【作者】 李超；

【导师】 刘陈；

【作者基本信息】 南京邮电大学 ， 电子与通信工程（专业学位）， 2018， 硕士

【摘要】 目标定位和识别是机器视觉研究中的一项重要内容。目标定位和识别应用在生产线上,不仅能够高效地实现工件检测和筛选,而且能够更好地掌控生产模式,有利于工业生产向智能化方向转变。模板匹配作为目标定位和识别操作的常用手段,已经被广泛应用于工件缺陷检测和测量、目标跟踪等场合。但常用的模板匹配算法对于非线性光照、遮挡等敏感问题,以及目标存在旋转、缩放的状况,效果并不理想,不能应用在复杂的工业生产环境。本论文研究了基于梯度矢量的边缘特征匹配算法,利用图像的边缘能直接反映物体的轮廓和拓扑结构信息,以及对非线性光线不敏感的特性,匹配出物体位置。并基于Xilinx Zynq-7000平台,通过软硬件协同设计的方法,将边缘特征匹配算法进行软硬件拆分。其中边缘检测模块在Zynq平台的可编程逻辑部分实现,模板训练模块和图像匹配模块则在Zynq平台的处理器系统中实现,并通过作业文件的方式动态配置算法输入的模板区域、搜索区域等相关配置参数,具有较高的实用价值。本论文主要内容包括以下几个方面:(1)介绍了当前常用的匹配方式,并详细描述了算法匹配原理、图像金字塔加速搜索原理以及旋转、缩放情况下模板匹配的解决思路。(2)通过研究Zynq平台的技术特点与开发流程,以及sobel边缘提取算子的理论知识,利用高层次综合工具Vivado HLS设计实现边缘提取操作的硬件加速。(3)利用后台PC界面配置模板区域、搜索区域等相关配置参数,生成XML格式的作业文件,下发到智能相机中。智能相机通过对作业文件进行数据解析,动态修改Zynq中图像处理流水线的算法参数,实现算法功能的调整。(4)在Zynq-7000的ARM部分实现边缘特征匹配算法的软件设计,详细描述了算法的具体实现细节如边缘特征点的筛选、旋转状况下的模板处理、匹配过程等。最后,对算法在ZC702开发板上进行了软硬件联合测试,得到非线性光照、遮挡、旋转、缩放情况下的性能数据。数据表明,基于梯度矢量的边缘特征匹配算法基本满足工程需要,对非线性光照、遮挡、旋转、缩放等问题,能够得到较好的结果。更多还原

【Abstract】 The location and identification of target are important parts of machine vision research.The applications of location and identification in the production line can not only achieve accurate and real-time identification which is online to improve production efficiency but also conducive to the production of information management to adapt to the changes in the form of modern production better.Template matching,an effective method in the field of target’s location and identification,has been widely used in workpiece localization,defect detection,target tracking.However,the common algorithm for template matching are not effective for the sensitive problems such as nonlinear illumination and occlusion,as well as the rotation and scaling of the target.The results are not ideal and can not be applied in complicated industrial production environments.The thesis choose the matching algorithm based on edge feature’s gradient vector to research.It can be used to match objects because of the edge of the image which can directly reflect the contour and the insensitive property for nonlinear illumination.Based on the Xilinx Zynq-7000 platform,the matching algorithm of edge feature is split to the software part and the hardware part based on the software hardware co-designe.The edge detection module is implemented in the programmable logic part of the Zynq platform.The template training module and the image matching module are implemented in the processor system of the Zynq platform.It has high practical value to dynamically configure parameters such as the templated area,the search area by job file.The main contents of this thesis include the following aspects:(1)The thesis introduces the matching method which is commonly used at present,and describes the algorithm’s matching principle,image pyramid acceleration search principle and the solution of template matching in rotation and scale in detail.(2)By studying the technical features and development process of Zynq platform and the theoretical knowledge of sobel edge extraction operator,implementing the hardware acceleration of edge extraction operation by using Vivado HLS which is high-level synthesis tools.(3)Through the background PC interface,we can configure related parameters included templated area,search area to generate job file based on xml format.Then we send the job file to the smart camera to resolve file data.The smart camera can dynamically modify the algorithm parameters in the image processing pipeline to achieve the adjustment of the algorithm functions.(4)Achieveing the software design for matching algorithm of edge feature in the Zynq-7000’s ARM.It describes the implementation details such as the edge of the screening,the template processing under rotation,matching process.Finally,the algorithm is used in hardware and software joint test by the ZC702 to get the performance data with the nonlinear illumination,occlusion,rotation,scaling.The data shows that the matching algorithm based on edge feature’s gradient vector meets the needs of the project.It can obtain good results in the nonlinear illumination,occlusion,rotation and scaling.更多还原