【作者】 张蕾；

【导师】 普杰信；

【作者基本信息】 河南科技大学 ， 计算机应用技术， 2008， 硕士

【摘要】 计算机视觉在众多领域都有广泛的应用,比如家庭智能机器人、仪表自动监测、汽车低速自动导航驾驶和航空图片中的物体识别,并且随着计算机视觉技术的发展,计算机视觉将具有更广泛的应用前景。而计算机视觉的重要研究课题之一是物体识别,并且特征提取和分类是物体识别的关键步骤。在识别物体的方法和过程中,还存在许多问题和挑战,比如如何从2D图片中快速而准确的识别出物体。人类的视觉系统就能够轻易地快速识别2D图片中的物体,这实际上是一个由2D信息出发辅以先验知识识别物体的过程。本文就从物体的形状信息出发,提出一种基于改进BP神经网络的物体识别方法。在特征提取方面,利用矩算法提取物体的不变性特征,并详细讨论了Hu矩及其修正算法。不变矩方法,能够反映物体的形状信息,并具有较好的抗噪性能,同时因不受被识别物体大小、位置、方位的影响而被广泛应用于物体识别、景物匹配、图像分析及字符识别等许多方面。并且修正的Hu不变矩,不管在连续的状态下还是在离散状态下都对平移、缩放、旋转具有不变性,而且具有较小的时间复杂度,可以用来有效的识别物体。本文在MATLAB实验环境下对修正的Hu不变矩算法进行了实现。在分类识别方面,先分析了BP神经网络的结构,算法,存在的缺点并提出加入动量项、共轭梯度法、正则化方法、弹性BP算法、自适应学习速率动量梯度下降反向传播算法,这一系列改进的学习算法,以满足解决不同问题的需要。其中自适应学习速率动量梯度下降反向传播算法,可以有效避免BP网络收敛速度慢和存在所谓“局部最小值”问题。最后在MATLAB实验环境中,将该改进后的BP算法用于识别Coil-20(columbia object image library)图像数据库中的物体。并且该实验是在无噪声和有噪声两种情况下分别进行的。与基于传统BP算法的物体识别方法进行实验比较,该改进后的BP算法进一步提高了BP神经网络在处理非线性和不确定因素问题上的能力,并且该改进算法无论是在无噪声情况下,还是在有噪声情况下,都比传统的BP算法具有更高的识别率和更快的收敛速度。从而证明了该算法的可行性、鲁棒性和有效性。更多还原

【Abstract】 Computer vision finds wide application in multiple areas such as household intelligence robots, automatic inspection of instruments, automonous navigation of automobiles and object recognition of aero-sensing images, and sees a brighter prospect in the bloom of a variety of related techniques, Yet object recognition is one of the most important problems in computer vision.Furthermore,to object recognition, feature extraction and classification are pivotal.Object recognition faces a great many challenges in method.One problem is how to identify the object rapidly and accurately from two dimensional images.This task could easily resolved by human visual system which infers from an image of two dimensional and apriori assumptions. This paper puts forward an algorithm of object recognition based on modified BP Neural Network from the shape information of the object.About feature extraction,invariant feature is extracted by moments.Hu moments and the modified algorithm are dicussed amply in the paper.Invariant moments can reflect the shape of object,possess the capability of resisting noise,and is not influenced by the size, position and orientation.Therefore, invariant moments are widely applied on object recognition, scenery matching,image analysis, character recognition,and so on. The modified Hu invariant moments are invariant to the translation , rotating and scale of object,when object is in sequential state or discrete state.Moreover they have small time complexity.So they can recognize object efficiently. This method is examined in the MATLAB laboratory environment.About classification and recognition,the structure,algorithms and shortcomings of BP neural network are introduced.Moreover,adding momentum,conjugate grads, regularization, stretch BP algorithm, back propagation algorithm based on self adapting learning rate with momentum gradient reduction are presented in the paper to meet different problems. Especially back propagation algorithm based on self adapting learning rate with momentum gradient reduction can avoid efficiently the slow convergence and the problem of‘local infinitesimal value’of BP network. The modified BP algorithm is examined in the MATLAB laboratory environment, to recognize objects in Coil-20.The experiments are performed in the noise-free environment,as well as in the noisy environment.Compared with object recognition based on traditional BP algorithm,the modified BP algorithm improves the capability of processing nonlinear and uncertain problems.and it has higher recognition rate and quicker convergence in the noise-free environment as well as in the noisy environment. Consequently the feasibility, robustness, efficiency are proved in the paper.更多还原