【作者】 王雯霞；

【导师】 翁桂荣；

【作者基本信息】 苏州大学 ， 通信与信息系统， 2004， 硕士

【摘要】 如何提取图像的不变性特征，构造一个高效率的分类识别系统一直是计算机视觉研究领域的一项热门课题。 近年来，数学形态学已经被广泛应用于图像处理的各个领域。由数学形态学基本运算定义的形态谱不仅描述了图像的尺度分布，同时也间接描述了图像的形状。因此可以将其作为区别不同图像的特征参数。 本文中首先对数学形态学进行了简要的介绍，并利用其基本运算的性质，对形态谱关于图像的平移、旋转、缩放是否不变的特性进行了证明与实验分析。其次，鉴于对形态谱的上述分析，对形态谱的定义进行改进获得一种归一化形态谱，使之具有对图像的平移、旋转、缩放不敏感的特性，获得目标图像的不变性特征。对于无噪声的二值图像，归一化形态谱可以准确地描述图像的形状。然而对于有噪声的二值图像，归一化形态谱却无法对图像的形状进行准确地定量的分析。针对归一化形态谱对噪声的敏感性，采用了一种具有Robust特性的交替序列形态变换，即高阶形态谱。实验证明高阶形态谱同样具有对图像的平移、旋转、缩放不敏感的特性，同时也能有效地抵抗噪声对图像形状的影响。 在目标的分类识别方面，我们采用形态学与神经网络相结合的方法，分别以归一化形态谱与高阶形态谱作为识别特征，采用BP网络对样本进行训练和学习，实现对目标的分类与识别。 实验结果表明，采用形态学与神经网络相结合的方法，利用归一化形态谱与高阶形态谱，能很好地实现对目标物体的分类与识别。更多还原

【Abstract】 How to extract invariable features of images and construct a efficient system of classification and recognition is always a hot spot in the field of computer vision.Mathematical morphology has already been widely used in all fields of image processing in recent years. Pattern Spectrum based on mathematical morphology describes not only dimension distributing of an image but also shape of this image. So we can make it a parameter of distinguishing different images.In this paper, we introduce several operations of mathematical morphology and their characters at first. Then we prove and analyze that if pattern spectrum has invariable features for translation rotation and zoom of images. Secondly we provide an improved normalized pattern spectrum based on the above analysis. Of course this normalized pattern spectrum has invariable features for translation rotation and zoom of images. It can describe shapes of images well and truly except for those images those have noise. In order to resist its sensitivity to noise, we adopt another pattern spectrum-high-level pattern spectrum that is defined based on alternating sequential morphological transformation. Experiments prove that high-level pattern spectrum not only has invariable features for translation rotation and zoom of images but alsocan resist the influence of noise on the shapes of images.In the aspect of classification and recognition of targets, we combine mathematical morphology with artificial neural network. Here we utilize a back propagation algorithm (BP). And then we use the normalized pattern spectrum and high-level pattern spectrum separately as the characters of recognition and the inputs of BP. By training this map, we can classify and recognize all images.Experimental results show that the method that combining mathematical morphology with artificial neural network is also a good way to realize the correct classification and recognition.更多还原