【作者】 肖宇；

【导师】 温伟刚；

【作者基本信息】 北京交通大学 ， 机械制造及其自动化， 2020， 硕士

【摘要】 阵列孔板广泛应用于医疗、生物、化学等领域,常作为移液设备的操作对象。目前,大多通过人工或相关设备将试剂瓶或液体放入孔板,但设备大多通用性差,仅适用于某种或某几种孔板,因此实现对阵列孔板的非接触测量和定位将大大提高设备的工作效率和智能化程度。随着机器视觉在各领域的广泛实践和应用,以及深度学习技术在目标分割定位方面的较大优势,很多学者越来越倾向于使用深度学习技术结合传统图像处理技术解决实际工程应用问题。目前各视觉测量系统针对特定对象设计特定方法进行测量,无法满足对本文对象的测量需求。本文以阵列孔板为对象,从阵列孔板的区域提取、特征点提取和立体匹配方面展开研究,最终实现对阵列孔板的测量和定位。1.针对孔板区域的提取,提出了适用于本文的改进Mask RCNN方法。改进Mask RCNN以Mobile Net为backbone、以P4+P5两层特征层为FPN。基于对Mask RCNN算法的研究,结合本文分割定位的难易程度,提出改进措施,并利用自制的孔板数据集,分别对三种backbone和不同的FPN层进行实验,实验结果表明了改进的Mask RCNN能有效地提取孔板区域。2.针对阵列孔板的特征点提取,提出了优化的边缘提取和拟合的方法。首先对图像增强算法进行研究,选用SSR方法进行图像增强;然后对Canny边缘检测算法进行研究,针对其存在问题,对各部分的常用方法进行对比,分别选用双边滤波、4方向sobel算子、迭代法对Canny算法进行改进,提高其边缘提取性能;最后根据边缘拟合方法(Hough变换、最小二乘法)对直线和圆的提取效果,选用Hough变换提取直线、最小二乘法提取圆,同时提出同一边缘的多条直线检测和合并策略及冗余圆剔除策略,并通过实验验证了其可行性。3.针对阵列孔板的立体匹配,提出了改进的SGBM方法。该方法利用基于平面拟合的方法剔除孔板区域的误匹配,利用快速均值滤波自适应地填充空洞,并通过相关实验验证了改进措施在孔板区域的有效性,为阵列孔板的尺寸测量和定位奠定基础。4.利用本文提出的方法进行尺寸测量和定位实验。在尺寸测量方面,孔板长宽尺寸上精度达2mm,孔径和孔距尺寸上达0.5mm以上,验证了本文方法的可行性,基本满足工程需求。在空间定位方面,定位精度达到1.5mm左右,精度有待进一步提高。本文研究的成果,能促进双目视觉测量技术在液体工作站中的应用,同时为智能化人机协作设备的研发提供基础,促进医疗生化领域光机电一体化设备的发展。更多还原

【Abstract】 Array orifice is widely used in medical,biological,chemical and other fields.It is often used as the operation object of pipette and subpackage equipment.At present,the reagent bottles or liquid are ofen put into the orifice plate by artificial or related equipment,but most of the equipment is poor in generality,which is only applicable to one or several orifice plates.Therefore,the non-contact measurement and positioning of array orifice plate will greatly improve the working efficiency and intelligence of the equipment.With the extensive practice and application of machine vision in various fields,and the greater advantages of deep learning technology in target segmentation and positioning,many scholars are more and more inclined to use deep learning technology combined with traditional image processing technology to solve practical engineering application problems.At present,Specific methods are designed for specific objects on each vision measurement system,which can not meet the measurement requirements of the object in this paper.In this paper,the array orifice is taken as the object,and the area extraction,feature point extraction and stereo matching of array orifice are studied.Finally,measurement and location of array orifice are completed.For the extraction of orifice area,an improved Mask RCNN method is proposed.The improved Mask RCNN takes Mobile Net as the backbone and P4 + P5 as the FPN.Based on the research of the Mask RCNN algorithm,combined with the difficulty of segmentation and location in this paper,the improvement measures are put forward,and three kinds of backbones and different FPN layers are tested by using the self-made orifice data set.The experimental results show that the improved Mask RCNN can effectively extract the orifice area.For feature point extraction of the array hole plate,an optimized edge extraction and fitting method is proposed.Firstly,the image enhancement algorithm is studied,and the SSR method is selected for image enhancement.Secondly,the Canny edge detection algorithm is studied,and aiming at its existing problems,the common methods of each part are compared,and the Canny algorithm is improved by using bilateral filtering,four-way Sobel operator and iterative method respectively,so as to improve its edge extraction performance.Finally,according to effects of fitting of the edge fitting method(Hough transform,least square method),hough transform is used to extract straight line and circle,and the least square method is used to extract circle,and detection and merging strategy of multiple lines on the same edge and the elimination strategy of redundant circles are proposed,and the feasibility is verified by experiments.For stereo matching of array orifice,an improved SGBM method is proposed.This method uses plane fitting to eliminate the mismatching of the hole plate area,and uses the fast mean filter to fill the hole adaptively.The effectiveness of the improvement measures in the hole plate area is verified by the relevant experiments,which lays the foundation for the size measurement and positioning of the array hole plate.In the last,dimension measurement and positioning experiment of the hole plate are carried out.In terms of dimension measurement,the accuracy of the length and width of the orifice plate is up to 2mm,and the apertures and the hole distance are up to 0.5mm,which verifies the feasibility of our method and basically meets the engineering requirements;in terms of spatial positioning,the overall positioning accuracy is less than 2 mm,which needs to be further improved.The research results of this paper can promote the application of binocular vision measurement technology in the liquid workstation,provide the basis for the research and development of intelligent human-machine cooperation equipment,and promote the development of optical electromechanical integration equipment in the field of medical biochemistry.更多还原