SIA OpenIR  > 光电信息技术研究室
Alternative TitleResearch on Target Matching-Recognition Technology of imaging Guidance System
Thesis Advisor史泽林
Keyword自动目标识别 Hausdorff距离 模板匹配 迭代切距离
Call NumberTP391.4/W32/2010
Degree Discipline模式识别与智能系统
Degree Name博士
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract 目标匹配识别技术是数字图像处理和模式识别中的一个重要研究方向,无论在军事上还是在民用方面都有着重要的应用价值。该问题的有效解决,可以使自动目标识别和目标跟踪等相关任务得到有实效的推进。本文在简要分析自动目标识别技术的国内外研究现状基础上,对成像制导目标匹配识别技术中的几个关键技术问题进行了研究,主要包括:(1)如何有效地提取红外图像的边缘特征;(2)如何利用可变形模板匹配技术进行仿射变换条件下的目标识别;(3)如何对图像匹配的相似性进行度量;(4)如何降低匹配过程中的计算量,提高实时性。主要研究内容如下:(1)在红外图像的边缘特征提取研究中,针对红外图像边缘模糊、对比度差的特点,利用侧抑制网络的突出边框、增强反差的功能,构造了基于边缘可能性度量的自适应侧抑制网络,用于红外图像的增强。对图像增强后噪声也得到增强的缺陷,利用小波的多尺度分析功能有效地抑制噪声进而提取边缘。将侧抑制网络和小波变换有机结合,准确、有效地提取低对比度、低信噪比、边缘模糊的红外图像的边缘特征。(2)在变形模板匹配技术研究中,针对仿射变换条件下的几何变换,在对切距离的分析基础上,提出多分辨率迭代切距离的目标识别方法。识别过程中采用迭代的方法消除切距离的一阶近似,使求解的迭代切距离更逼近模板和实时图像的真实距离。同时,针对迭代过程收敛域小、容易陷入局部最优的缺点,将迭代切距离嵌入多分辨率框架,扩大了收敛域进而提高算法性能。所提出的多分辨率迭代切距离具有更高的不变性,提高了识别率。实验结果验证了算法的有效性,为几何变换下的目标识别提供了新的研究思路。(3)在图像匹配的相似性度量研究中,在对基于边缘特征图像匹配算法研究的基础上,提出引入黎曼度量的Hausdorff距离图像匹配识别方法。利用边缘点的灰度和邻域梯度信息构造边缘结构张量,用其对Hausdorff距离进行加权,改进了现有的Hausdorff距离匹配方法,增强匹配的容错性,较好地解决了传统Hausdorff距离匹配方法中因噪声点、伪边缘和光照变化造成的误匹配问题。(4)针对Hausdorff距离图像匹配方法计算量大的问题,提出基于增量群体学习(PBIL)搜索策略的Hausdorff距离图像匹配方法,该算法集成了基于函数优化的遗传搜索和竞争学习两种策略,具有较强的鲁棒性和并行搜索能力,因此能获得较快的收敛速度和理想的计算结果。该方法将进化过程视为学习过程,首先从一个均匀分布的初始概率出发,产生若干个随机的可行解种群,通过对可行解种群中的个体进行分析与评价获得知识,根据学习所获得的知识修正生成概率,进而指导后代的生成,在匹配速度得到保证的基础上,提高匹配的正确率。
Other AbstractResearch on target matching-recognition technology is one of the important subjects on digital image processing and pattern recognition, which plays very high roles both in military field and daily life. The core idea of recognition is to match, and if it is sufficiently solved, which will improve the efficiency of ATR, target tracking and other similar tasks. On the basis of the chief study about the current research of automation target recognition technology based on template matching in the world, the thesis is to do further research on some principal technological questions on automation target recognition technology based on template matching. It includes: (1) How to extract effectively the edge characteristics of infrared image; (2) How to match the deformable template under the conditions of affine transformation; (3) How to measure the similarity between images;(4) How to reduce the calculation of the matching process and improve the performance of real-time matching. The main contents as follows: Research on the edge extraction of infrared image based on lateral inhibition network of wavelet, we propose an adaptive lateral inhibition based on the possibility of edge, and use it in the enhancement of infrared image because of the characteristics of poor contrast and blur. At the same time, the defective effects of noise enhancement for the image enhancement processing will be restrained by the utilization of multi-scale wavelet analysis. Based on the multi-resolution of wavelet, it gives a algorithm to combine the noise insensitiveness ability of wavelet and edge enhancement ability of the lateral inhibition network. Then we can use it extracting the edge of image accurately and effectively. On the study of deformable template matching technology, we get invariant of distance measure and use it for target identification under affine transformation. On the basis of tangent distance analysis, we put forward a multi-resolution iterative distance method. In the course of detection, the first-order approximation of tangent distance will be cut off with the help of iterative method, so that the detecting iterative tangent distance is much closer to that of template and real image. At the same time, the limit convergence domain of the iterative method is easy for local optima, which is convenient to insert into a multi-resolution frame, which will enlarge the convergence domain physically, so as to improve the efficiency of calculation. The comparison among multi-resolution iterative tangent distance, tangent distance and iterative tangent distance in the thesis is of higher consistency for affine transformation and it promotes detection rates. The experiments verify the efficiency of such method and it will provide new sight in the study of target detection under geometric transformation. Research on Hausdorff distance and Riemannian metric, the edge structure tensor for gray edge and adjacent gradient information will be created, so as to weight Hausdorff distance and to improve the effect of Hausdorff distance matching methods, and to enhance the fault tolerance, therefore, the matching errors from the noise, fake edge and change of light in the course of formal Hausdorff distance matching methods will be solved.For the large computation from Hausdorff distance matching methods, Hausdorff distance matching methods based on PBIL is put forward. The method integrates function optimization genetic search and competitive learning strategies, which processes strong robustness and parallel algorithm, so that it can make better convergence rates and ideal calculation results. It develops the evolution as a process of study: first, take the equally distributed initial probability as a start; then feasible solution groups are produced, which will be canalized and evaluated separately ;finally, the modified definition of probability will be created, which will instruct the latter study. On the guardians of high speed, the accuracy will be improved.
Contribution Rank1
Document Type学位论文
Recommended Citation
GB/T 7714
王国刚. 成像制导目标匹配识别技术研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2010.
Files in This Item:
File Name/Size DocType Version Access License
成像制导目标匹配识别技术研究.pdf(4253KB) 开放获取LicenseApplication Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[王国刚]'s Articles
Baidu academic
Similar articles in Baidu academic
[王国刚]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[王国刚]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

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