【作者】 张虹；

【导师】 陈文楷；

【作者基本信息】 北京工业大学 ， 模式识别与智能系统， 2004， 硕士

【摘要】 本文重点研究了小波分析及不变矩在图像目标识别中的应用。利用矩不变量进行形体识别一直是图像识别的研究重点。本文详细介绍了几何矩、Hu 矩不变矩的概念、性质及在图像识别中的应用，针对这几种传统不变矩的缺点：它们都是在整个图像空间中计算得到的图像的全局特征，容易受到噪声的干扰。本文引入了新的矩不变量—小波矩不变量，它将小波分析用于不变矩中，使其具有平移、旋转和比例不变性。利用小波矩不仅可以得到图像的全局特征, 而且也可以得到图像的局部特征, 还增加了小波对图像结构精细特征把握能力强的优点，因而在识别相似物体方面具有更高的识别率。本文以小波不变量为特征，并对所提取的特征进行优化选择，最后用优化后的特征与 BP 神经网络分类器相结合进行目标图像识别。从小波矩的整体性特点出发，本文提出用小波模极大值法进行图像预处理，提高目标区域分割的精度，并在此基础上进行图像区域分割。本文通过仿真实验，对三类飞机六十个样本进行识别，得出了 Hu矩、小波矩在是否含有噪声的情况下的识别结果并对识别性能进行了比较。实验表明，小波矩不变量的识别效果要优于 Hu 矩，尤其是对于有噪声的图像。更多还原

【Abstract】 This paper focuses on the research of wavelet analysis and moment invariants’application in image object recognition. The concept of geometry moment and Hu’smoment, its property and its application in image recognition are described. Thetraditional moment invariants have the defect: these moments are the wholefeatures calculated from the whole image space, which are apt to disturb by noise.Aim at the above defect, a new moment invariant—wavelet moment is presented,which apply wavelet analysis to moment invariant. Thus wavelet moment possessthe image object’s invariant to translation, scaling and rotation. By using waveletmoment invariant, not only the local feature of image object is obtained, but alsothe description ability for the fine features of image construct is improved. Thus,the higher recognition rate is obtained, especially similar images. In this paper,wavelet moments are used as the image features, then the features abstracted areoptimized and finally the features optimized are combined with BP neural networkclassifier to make object image recognition. Moreover, in the view of the whole,the method of local maximal modulus of wavelet transform is used for imagepre-process, based on which image region segment is made, to improve theaccuracy of region segment. In order to ensure veracity of the recognition result, this paper use 60 samples of3 classes of planes for training and testing. In the case of free-noise andnoise-added, we use Hu’s moment and wavelet moment as features to test therecognition veracity. The experimental results demonstrate that the recognitionaccuracy and anti-noise capability of wavelet moment have been much increasing.更多还原