【作者】 安国平；

【导师】 吴庆宪；

【作者基本信息】 南京航空航天大学 ， 模式识别与智能系统， 2008， 硕士

【摘要】 现代战争中,空中精确打击和空地一体化作战已经成为最重要的作战形式。因此,对空中目标的识别就显得非常重要。本论文针对飞机目标,在基于DSP的图像处理平台上进行了飞机目标图像处理与识别技术研究。主要完成了以下几个方面的工作:首先分析了本课题国内外研究现状和发展方向,对DSP技术发展进行了简要介绍,着重探讨了DAM6416P在图像处理与识别应用过程中程序结构的软硬件设计和数据传输问题。其次研究了了飞机目标的分割和运动目标检测技术。综合利用图像灰度及邻域空间信息,研究了基于进化规划的二维最小交叉熵法和基于神经网络的图像分割法;分析了常用的运动目标检测算法的优缺点,研究了基于自适应混合高斯模型的目标检测技术。实验表明,这些方法能较好的提取飞机图像,为目标识别做好了准备。接着研究了对分割目标的特征提取,主要进行了仿射不变矩,相对不变矩以及修正后的Hu不变矩特征提取,将几种基于矩的特征提取方法应用于目标识别,对实验结果进行了对比与分析。然后将目标特征向量输入到分类器中训练,再运用测试样本进行测试,得到了良好的识别结果。识别方法主要运用了自组织特征映射神经网络和支持向量机技术。自组织特征映射神经网络主要对网络模型进行了改进;支持向量机主要研究了改进的序贯最小优化算法和参数初始化设置问题。最后简单地总结了本课题的内容,并从硬件及软件方面对本课题的研究方向进行了展望。更多还原

【Abstract】 In the modern war, precision air striking and the integrated air-to-ground combat have been the most important combat form, so it is getting more and more important to recognize the sky goal. Directed against the aircraft, this paper has discussed the objective research aircraft based on DSP image processing platform. Following the completion of several major aspects:Firstly, research actuality and development direction of this task have been analysed. The technical development about DSP, the procedure of structure design about DAM6416P and the problem of data transmission have been introduced.Secondly, this paper have studied segmentation technique of plane target, including the two dimentional minimum mix entropy based on Evolutionary Programming and neural network methods. Analysed the advantages and disadvantages of moving objects detection algorithm. Then the algorithm based on mixture gauss model have been studied. The test shows that these methods can distinguish the object and background perfectly to make preparation for the following recognition.Thirdly, the feature extraction algorithms have been studied. The improved Hu invariable moment, the relative invariable moment and the affine invariable moment have been discussed in this paper. Then using feature extraction algorithm to targets recognition, analyse and compare the results.Fourthly, target recognition has been studied in which target samples are first input to interpolator to be trained, then good recognition results have been gotten by using testing samples. Self-organizing feature map neural network and support vector machine methods have been provided. In the self-organizing feature map neural network, mainly the improving of model structures has been introduced. In the support vector machine, mainly the improving of sequential minimal optimization method and the initial algorithm of parameters have been introduced.Finally, primary content of this task has been simply introduced and research aspect has been expounded.更多还原