【作者】 牛轶峰；

【导师】 沈林成；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2007， 博士

【摘要】 随着现代防御技术的发展,攻击型无人飞行器(Attack Unmanned AerialVehicle,AUAV),需要不断提高其自主控制能力,而基于图像融合的态势感知是实现自主控制的前提。由于战场环境日益复杂,如何设计有效的图像融合方法,提高AUAV对目标的识别能力和对威胁的规避能力,增强其智能程度,是一个具有重要理论价值与军事意义的课题。图像融合在多维空间综合处理图像信息,形成具有决策意义的融合结果。目前的研究方法多是基于先融合再评价的模式,因而无法使图像融合的结果达到最优。为了提高AUAV多传感器图像融合的性能,论文采用多目标优化理论,对AUAV图像融合的关键环节(包括预处理、融合算法、目标定位等问题)展开研究。论文的研究工作和主要创新点如下:1)研究了基于图像融合的AUAV自主控制系统。介绍了自主控制的概念和层次划分,分析了基于图像融合的AUAV自主控制系统的组成、功能和结构,并设计了相应的系统流程,并将其分为态势感知、动态重规划和任务执行三个阶段,其中态势感知包括图像预处理、图像配准、图像融合、目标识别与定位等环节;动态重规划包括控制策略选择和实时航迹规划两步。2)提出了基于离散小波变换(Discrete Wavelet Transform,DWT)和快速离散曲波变换(Fast Discrete Curvelet Transform,FDCT)的AUAV图像融合预处理方法。针对目前图像去噪方法多不能实现最佳的去噪效果等问题,研究了基于多目标优化算法的图像去噪方法,给出了图像去噪的评价指标,提出了基于DWT和基于FDCT的图像去噪优化算法。实验结果表明,两种方法都可以实现Pareto最优的图像去噪效果,且两者各有优势,DWT法速度较快,而FDCT法去噪效果较好。在一定条件下,如机载传感器成像条件保持不变,去噪阈值可以设为定值,这样得到图像可实时进行处理并输出结果,满足实际应用的需要。3)提出了基于多目标优化理论的AUAV图像融合方法。首先设计了图像融合的有效评价准则,即分别从单幅图像质量、与源图像关系和与参考图像关系三个角度出发,选择并设计了一组有效的评价准则,设计了新的交互信息量指标,避免了信息的重载,并根据指标相关度和指标选取原则对图像融合的指标进行评价和选取;然后以多个评价指标作为优化目标,进行融合参数的优化搜索,并对图像融合的空域模型和小波域模型进行了统一,提出了两种融合规则,简化了图像融合模型的设计;最后进行了多聚焦图像融合、盲图像融合、多分辨和彩色图像融合实验,并根据不同的融合图像进行相应的算法设计和评价指标设计,实验结果表明所提方法不但有效,而且能够得到最优的评价指标。4)提出了最优的图像融合小波分解层数计算方法,并研究了图像融合小波基函数的选取问题。为了实现最优的图像融合,小波分解层数和小波基函数的选择都是很重要的。基于多目标优化理论,提出了用于图像融合的小波分解层数优化方法。实验结果表明最优的分解层数并不是一个固定值,它随源图像的特性发生改变。一般地,最优分解层数随着源图像的尺寸而增加。但无论源图像有何种特性,用多目标优化的方法都可以获得最优的分解层数,而且该方法可以获得Pareto最优的融合结果,使各评价指标最大化。对最优的小波域图像融合而言,小波基函数的选取也是非常重要的。根据图像融合的特点给出了小波基函数的选取原则,为使结果更有意义,计算较为简便,并避免图像重构失真,采用Haar小波函数作为基函数。5)提出了基于变形模板的融合图像目标识别与跟踪方法。在融合图像基础上,研究了目标的提取与定位问题。首先分析了变形模板用于目标识别与跟踪问题的优势,由于变形模板具有形式化的描述参数,因此可以很好的定义各种形状的目标。然后设计了典型目标的解析参数式变形模板,定义了变形模板的能量函数,将其化为单目标优化问题进行求解。提出了基于变形模板的目标识别与定位方法,通过对能量函数的优化达到目标匹配目的,给出了不同目标定位方法及其使用条件。最后研究了基于变形模板的目标跟踪算法,实验表明变形模板方法能够实现快速的目标搜索、识别、定位和跟踪。6)提出了基于PSO的多目标优化算法MOCPSO。借鉴不同的多目标优化算法思想,提出了适合不同优化目标和优化问题的MOCPSO算法(Multi-objectiveConstriction Particle Swarm Optimization),放弃用自适应网格进行种群Pareto最优解的控制,设计了一种新的分层拥挤算子进行种群的多样性保持,并引入自适应变异操作,从而提高算法寻优能力,避免早期收敛,使Pareto最优解尽可能均匀分布于Pareto前端,最后用均匀设计进行算法控制参数的优化设计。分析了多目标优化算法的评价准则,并给出了典型的测试函数,最后通过实验考察了算法的性能,结果表明算法对不同的多目标优化问题都具有较强的适应性。更多还原

【Abstract】 With the development of modern defense techniques, the abilities of autonomous control of attack unmanned aerial vehicles (AUAV), such as cruise missile, attack unmanned plane, urgently need to be improved, which can help to enhance the intelligence of AUAV. In order to realize the autonomous control of AUAV, the situation awareness based on image fusion is the first step. Because of the high complexity of battle environment, how to design an effective method of image fusion has become a problem with both great significant theoretical value and great practical value, which can improve the abilities of target recognition and threat elusion, and enhance the intelligent degree of AUAV.Image fusion can deal with the images and information in multiple dimensions, and form a result for decision making. Aiming to overcome the limitations of traditional image fusion that cannot realize the optimal fusion because of the mode of the fusion before the evaluation, the multi-objective optimization image fusion methods where the evaluation before the fusion are proposed and designed in this thesis. In this architecture, image fusion is divided into preprocessing, fusion and target location to realize the optimal image fusion. The main work and the creative contribution of this thesis are as follows:1) The system structure of autonomous control of AUAV based on image fusion is built. First the concept of autonomous control is defined and the autonomy level is classified; then the structure, components, functions of autonomous control system of AUAV are analyzed, the system flow of autonomous control based on image fusion, which are divided into three parts, i.e. situation awareness, dynamic replanning, and mission conduction, where situation awareness includes image preprocessing, image registration, image fusion, image recognition and target location, and dynamic replanning includes strategy selection and real-time route planning.2) The methods of image fusion preprocessing based on DWT (Discrete Wavelet Transform) and FDCT (Fast Discrete Curvelet Transform) are proposed. Aiming at the shortcomings of current image denoising methods that can’t realize the best denoising effect, the image denoising method based on multi-objective optimization is researched, the multiple criteria for image denoising are presented, the methods of image denoising based on DWT and Fast DCT are proposed. The experimental results show the two method can realize the Pareto optimal denoising effect, but the two methods have their advantages themselves, the speed of DWT methods is higher, while the effect of FDCT methods is better. On condition that the imaging environment of AUAV is invariable, the thresholds of denoising can be set a fixed value, which can meet the requirement of applications, and give the effective solutions in time. 3) A multi-objective optimization image fusion is proposed. The effective evaluation metrics is proposed from the view of image quality, the relation of the fused image to source images, and the relation of the fused image to the standard image. New conditional mutual information is proposed, which can avoid the information overloaded. The evaluation metrics are selected and designed according to the metric relativity and a certain selection rules. The multiple metrics are the optimization objectives, and the fusion parameters are selected as the decision variables and optimized by a multi-objective optimization algorithm. The fusion models in space domain and DWT domain are uniformed and simplified, two fusion rules are designed, finally the experiments for multi-focus image fusion, blind image fusion, multi-resolution image fusion, and color image fusion are conducted, and the methods and evaluation metrics of image fusion is designed according the characteristics of different images. Experimental results indicate that the fusion method based on multi-objective optimization is suitable for many types of pixel-level image fusion and can realize the Pareto optimal image fusion.4) The method of optimizing the decomposition levels for wavelet-based image fusion is presented and the selection of wavelet bases is analyzed. In order to realize the optimal wavelet-based image fusion, the decomposition levels and wavelet base selection are both important. A multi-objective optimization approach to determine the optimal number of decomposition levels is presented. In the experiments it is found that the optimal decomposition level is not a fixed value, but rather, changes with the characteristics of the original images. In general, fusion of images with larger resolution requires a higher number of decomposition levels. The experimental results also show that using the multi-objective optimization can effectively obtain the optimal number of decomposition levels make each metric be maximized. For there exist too many wavelet bases, the selection criteria of wavelet bases are given. In order to make more meaningful results, simplify the computation, and avoid the reconstruct distortion, the "Haar" wavelet is selected the wavelet base.5) A method of target recognition and tracking based on deformable templates in a fused image is proposed. First the advantage and challenges of deformable templates in target recognition and tracking are analyzed. Because the deformable template has formalized description parameters, it can define different targets in all kinds of shapes. Then different target templates are designed, and the energy functions are defined, the template matching is converted into single objective problem. The target recognition based on the deformable template is presented, which can reach the target matching through optimizing the energy functions, and the different methods of target location are analyzed. The target tracking based on the deformable template is also researched, and experimental results show that the method based on the deformable template can search, recognize, locate, and track the target quickly. 6) A novel multi-objective constriction particle swarm optimization (MOCPSO) is presented. Using different multi-objective optimization algorithms as reference, MOCPSO is proposed, which not only uses mutation operator to avoid earlier convergence, but also uses a new crowding operator to improve the distribution of nondominated solutions along the Pareto front, and uses the uniform design to obtain the optimal parameter combination. The sound evaluation criteria for multi-objective optimization algorithm are given, and some typical test functions are introduced. Experimental results show that MOCPSO has faster convergent speed and better search capacity than other multi-objective particle swarm optimization algorithms, especially when there are more than two objectives. MOCPSO is suitable to solve different multi-objective optimization problems.更多还原