【作者】 张阳；

【导师】 王宣银；

【作者基本信息】 浙江大学 ， 机械电子工程， 2011， 博士

【摘要】 在现代工程技术领域中,视觉信息采集与处理具有重要的科学意义。研究人员参考人眼视觉原理,发明感知视觉图像信息的传感器,建立基于计算机技术的图像处理理论体系,开创机器视觉这一崭新研究领域。机器视觉技术已经在先进制造系统、机器人系统、智能监控、航天军工等领域广泛应用。但是多数视觉平台仅实现基本视觉功能,与真实人眼功能差异大。由于人眼具有强大的视觉信息处理能力,因此对人眼生理特性深入分析将使机器视觉研究具有更加充实的理论基础,推动机器视觉技术的发展。本文针对人眼特性研究,提出并联驱动仿生人眼机构模型；对前庭系统图像稳定机制、视觉信息注意机制以及单眼立体成像功能等方面进行深入分析和应用研究,提出仿入视觉特性的机器视觉处理理论和方法,进行建模仿真和实验分析,主要内容如下：第一章,阐述本课题的相关研究背景、目的和意义,介绍了人眼视觉系统的基本组成及特性,对国内外仿生人眼系统及仿人眼视觉理论研究的内容和现状进行了阐述,最后分析了仿生视觉系统及拟人视觉应用研究所面临的主要问题,给出了本课题的主要研究内容。第二章,针对广义人眼系统功能特性进行分析,建立了仿生人眼机电系统模型；提出了仿生人眼结构的假设条件和模拟人眼的多孔道球支撑结构；依据仿生学设计原理,提出了一种基于柔性执行器并联驱动的仿生人眼结构模型,包含六条模仿眼外肌的执行元件、视觉传感器及姿态检测传感器,并且根据机械加工需求选取了尺寸系统参数；针对人眼最重要的扫视运动特性进行分析,利用并联仿生人眼运动学模型和扫视运动特性,以及所设计的仿生人眼机构对典型的扫视运动进行了模拟仿真,得到了双眼运动路径及执行器运动参数；第三章,针对摄像机异常运动对视觉连续采集的影响进行研究,分析了保证视觉连续采集稳定处理原理和方法,研究视频连续采集稳定原理。在人眼前庭动反射的启发下,分析人眼视频稳定的机理,研究了摄像机异常运动在成像平面上所反映出的规律,提出了异常运动短时预测方法。在平均分布的子空间提取特征,利用预测结果确定搜索起始点进行匹配运算,有效减小图像内容中运动目标对稳像结果的影响。分析视频稳定后导致图像部分边缘信息丢失的现象,研究人眼的视觉暂留机理,提出以时间间隔为依据对当前帧空域所丢失内容进行线性化叠加恢复,保证视频主观判断的完整性。第四章,对人眼视觉注意机制和视网膜成像的中心-四周机制进行分析,建立基于初级视觉特征融合的视觉注意机制计算模型,提出了一种基于非线性中心-四周多尺度特征提取的计算方法,得到分辨率非线性变化的各尺度图像提取特征图。根据各初级视觉特征的自身特点及对视觉注意机制的不同作用,提出了一种基于权重系数的显著图叠加方法。对注意感知机制在视觉信息压缩处理中的应用进行研究,提出一种基于图像注视区域的压缩系数可变图片压缩算法,利用人眼视觉成像的多分辨率特性,采用Log-Polar函数设定压缩系数过渡策略,有效的降低了图片的存储空间。第五章,分析人眼视觉注意机制对运动目标的初级特征提取特性,提出了利用高斯模型进行背景建模提取运动目标作为感兴趣区域的算法,将背景与目标分割。经典DPCM编码压缩模型中整幅图像量化矩阵固定,本文提出了一种依据关注目标设置压缩量化矩阵的算法,设计量化矩阵过渡策略,降低视觉噪声,通过编解码端传递运动区域坐标和几何信息等少量数据的策略,保证编解码端数据一致性。第六章,分析人眼立体视觉成像机理,研究分析单眼立体视觉功能,进行合理假设,提出了一种基于单目视觉的立体建模和测量模型。通过光学原理和几何学基础,分析单一像平面对空间目标进行立体成像的可行性。对立体测量中的特征提取等数字图像处理方法进行研究,提出了一种新的利用颜色空间特性的圆形特征标记提取与亚像素中心定位算法。依据空间目标特征点在多幅变倍图像中投影的几何关系和特性,提出了一种1-D空间特征匹配搜索算法,将二维图像内容的匹配降低到一维空间。建立虚拟摄像机及实验环境模型,通过建模参数与单目立体视觉图像处理算法结果进行比较,验证了算法的正确性和可行性；与传统立体视觉方法进行比较分析,提出课题算法的特点和适合应用场合。第七章对全文作了总结,阐述了本课题的研究结论和创新点,对后续研究工作做出了展望。更多还原

【Abstract】 Visual system of human-being provides more than 80% of the information we used during daily life, thus it means the eyes, which are the key component of the visual system, are extremely important. In the modern industry, the visual information, which is acquired by different types of sensor, plays an important role as well. With rapid development of the science and technology, researchers had invented and manufactured the vision sensor based on the inspiration of the human visual system. Such sensors could percept the light signal and information of the objects. With these sensors and digital process technology, machine vision, a fresh research area, boomed rapidly. Nowadays, it’s already widely utilized. In fact, human visual system includes a comprehensive information acquisition and processing mechanism. However, most of the visual systems inspired by human eyes are simple version. Analyzing the perception mechanism and properties of the human eyes could provide the theory fundamental and enhance its application in machine vision area. In the thesis, we would analyze the components of general human visual system, provided a bionic mechanical eye’s model. Then, with the inspiration of the human vestibular mechanism, a video stabilization algorithm is introduced. Meanwhile, the visual attention mechanism and stereo vision by monocular system are our research topics as well. The simulation and experiments are designed to prove the algorithms above. The main contents are as follow:In chapter 1, the related study background and significance are introduced. On the basis of analyzing the components and mechanism of human visual system, both the domestic and international research achievements are investigated broadly. Finally, we listed the important research problem and introduced our research topics.In chapter 2, the properties of the generalized visual system are analyzed. Under the hypothesis, a novel type of bionic visual system is designed. The new structure of the bionic eye is inspired by human eyes, which includes six linear motion actuators to simulate the muscles around the eye of human and sensors for vision and orientations of eyeball. The dimension of the structure is designed and optimized. Finally, we utilized the designed eye model to simulate the saccadic movement of human, the motion of each actuator are solved by kinematics.In chapter 3, the influence of video record during abnormal moving of the camera is introduced. Firstly, we analyzed the basic video stabilization theory and methods. Inspired by the using of vestibular system to stabilize and compensate the head motion of human, a novel video stabilizing method is proposed. The features are extracted from sub-area of the entire image, and with the prediction of the further moving vector, the influence of the moving features are decreased. After video stabilization, the bounder area of the each frame is missed more or less, by simulating the mechanism of human eyes, a novel data recovery method based on visual delay is proposed to display the whole image area.In Chapter 4, we analyzed the visual attention mechanism and center-surround visual sampling on retina. A model based on low-level visual features for visual attention is proposed. In this model, the feature extracting algorithm is non-linear methods and multi-scale image of each feature are acquired. The conspicuity maps are generated with adjustable scale number, which is based on size of the area. The application of visual attention mechanism is analyzed and an image compression algorithm based on visual attention area and conspicuity maps is proposed. In that algorithm, the compression rate is adjustable, which could decrease the memory consumption.In chapter 5, the perception mechanism of human visual system for moving objects is analyzed. Based on the properties of the video content, the moving objects segmentation algorithm by Gauss model is proposed. In the typical DPCM video compression model, the quant matrix for compression the block is static. In this thesis, a quant matrix scaled algorithm based on content is proposed. The quant matrix slides from region of interested to background consecutively. The method proposed could reduce the memory request of compressed video.In chapter 6, models of stereo vision system of human is introduced. Compared with widely used binocular type for stereo vision, in the thesis the stereo cues for monocular vision is investigated. By the hypothesis, a novel model for 3D measurement by single camera is proposed. The digital image processing algorithms are important in stereo vision. A novel feature extraction algorithm based on color space and circle shape is proposed. After analyzing the properties of the zoomed images, the 1-D matching algorithm could reduce the 2-D full searching to 1-D on a selected line segment. The virtual environment is constructed for simulation and experiment.In chapter 7, the major work of the thesis is summarized. The conclusion and innovations of this thesis are introduced. Finally, the future development topics are presented in order to provide guidance for researchers, who are interested in such kind of projects更多还原