【作者】 高阳；

【导师】 李言俊；

【作者基本信息】 西北工业大学 ， 导航、制导与控制， 2006， 博士

【摘要】 作为一种自主式智能精确制导技术，红外成像制导在现代空战中发挥着越来越重要的作用，而自动目标识别则是其中的核心技术。但目前目标识别技术还有许多困难需要解决，而且这些困难分别存在于识别系统中的预处理、特征提取和分类识别等环节中。在此背景下，本文以红外成像制导的图像处理分析和目标识别为主线，针对各个环节所存在的困难，系统研究了目标识别系统中的图像滤波、目标分割、二维目标特征提取、三维目标特征提取和分类识别等问题。论文所取得的主要研究成果如下：1．在图像滤波技术中，现有滤波方法存在无法同时实现噪声去除与边缘保持的问题。针对这个问题，本文提出了一种各向异性高斯分段滤波器，在对滤波器的参数进行估计的基础上，使滤波器仅在有效邻域滤波。但是各向异性高斯分段滤波器又存在着计算量较大的缺点，为此本文又提出了一种分段矩滤波来解决这些问题，它在快速局部矩卷积算法和边缘阶梯模型的基础上实现了快速的有效邻域滤波，计算机仿真结果表明该方法能在有效去除噪声的同时较好地保持边缘。2．图像分割是目标特征提取的基础，针对现有自适应阈值分割方法在非均匀背景下不能有效分割出目标的问题，提出了一种基于背景拟合的图像自适应分割方法，首先根据噪声统计信息估计出图像的噪声均方差，然后据此对图像背景进行了光顺限制的曲面拟合。通过与其它自适应阈值方法分割结果的比较表明，在非均匀背景下本文所提出的这种背景拟合分割方法具有较好的目标分割效果。3．在实际目标识别系统中，为了提高计算速度并简化系统常采用二维目标识别方法。本文针对常用的二维RST(旋转、尺度、位移)不变特征维数小且对噪声往往不稳定、而Trace变换较稳定但计算量过大的问题，提出了一种极坐标投影矩，它可以通过增减不变函数来调整不变量的特征维数。仿真实验结果表明这种不变量在尺度、旋转及噪声变化情况下仍具有很强稳定性，而分类实验则表明其具有良好的分类效果。4．针对二维目标识别系统中常用的仿射不变特征存在着特征维数较少且不变性不稳定，而规格化轮廓虽稳定但匹配过程复杂等缺点，提出了一种仿射投影矩特征，这种特征通过规格化和尺度归一化方法实现了对仿射变换的不变性，而且其特征维数的大小同样可以通过改变不变函数的数量来调整。计算机仿真结果验证了这种仿射投影矩对仿射变换的稳定性，并且分类实验表明了这种仿射投影矩的良好分类效果。5．针对三维目标识别系统中常用的正交矩没有分析特征之间相关性的情况，本文通过将K-L(Karhunen-Loeve)变换引入正交矩得到了具有最小相关性的最优正交识别矩——K-L傅立叶矩，但是K-L变换存在着正交基不固定、且对多类别数据识别效果不佳的缺点。为此，本文基于余弦变换是一阶马尔可夫过程中对K-L的最优近似的理论，提出了用余弦傅立叶矩来提取特征的方法，通过将余弦傅立叶矩分解为独立的余弦变换和傅立叶变换的方法，从而提高了计算速度。此外理论分析和实验仿真表明，余弦傅立叶矩的不变性对RST变换具有很强的稳定性，并且这种余弦傅立叶矩对于分类识别也有较好的效果。6．在分类算法中，支持向量机因能对样本进行结构风险最小化的分类而得到了广泛的应用。但现有多分类支持向量机方法在处理目标类别较多的情况时，存在着分类器训练速度慢、分类器较多、决策结构复杂且不稳定的问题。针对这种情况，本文提出了一种基于风险分析的最小随机风险多分类方法，当使用同类型的支持向量机对目标进行分类时，决策导向非循环图结构下的系统随机结构风险最小，计算机仿真结果表明其具有良好的分类效果。此外，将余弦傅立叶矩和支持向量机结合，用于对三维目标识别也取得了较高的识别精度。更多还原

【Abstract】 As an autonomous intelligent accurate guidance, infrared imaging guidance is playing an increasingly important role in air war. In this system, targets recognition is key to guidance processing. But there are many problems need to be solved in targets recognition and these problems lie in different parts, such as preprocessing, feature extraction and classification. Under this background, this paper is organized on infrared image analyzing and targets recognition. To solve these problems, the image filtering, target segment, feature extraction of 2D and 3D targets recognition are investigated systemically in this dissertation. The important contributions and creative achievements are summarized as follows:1. As the existent filters haven’t good effect in denoising and preserving edge, an anisotropic piecewise gaussian filter has been put forward. In this method, the parameters of filter are estimated to make it smooth only in effective neighbor. But the computation amount of this anisotropic piecewise gaussian filter is large, so a piecewise moment filter is developed in this dissertation. Based on fast convolution and stair edge model, the moment filter can make a fast smoothing in effective neighbor. The experiment result shows that this new method has better capability than other common methods in suppressing noise and preserving edge.2. Image segment is the base of feature extraction, but the standard segmentation methods are ineffective for extracting target in nonuniform background. To solve the problem, a surface fitting segmentation method is proposed. Firstly, a statistical model of noise has been used to estimate the variance of noise. Then the fairing constrained surface fitting is made for the image background. The experimental result shows that the new method has better capability than other common methods in removing background and extracting target.3. To increase the speed of computation and simplify the system, 2D targets recognition system is often used. In this dissertation, a polar-projection moment is proposed to solve the problems that the invariant features aren’t stable and the computational amount of trace transform is large. In the polar-projection moment, dimension of these features can be adapted by the number of invariant functions. The simulation shows that these features have mighty stability when the target have scale, rotational and noise changes and classification experiment demonstrated that these features have good effect in classification.4. Furthermore, affine invadant features also have the problems that invariant features aren’t stable and the matching process of normalized contour is complex. To solve these problems, a affine-projection moment is proposed. These features can realize invariance by normalization and scale average. After that, dimension of features also can be adjusted by the number of invariant functions. Simulation explains that these features are stable and the classification experiment displays that the data can be classified well by this affine projection moment.5. To get the least information redundancy of orthogonal moment, a K-L fourier moment is put forward by involving K-Ltransform to get the best orthogonal moment in recognition. But the K-L orthogonal bases are not fixed orthogonal bases and it is not good at recognizing the multi-classification data, so a cosine fourier moment is proposed. Because the cosine transform is the best approximation to K-L transform in one order markov process, the cosine fourier moment is the suboptimal orthogonal moment. As the cosine fourier moment can be divided into independent cosine transform and fourier transform, the speed of computation can be increased. The analysis in theory and experiments show that this feature can obtain stable invariance and classification ability.6. To reduce the excessive sub-classifier and complex structure of SVM multi-classification, a minimal random risk method is proposed by analyzing the risk of structure. When the same SVM classifiers are used in classification, the random structural risk of system is minimized. The experiment presents that it can get better result than DAG method and it’s speed is high. Finally in the experiment, the recognition system composed of cosine fourier moment and SVM gets higher classification accuracy in processing 3D target.更多还原