基于BP网络特征级信息融合及在目标识别中的应用研究
|
摘要
在社会信息化的发展过程中，遥感应用技术占据着越来越重要的地位。而遥感影像中的特定地物目标的识别和提取又是遥感应用技术的热点。利用多光谱遥感影像进行地物目标识别和提取已经有20多年的研究历史，取得了一定的研究成果，但是已有的地物目标识别提取方法所用到的地物目标的特征信息还比较单一，因此地物目标的识别率和提取效果还不够理想。所以本文在特征级信息融合的理论基础上，将地物的光谱特征、几何形状特征和纹理特征进行有机融合，然后利用这些融合信息进行特定地物目标识别和提取，提高了地物目标识别率和增强了地物目标的提取效果。
信息融合技术是多源信息综合处理的一项新技术，在遥感图像处理领域中，象素级融合技术已经相对成熟，而涉及到遥感图像特定地物目标识别和提取的研究时，人们越来越清楚地认识到特征级融合的重要性。在以特征级信息融合技术为基础的目标识别技术中，地物目标的光谱特征、几何形状特征和纹理特征都是不相关的特征向量，把这些不相关特征向量数据非线性关联成一个融合矢量比较困难。而人工神经网络具有对数据类型和分布函数没有限制、对数据的要求更加灵活、容忍度更高等优点。基于人工神经网络的这些优点，本文在特征级信息融合的理论基础上，提出一种基于BP网络信息融合的目标识别提取方法，然后以中等分辨率遥感影像中的机场目标识别为例，验证了本方法的有效性和可行性。
本文的主要研究内容有：
首先搜集、整理和总结了近年来国内外在信息融合领域的最新研究成果进展，并对特征级信息融合的优缺点和适用范围进行了详细总结：其次介绍了人工神经网络技术的理论基础，详细讨论了BP网络结构与参数的设计；接着概述了目标特征选择与提取的方法，并且针对本文选用的三种主要特征——光谱特征、几何形状特征和纹理特征给出了不同的选择与提取的方法，并给出了部分目标特征选择和提取的试验结果；然后重点研究了基于BP网络的特征级信息融合技术，并提出了以此技术为基础的地物目标识别系统，该系统充分结合了信息融合的多源性和神经网络强大的非线性处理能力等优点，并以中等分辨率遥感影像中的机场识别为例，来验证该系统的有效性和可行性；最后本文回顾了作者在理论和应用方面所做的工作，提出了今后进一步研究和改进的方向。
Automatic target recognition and useful information extraction is a key application of remote sensing data. Though more than two decades has been past since the utilization of multispectral remote sensing images for object recognition and extraction, the accuracy as well as efficiency of target recognition is still far from satisfaction, partially due to the fact that relatively simple target features are used in the existing approaches. This paper expounds a methodology which employs spectral reflectance, geometric property and texture feature of a given target synergistically under the framework of feature-level information fusion. The results show the accuracy of recognition is increased in general and the efficiency improved in particular.
Information fusion technology is a newly arisen technology of comprehensive processing of multiple source information. In the field of image processing of remote sensing, pixel-level fusion technology is very ripe. When studing object recognition and extraction from remote sensing images, we realize the importance of feature-level information fusion more and more clearly. For spectrum feature, geometry shape feature and texture feature are irrelevant characteristic vector each other, it is difficult to integrate these vectors into a criteria vector. Neural network doesn't have the limitation of data type and distribution function. More over, it has more flexible requisitions for data and more high tolerant degree. According to these merits, this paper puts forward the feature-level information fusion system based on BP Network, besides studies its application in target recognition, and regards airport from the remote sensing images as the goal of target recognition.
The detailed research work can be sumed up as the following:
At first, the thesis collects and summarizes the latest research results and progress of information fusion technology, especially feature-level fus on technology and its advantages and disadvantages. Secondly, the theory of the artificial neural network (ANN) technology is introduced. This thesis regards BP network as the main research object, and discussed the designs of BP network structure and parameter in detail. In succession, a list of classical feature chosen and extraction methods is explained, and the experiments of those methods that facing to spectrum feature, geometry shape feature and geometry texture feature are given. Then feature-level information fusion technology based on BP network are discussed in detail, and brings forth target recognition system based on the technology, which integrated the multi-source characteristic of information fusion and strong non-linear processing ability of neural network. That system has strong feasibility and efficiency proved by airport recognition and extraction from middle resolution remote sensing images. In the end, this thesis reviews the work that had been done by the author in theory and application, pointes out the research direction and improvement direction in the future.
信息融合技术是多源信息综合处理的一项新技术，在遥感图像处理领域中，象素级融合技术已经相对成熟，而涉及到遥感图像特定地物目标识别和提取的研究时，人们越来越清楚地认识到特征级融合的重要性。在以特征级信息融合技术为基础的目标识别技术中，地物目标的光谱特征、几何形状特征和纹理特征都是不相关的特征向量，把这些不相关特征向量数据非线性关联成一个融合矢量比较困难。而人工神经网络具有对数据类型和分布函数没有限制、对数据的要求更加灵活、容忍度更高等优点。基于人工神经网络的这些优点，本文在特征级信息融合的理论基础上，提出一种基于BP网络信息融合的目标识别提取方法，然后以中等分辨率遥感影像中的机场目标识别为例，验证了本方法的有效性和可行性。
本文的主要研究内容有：
首先搜集、整理和总结了近年来国内外在信息融合领域的最新研究成果进展，并对特征级信息融合的优缺点和适用范围进行了详细总结：其次介绍了人工神经网络技术的理论基础，详细讨论了BP网络结构与参数的设计；接着概述了目标特征选择与提取的方法，并且针对本文选用的三种主要特征——光谱特征、几何形状特征和纹理特征给出了不同的选择与提取的方法，并给出了部分目标特征选择和提取的试验结果；然后重点研究了基于BP网络的特征级信息融合技术，并提出了以此技术为基础的地物目标识别系统，该系统充分结合了信息融合的多源性和神经网络强大的非线性处理能力等优点，并以中等分辨率遥感影像中的机场识别为例，来验证该系统的有效性和可行性；最后本文回顾了作者在理论和应用方面所做的工作，提出了今后进一步研究和改进的方向。
Automatic target recognition and useful information extraction is a key application of remote sensing data. Though more than two decades has been past since the utilization of multispectral remote sensing images for object recognition and extraction, the accuracy as well as efficiency of target recognition is still far from satisfaction, partially due to the fact that relatively simple target features are used in the existing approaches. This paper expounds a methodology which employs spectral reflectance, geometric property and texture feature of a given target synergistically under the framework of feature-level information fusion. The results show the accuracy of recognition is increased in general and the efficiency improved in particular.
Information fusion technology is a newly arisen technology of comprehensive processing of multiple source information. In the field of image processing of remote sensing, pixel-level fusion technology is very ripe. When studing object recognition and extraction from remote sensing images, we realize the importance of feature-level information fusion more and more clearly. For spectrum feature, geometry shape feature and texture feature are irrelevant characteristic vector each other, it is difficult to integrate these vectors into a criteria vector. Neural network doesn't have the limitation of data type and distribution function. More over, it has more flexible requisitions for data and more high tolerant degree. According to these merits, this paper puts forward the feature-level information fusion system based on BP Network, besides studies its application in target recognition, and regards airport from the remote sensing images as the goal of target recognition.
The detailed research work can be sumed up as the following:
At first, the thesis collects and summarizes the latest research results and progress of information fusion technology, especially feature-level fus on technology and its advantages and disadvantages. Secondly, the theory of the artificial neural network (ANN) technology is introduced. This thesis regards BP network as the main research object, and discussed the designs of BP network structure and parameter in detail. In succession, a list of classical feature chosen and extraction methods is explained, and the experiments of those methods that facing to spectrum feature, geometry shape feature and geometry texture feature are given. Then feature-level information fusion technology based on BP network are discussed in detail, and brings forth target recognition system based on the technology, which integrated the multi-source characteristic of information fusion and strong non-linear processing ability of neural network. That system has strong feasibility and efficiency proved by airport recognition and extraction from middle resolution remote sensing images. In the end, this thesis reviews the work that had been done by the author in theory and application, pointes out the research direction and improvement direction in the future.
引文
[1] Mahmood R.Azimi-Szdjadi, Mukhtiar A.Shaikh, Bin Tian, Kenneth E.Eis, Donald Reinke. Neural Network-Based Cloud and Spectral Features. IEEE. 1996:1105～1107
[2] Ljiljana V.Stoiljkovic, Nenad D.Stojanovic, Dejan R.Milenovic. An Approach in texture segmentation using Multi-valued Neural Network. IEEE 1998, 62～67
[3] Robert J.Pawlak, A new neural network architecture for the fusion of the fusion of independent or dependent sensor decision. SPIE Proc.Of Sensor Fusion 1994: 22—32
[4] Merigeault. S, Batarirer.M, Patillon. J.N. Data fusion based on neural network for the mobile subscriber location. IEEE. 2000(9): 536～541
[5] 康耀红．数据融合理论与应用．西安：西安电子科技大学出版社，1997．
[6] 刘宁宁，田捷，胡志刚，诸葛婴 基于神经网络的纹理和灰度信息融合方法．软件学报Vol．10，NO6，1999，575～579
[7] 彭家欢，杨建刚，基于神经网络的多传感器数据融合的分散型网络结构系统．计算机工程．Vol．27，NO．4，2001．29～31
[8] 刘印华，刘韧．神经网络在字符识别中的应用．天津理工学院学报．Vol．12，NO．4，1996，1～4
[9] 蔡自兴，成浩．一种基于骨架特征和神经网络的手写体字符识别技术．计算技术与自动化Vol．20，NO．3，2001，59～65
[10] 胡健，汪庆宝，涂承宇．多层前向神经网络在手写体数字识别应用中的研究[J]．北京工业大学学报，1996，22 (4)
[11] 曹迪铭，宣国荣．神经网络在车辆牌照字符识别中的应用．微型电脑应用．Vol．16，No．6，2000，5～8
[12] 吴连喜，严泰来，张纬．基于多层感知器神经网络对遥感融合图像和TM影像进行上地覆盖分类的研究Vol．32，No．So，2001，33～36
[13] H. Bischof, W. schneider, A.J. Pinz. Multispectral Classification of Landsat images Using Neural Networks. IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENING, VOL. 30, NO.3, May 1992
[14] 王强，胡维平，胡建平，覃少华．基于儿何不变性和B P网络的二维目标识别算法．北京航空航天大学学报．Vol．27，NO．4，2001 (8)：413～416
[15] 叶斌，彭嘉雄．基于结构特征的军用机场识别与理解．华中科技大学学报．Vol．29 No．3，．2001 (3)：39～42
[2] Ljiljana V.Stoiljkovic, Nenad D.Stojanovic, Dejan R.Milenovic. An Approach in texture segmentation using Multi-valued Neural Network. IEEE 1998, 62～67
[3] Robert J.Pawlak, A new neural network architecture for the fusion of the fusion of independent or dependent sensor decision. SPIE Proc.Of Sensor Fusion 1994: 22—32
[4] Merigeault. S, Batarirer.M, Patillon. J.N. Data fusion based on neural network for the mobile subscriber location. IEEE. 2000(9): 536～541
[5] 康耀红．数据融合理论与应用．西安：西安电子科技大学出版社，1997．
[6] 刘宁宁，田捷，胡志刚，诸葛婴 基于神经网络的纹理和灰度信息融合方法．软件学报Vol．10，NO6，1999，575～579
[7] 彭家欢，杨建刚，基于神经网络的多传感器数据融合的分散型网络结构系统．计算机工程．Vol．27，NO．4，2001．29～31
[8] 刘印华，刘韧．神经网络在字符识别中的应用．天津理工学院学报．Vol．12，NO．4，1996，1～4
[9] 蔡自兴，成浩．一种基于骨架特征和神经网络的手写体字符识别技术．计算技术与自动化Vol．20，NO．3，2001，59～65
[10] 胡健，汪庆宝，涂承宇．多层前向神经网络在手写体数字识别应用中的研究[J]．北京工业大学学报，1996，22 (4)
[11] 曹迪铭，宣国荣．神经网络在车辆牌照字符识别中的应用．微型电脑应用．Vol．16，No．6，2000，5～8
[12] 吴连喜，严泰来，张纬．基于多层感知器神经网络对遥感融合图像和TM影像进行上地覆盖分类的研究Vol．32，No．So，2001，33～36
[13] H. Bischof, W. schneider, A.J. Pinz. Multispectral Classification of Landsat images Using Neural Networks. IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENING, VOL. 30, NO.3, May 1992
[14] 王强，胡维平，胡建平，覃少华．基于儿何不变性和B P网络的二维目标识别算法．北京航空航天大学学报．Vol．27，NO．4，2001 (8)：413～416
[15] 叶斌，彭嘉雄．基于结构特征的军用机场识别与理解．华中科技大学学报．Vol．29 No．3，．2001 (3)：39～42