【作者】 刘钊；

【导师】 孔建益；

【作者基本信息】 武汉科技大学 ， 机械设计及理论， 2011， 博士

【摘要】 随着信息获取和信息处理技术的快速发展,计算机视觉,即如何利用计算机技术高效准确地从环境图像或者视频中获取相关信息,进而对客观世界中的事物及发生的现象进行分析、判断和决策,已经成为一个非常重要的研究课题。生物的视觉和智能,目前还是人类认知的处女地。在生物视觉研究中已经发现大量有趣的现象、结构和功能,但是对于总的工作机制和机理还知道的很少。从信息处理的角度,对于已有的关于生物视觉和智能的研究成果进行总结、分析与综合等研究,必将对计算机视觉和人工智能的发展产生强有力的推动作用,同时计算机视觉和人工智能的研究能为生物视觉和生物智能的工作机理提供重要线索。计算机视觉基于物理计算设备。七十年来,对于计算设备的能力上限,主流的科学家持悲观的态度。究竟什么是计算?计算到底能干什么?这是计算机科学、人工智能、计算机视觉的基本问题,是彻底解决计算机视觉和人工智能的前提条件。自适应性和智能化已经成为困扰当前计算机视觉的最大障碍。如何在复杂的背景下,从样本图像和指定目标中自主获取目标的模式,并利用这些模式对未知环境和图像进行分析和判断;如何利用计算产生自适应性;视觉和智能在信息处理的层面上究竟有什么样的关系等问题,是彻底解决计算机视觉的关键。本学位论文从生物视觉中总结出能代表视觉信息处理的共同特征的、计算机视觉可利用的生物学证据;在此基础上,提出并论证了演化计算具有比算法更强的表达能力的观点;将演化计算、人工神经网络方法、并行计算等技术融合起来,找到一种适合于目标识别过程中使用的知识表示、知识进化、知识使用的新的计算智能方法;将该方法应用于机器视觉技术研究,解决带钢表面缺陷检测过程中的准确度和实时性问题。本文研究工作主要贡献是:(1)生物视觉对于高等哺乳动物视觉领域中已经取得的一些相关研究成果进行归纳、总结和假设,从中找出能代表视觉信息处理的生物学证据,包括视觉与智能的同源性、演化机制、并行处理机制、自适应机制、层次处理机制、和竞争机制等等。(2)演化计算基本理论指出了算法和图灵机的缺陷,对于“图灵机是任意物理计算设备的精确模型”的论断提出了质疑;提出了演化计算的概念和相应的自动机模型,并对它的表达能力进行了论证,证明了演化计算方法比算法或图灵机的表达能力更强。(3)基于演化计算的模式识别将演化计算用于结构模式识别和统计模式识别,将演化计算方法用于统计模式识别的特征选择、分类器参数,以及结构模式识别中的基元结构、基元关系选择、基元关系提取。该方法不仅使得编程人员的工作强度大大降低,而且得到的参数更为合理。(4)计算机视觉仿真实验平台设计为了提高视觉系统的开发速度,便于程序的学习、训练、演化,提出将虚拟现实技术引入到计算机视觉系统的设计和知识获取过程之中,设计并建立缺陷检测视觉仿真实验平台,以及目标跟踪视觉仿真实验平台。在仿真实验平台可以进行目标识别、分类训练、缺陷知识获取等研究和开发,具有安全、经济、可重复及不受场地条件限制等优势。(5)基于演化计算的目标识别采用演化计算的方法获得关于钢坯板表面缺陷目标的显性知识和隐性知识,并利用GPU并行计算对表面缺陷目标进行分割、分类和判断;提出了基于互信息熵的钢坯板速度检测方法。更多还原

【Abstract】 With the enormous increase in popularity of information acquisition and processing technologies, Computer Vision, which effectively analyzing images or videos from our real world, getting information we are interested in, figuring out what is happening and what should to do, has become a very important research topic.Biological vision and intelligence full of mystery. Research on biological vision has found a large number of interesting phenomena, structures and functions, however, the working mechanism about them be known is very little. Summary, analysis and synthesis of research results is bound to play a strong role in the development of computer science, artificial intelligence, and computer vision. Computer vision is based on physical computing devices. The maximum capacity of computing devices is mistrusted by the most scientists in the last 70 years. What is computation, and what can they do in the end? Theory of computation is the basic problem and precondition on which computer science, artificial intelligence, computer vision is to dependent.Adaptively or intelligence is the biggest obstacle in current computer vision area. Adaptively or intelligence is the biggest obstacle in current computer vision area. The answer to those questions, includes how to obtain models autonomously from images of complex environment, how to analyze and make decision under unknown environment by these models, how to get adaptively or intelligence by computation, and what kind of relationship it is between vision and intelligence when both of them act as information processing, is the key to solve the computer vision.This thesis summarized biological available evidence about vision and information processing from biological vision. Based on that, proposed and demonstrated evolutionary computation (EC) is more powerful methods than algorithm. A hybrid of EC and ANNs, in which knowledge representation, knowledge evolution, and knowledge utilization based on EC, are proposed. The method of computer vision based on EC is used in Strip surface defect detection. In summary, the works in this thesis can be categorized into the following four aspects.(1) Biological vision Introduce and summarize the related research about biological vision of mammals, find out biological evidence about information processing, including homology of vision and intelligence, evolution mechanisms, parallel mechanisms, adaptive mechanism, hierarchical processing mechanisms, competition mechanism and so on.(2) Theory of computation The defects of algorithm or Turing machine are proposed and proved that, Turing machine is not an accurate model of all physical computing devices. Put forward the concept of EC and its automaton models, which are more powerful than algorithm or Turing machine.(3) Pattern recognition based on EC Evolutionary computation is used in structural pattern recognition and statistical pattern recognition. Many problems, include pattern features selection, classification parameters selection in statistical pattern recognition, and primitives structure, primitives relationship choice, primitives relationship extraction in structural pattern recognition, are benefit from EC. This method not only allows programmers to greatly reduce the intensity of the work, but also to get more reasonable parameters.(4) Visual simulation platform To improve the development speed of the vision system and to facilitate the learning, training, and evolution process, a visual simulation platform for defect detection and a visual simulation platform for target tracking, which introduce virtual reality technology to computer vision system design and knowledge acquisition process, is proposed, design and build. Under the simulation platform, classification, training, knowledge acquisition in the research and development are more security, economic, repeatable and independent.(5) Target recognition based on EC To solve the problem of surface defects detection, EC method are used to obtain the explicit knowledge about defects, and ANN method to obtain the recessive knowledge about defects, and and GPU parallel computing is used in defects segmentation, classification and judgments. Proposed a plate speed detection method based on mutual information entropy.更多还原