【作者】 杨柳；

【导师】 张红兵；

【作者基本信息】 兰州交通大学 ， 模式识别与智能系统， 2010， 硕士

【摘要】 伴随着计算机技术的不断开拓和对人类感知机理研究的不断发展,智能移动机器人的先进控制成为了学术界关注的热点。赋予机器人“看”的能力是机器人视觉研究的主要问题之一。机器人利用视觉实现对环境的识别是一个复杂的过程,因此,研究识别率高、抗噪能力强、鲁棒性好的图像识别方法对移动机器人研究领域具有重要意义和理论价值。本文首先对机器人采集环境图像的预处理方法进行了研究：(1)研究了对原始图像的增强方法。采用分段线性函数和直方图处理对光线暗的图像增强对比度；然后对含噪声的模糊图像进行去噪研究,对比了常用的几种去噪方法,并采用了一种改进的阈值折中小波去噪的方法进行降噪。实验证明,该方法能够得到较理想的降噪效果。(2)采用了不同的模板对图像进行边缘提取以及二值化处理,并进行了实验比较及分析,为后续图像识别做好了充分准备。其次,对环境图像识别、分类的方法进行了研究：(1)采用SIFT (Scale-invariant feature transform)特征提取的方法,建立图像金字塔,提取图像的边缘特征,然后采用特征点之间的最近欧氏距离进行识别。而后在此基础上,本文从实时性方面进行了改进：①在特征点进行欧氏距离的匹配的时候加入一个阈值,将那些不匹配的点可以直接滤除,不必进行复杂的128次差平方运算；②在建立图像金字塔的时候没必要一直进行到最后一组足够小为止,本文通过实验得到第4组的时候特征点的总数就不再增加了,不需要一直计算下去,所以节省了运行的时间。实验证明了这两种方法的有效性。(2)本文还研究了基于Gabor滤波器和不变矩的神经网络的图像分类方法,研究了Gabor滤波器以及不变矩提取图像特征的原理,使用Gabor滤波器组提取32个局部特征,不变矩提取7个全局特征,将特征值通过BP神经网络学习,最终实现对环境图像的分类。仿真实验验证了该算法提高了识别精度的同时也提高了识别的鲁棒性,满足智能移动机器人对环境识别的要求。更多还原

【Abstract】 With the development of computer technologies and human sensation researches,the advanced control of intelligent robot has become a hot spot in academic.Giving the robot the function of look is one of the main questions of robot vision research. And the robot use vision to recognition the environment is a complex process.Therefore, the research of real-time vision recognition system with a high recognition rate, a strong anti-noiseability and highly robust performance has a vital significance and the theoretical value.First of all in this dissertation,the environment image’pretreatment methods are studied:(1) The enhancement methods of original environment images are studied.Though using piecewise-linear function and histogram processing,the light-dark images’contrast are enchanced.Then, for the fuzzy images with noise,comparing several general used de-noising method,an improved wavelet mid-threshold denoising method is proposed.The experiment result shows that this method can achieve superior de-noising performance.(2) Image edge extraction based on a variety of templates and binarization processing methods are studied,and the experiment result is compared and analyzed. This process is well prepared for follow-up image recognition.Secondly,the environment images recognition and classification methods are studied:(1) Using SIFT(Scale-invariant feature transform)feature extraction methods, building the image pyramid, extracting image’s edge features and then the recognition between the two images is though their features’ Euclidean distance.In addition,on the basis of SIFT,the dissertation has two improvemwnt in terms of real time:first,a new threshold is added to the process of the match between the two features,in order to exclude directly the features which are not matched,so that the complex 128 deviation square operation is no need to be calculated.Second,it is not need to continue until the last enough small group when the image pyramid is established,and though the experiment it is found that the total number of features will no longer be increased over the first 4 group.Therefore,this method saves running time.Simulation experimental results show the two methods are effectiveness.(2) Neural network classification method based on Gabor filter and moment invariant is studied in this dissertation.The principle of image feature extraction based on Gabor filter as well as moment invariant is described.32 local features are extracted though Gabor filter and 7 global features are extracted though moment invariant.Then,the 39 features value is learned by neural network and adjustable parameters of neural network are trained to optimize.Simulation results show that the algorithm improves the recognition accuracy at the same time also cuts down the recognition time.This algorithm is able to meet the requirement to recognize the environment for intelligent mobile.更多还原