【作者】 刘晓玉；

【导师】 方康玲；

【作者基本信息】 武汉科技大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 智能机器人之所以区别于传统的示教-再现型机器人,是因为其有感知外部世界变化并作出自适应调整的能力。在众多的机器人感觉中,尤以视觉应用最广。而在复杂工业现场环境下,存在诸如粉尘、光照变化、摄像机抖动、遮挡、信道传输噪声等干扰因素,不可避免地会对机器人的视觉产生影响,造成图像信号的降质。此外,复杂多变的工作环境,也给传统的需要标定技术才能完成的机器人控制器的设计造成困难。针对上述问题,本文以基于视觉的机器人智能分拣和抓取为应用背景,利用神经网络研究复杂环境下图像的恢复和目标识别以及机器人视觉伺服控制。神经网络固有的非线性逼近能力、自适应和泛化能力,以及联想记忆能力,使其成为贯穿本文各个研究部分的关键技术。其中,复值神经网络以其自然的复数处理能力,使得图像等常需频域处理的信号有了直接的表达和处理方式。为解决恶劣环境下图像降质严重的问题,本文利用复值Hopfield神经网络的联想记忆能力完成对降质图像的恢复。为此本文首先开展复值Hopfield神经网络的分析和综合研究。在此基础上,讨论复杂环境下图像识别和恢复的实现问题,最后基于模糊行为规则和神经网络对机器人智能视觉伺服控制器进行设计。归结起来,本文的主要研究工作包括:1)针对一类复值离散时间Hopfield神经网络,提出一种兼顾平衡条件和稳定条件的网络综合方法。该方法首先根据稳定性分析,得出系统局部渐近稳定性判决条件,而后通过求解网络平衡方程和调整平衡解中的激活函数增益,确保最后的网络综合结果满足稳定性判据,从而使每个待记忆模式在网络中是稳定且吸引的。2)针对一类复值连续时间Hopfield神经网络,提出一种吸引域受限条件下兼顾平衡条件和稳定条件的网络综合方法。该方法首先在复值域中构造基于吸引域参数的Lyapunov函数,而后开展系统的稳定性分析,并分别在网络时间常数已知和未知的情况下,进行吸引域受限条件下的网络综合研究。其中平衡方程的求解分别基于伪逆规则和奇异值分解技术,所得网络参数约束和给定的吸引域参数均包含在用线性矩阵不等式表示的渐近稳定判据中,所以网络参数易于求解,且所得结果能同时满足给定吸引域内记忆模式的稳定性和吸引性。3)针对一类可直接表示灰度等多值图像的复多值Hopfield神经网络,提出一种兼顾平衡条件和能量函数逐步递减要求的网络综合方法。综合中基于奇异值分解技术求得的网络平衡方程通解,依能量函数递减原则进行调整,并最终得到能同时满足稳定性和吸引性要求的网络权值解,保证了图像联想记忆的可靠性。4)针对复杂环境下图像易出现模糊、缺损、噪声污染等严重降质的问题,通过引入复多值单层感知机完成对降质图像中各种形态工件的图像识别;利用3)中提出的一种复多值Hopfield联想记忆器完成对工件降质图像的恢复,进而实现对工件上角点信息的可靠提取,为机器人的视觉伺服控制提供可靠的图像反馈信息;5)针对复杂多变的非结构化环境中,常规的基于图像雅克比矩阵的视觉伺服控制存在建模困难且适应能力差的问题,提出一种基于模糊行为和神经网络的两阶段机器人智能抓取控制方案。其中第一阶段的模糊行为控制模仿了人的控制经验,可完成对机器人手爪的粗略定位;第二阶段的神经网络控制则利用BP神经网络的非线性逼近能力,实现对从图像空间到机器人关节空间的复杂非线性系统建模,完成对机器人手爪的准确位姿控制。更多还原

【Abstract】 Intelligent robot has the ability to perceive and acclimatize itself to the environment around it. So it is very different from conventional playback robot. Among the various senses of robot, vision is the most popular one. However, when robot works in complicated industrial environment, there are many disturbance factors to degrade the image signals, such as dust, variation of illumination, vibration of camera, shielding, noise in signal channel, etc. Besides these problems, the complicated and variable work environment makes it very difficult to design robot controller by conventional calibration technique. For these reasons, considering the sorting and griping manipulation of robot in the complicated industrial environment, this thesis discusses the problem of image restoration and recognition and visual servoing control for robot based on neural networks.Neural networks are herein the key technology permeating throughout this thesis for their abilities in nonlinear approximation, adaptation, generalization and associative memory, among which the complex-valued neural networks are capable naturally to process complex number, thus can be used to denote and process frequency signals directly, such as the image expressing in frequency domain.In order to solve the image degraded problem resulting from bad visual environment, complex-valued Hopfield neural network functioned as associative memory is utilized for image restoration in this thesis. Thus the analysis and synthesis of several kinds of complex-valued Hopfield neural networks are studied firstly. And based on such research results the image recognition and restoration are completed in complicated environment. This thesis also discusses the designation of visual servoing controller for robot by fuzzy behavior rules and neural network control. The main work in this thesis can be summarized as follows:1) Propose a synthesis method for a class of complex-valued discrete time Hopfield neural network, where the equilibrium condition and stability criterion are both satisfied. In this method, a local asymptotic stable condition is derived firstly and used to decide whether the solution of equilibrium equations is accepted. And if no, a gain-regulation of the neuron activation function is carried and guarantees finally both the stability and attractive ability for every storage patterns.2) Propose a synthesis method for a class of complex-valued continuous time Hopfield neural network under constrained attractive domain, which can satisfy the stable condition and equilibrium condition. In this method, this thesis constructs a Lyapunov function in complex-valued domain where the parameter of attractive domain is contained, and then analyses stability of the network. Considering two situations about time constants of the system, i.e. unknown or known, the corresponding solving algorithm is Pseudo-inverse and singular values decompose, respectively. The constraints of network parameters derived from equilibrium equations, together with the parameters of attractive domain, are denoted in the asymptotic stable condition which can be rewritten as linear matrix inequations. As a result, the network这“parameters can be solved easily and guarantee the storage patterns are both stable and attractive in the given attractive domain.3) Propose a synthesis method for a class of complex multi-valued Hopfield neural network used to express and process multi-valued information, for example, gray image where both the equilibrium condition and the requirement for energy function decreasing are considered. In this method a general solution of network equilibrium equation solved by singular value decomposition technique is adjusted according to the rule of decreasing the energy function. The final network weights can satisfy both stability and attraction, thus guarantee the reliable association memory of gray images.4) Study the image recognition for workpieces with different shape where the images are badly degraded by blur, defect or noise pollution because of the complicated industrial environment. The solution for degraded image recognition is to apply a class of complex-valued single layer perceptron. The proposed energy decreasing synthesis method by 3) for complex-valued Hopfield association and memory is utilized for restoring the degraded image, which can provide more reliable corner point information of the workpiece for the next grasping manipulation of robot.5) Propose a fuzzy behavior and neural network based visual servoing control scheme for intelligent grasping of the robot by the reason that the conventional visual servoing control method, i.e. image Jacobian matrix based method, has difficulties in modeling and its adaptive capacity is bad when robot works in a complicated and multivariate environment. The first phase of the scheme is fuzzy control which simulates the human control experience and be used to rough position for the clamp holder of the robot. And the second phase is neural network based control which applies the nonlinear approximation capability of BP network to model the nonlinear mapping from the image space to the joint space of the robot, and thus to complete the accurate position and orientation control of the clamp holder of robot.更多还原