genetic Programming is a new type of search optimization methods. It followed in the evolution of biological and genetic process, in compliance with the "survival of the fittest, su...更多rvival of the fittest" principle, from a set of randomly generated initial feasible solution to start, through reproduction, crossover and mutation, such as genetic manipulation, and gradually approaching the problem of iterative optimal solution.First of all, the characteristics of the text object extraction is used to extract the pixel statistics of the feature extraction window to do a different region. Experiments show that feature extraction science division of the window in the detection of different objects can be more effective to extract the characteristic values of objects to enhance the efficiency of detection of objects.Secondly, the analysis of genetic programming based on objects of three different detection methods: direct detection, object classification detection method, hybrid method. Research to find a way of thinking is not only to detect the effect can be maintained, but also can reduce training / evolutionary time. Hybrid method which is adopted in the training / evolution phase of the use of two training methods, training in the use of the first classification method, the initial training / evolution that can detect objects in the correct initial population and narrow in the second training / evolution the search space. The second training is training in the first population to be further training / evolution; evolution can correctly detect the target object of the optimal genetic programming. Experiments proved that the hybrid method can maintain the basis of detection rates to shorten the training / evolutionary time.Finally, we adopted two different fitness function of the comparison: one is to respond to the size of genetic programming parameters into the fitness function, a fitness function is no response to the size of genetic programming parameters. Experiments show that the fitness function by adding the size and procedures relating to the contents to ensure accurate detection of objects under the premise of reducing the process to achieve the purposes of scale. Reduction in the size of the procedure can reduce the possible evolution of space exploration to increase the detection efficiency of objects.

【中文关键词】 遗传规划; 物体探测; 特征提取; 适应度函数
【英文关键词】 Genetic programming; Object detection; Feature Extraction; fitness function
论文目录】
摘要 4-5
Abstract 5-6
1 引言 9-11
    1.1 研究背景 9
    1.2 国内外研究现状 9-11
2 遗传规划的基本原理 11-19
    2.1 遗传规划的基本原理 11-12
    2.2 遗传规划的结构 12
    2.3 初始群体的生成 12-14
    2.4 遗传算子及遗传操作 14-15
    2.5 遗传规划运行参数 15-16
    2.6 遗传规划程序产生方法 16
    2.7 适应性度量 16-18
    2.8 终止准则与结果标定 18-19
3 基于遗传规划物体探测的技术研究 19-32
    3.1 基于遗传规划物体探测的三种方法 19-22
        3.1.1 直接探测法 19-20
        3.1.2 物体分类探测法 20-21
        3.1.3 混合法 21-22
        3.1.4 图像特征提取方法 22
    3.2 函数集和终端集 22-25
        3.2.1 终端集 22-25
        3.2.2 函数集 25
    3.3 物体分类和探测的策略 25-28
        3.3.1 基于遗传规划的两类分类过程 25-26
        3.3.2 基于遗传规划的多类分类技术 26-28
    3.4 适应度的计算 28-32
        3.4.1 物体探测的适应度函数的设计 28-30
        3.4.2 适应度值的计算过程 30
        3.4.3 用于分类的适应度函数的设计 30-32
4 基于遗传规划的物体探测技术实验 32-43
    4.1 图像样本 32-33
    4.2 探测基本绘制图形的实验 33-35
    4.3 对英文字母探测的实验 35-39
        4.3.1 直接探测法的实验结果与分析讨论 35-37
        4.3.2 物体分类探测法的实验结果与分析讨论 37-38
        4.3.3 混合探测法的实验结果与分析讨论 38-39
    4.4 对硬币探测的实验 39-41
        4.4.1 直接探测法的实验结果与分析讨论 39-40
        4.4.2 物体分类探测法的实验结果与分析讨论 40-41
        4.4.3 混合探测法的实验结果与分析讨论 41
    4.5 实验总结 41-43
5 结论与展望 43-44
    5.1 总结 43
    5.2 展望 43-44
参考文献 44-47
在读期间发表的学术论文 47-48
作者简历 48-49
致谢 49

返回栏目页:生命环境论文

设为主页】【收藏论文】【保存论文】【打印论文】【回到顶部】【关闭此页